BK26 | When the World Says It's Impossible —— Lisa Su: From AMD's Near Collapse to AI Age

How did a struggling semiconductor company regain the world's confidence? From engineer to AMD CEO, Lisa Su's journey reveals more than a corporate turnaround. It is a story of technological ambition, difficult choices, long-term thinking, and the responsibilities that come with shaping the age of artificial intelligence.

When the World Says It’s Impossible

Lisa Su: From AMD’s Near Collapse to the Age of Artificial Intelligence

Fang Tianliang

BK26 · Civilizational Reflection Series

Editorial Note

This is an edited narrative biography, not an independently fact-checked scholarly edition. It distinguishes documented corporate developments from interpretation and retains some personal anecdotes and reported remarks whose original sources are not yet established. Where necessary, the text uses the concise notice “Currently unverifiable.” Readers should not treat such passages as independently confirmed historical records.

 

Contents

Preface · What Did She Really Overcome?

Part I · A Signboard and a Remote-Control Car

Chapter 1 · The Old House Beside Tainan’s Shuixian Temple

Part II · A Distant Family Connection, Two Different Paths

Chapter 2 · A Distant Family Connection, Two Different Paths

Part III · Thirteen Years at IBM: How an Engineer Became a Leader

Chapter 3 · Choosing the Hardest Major

Chapter 4 · Between Copper and Silicon

Chapter 5 · A Ten-Person Team

Chapter 6 · One Year, and a Complete Map

Part IV · Succession Planning and the Leadership Test

Chapter 7 · The First Time Responsible for Profit and Loss

Chapter 8 · Two Moves

Chapter 9 · The Decisive Test

Chapter 10 · AMD’s First Woman CEO

Part V · $1.61

Chapter 11 · $1.61

Chapter 12 · What She Inherited, What She Changed

Chapter 13 · Three Concentrations

Chapter 14 · The Cost of Long-Term Thinking

Part VI · When Bad News Arrives

Chapter 15 · When Bad News Arrives

Chapter 16 · 2017: Being Believed Again

Part VII · The $293 Million Question

Chapter 17 · Technical Facts

Chapter 18 · Commercial Interests

Chapter 19 · National Security and Moral Judgment

Part VIII · Beyond the Breakthrough: The Next Competition

Chapter 20 · February 2022

Chapter 21 · Making the World Willing to Pay for a Second Choice

Chapter 22 · Two Industrial Orders

Chapter 23 · Life Beyond the Balance Sheet

Final Chapter · As Technology Grows More Powerful, Do We Grow Wiser?

What Did She Defeat

A Larger Picture

But Excellence Does Not Automatically Equal Good

That Question That Refuses to Disappear

What Is Lisa Su’s Answer

The Last Sentence

 

Preface · What Did She Really Overcome?

First, look at a photograph.

One widely circulated account places Lisa Su among prominent technology leaders at a White House dinner in 2026. The date, occasion, and guest list are currently unverifiable. Whether or not that particular scene is accurate, her prominence in the semiconductor industry is beyond doubt.

To sit at that table, one’s weight must be comparable to Musk’s or Huang’s. Her public recognition grew as AMD recovered and expanded into new markets.

Another frequently repeated anecdote describes an F1 pit-lane interviewer failing to recognize Su. The precise year, location, dialogue, and whether the person was Su are currently unverifiable; the episode should not be treated as established fact.

By 2018, Su was already AMD’s president and CEO, even though she was far less familiar to the general public than many consumer-technology executives.

This is not a story about fame. But it does begin with a nearly absurd contrast: a chip company that Wall Street had sentenced to death, a company whose stock had fallen to $1.61 and that analysts publicly discussed as potentially bankrupt any day—why was it able to stand up again?

The easiest answer is “because of Lisa Su.” This answer is both right and dangerous.

It is plausible because different leadership in October 2014 might have taken AMD down another path—spin-off, sale, or slow attrition in a price war. It is dangerous because writing her as a lone hero makes everything truly important disappear: hundreds of engineers rewriting the CPU core for consecutive years; reconstruction already launched by previous management that had not been completed; cash flow from the game-console business in the hardest years; TSMC’s execution on advanced process nodes; capital markets’ renewed willingness after 2016 to assign valuations to chip stocks; and Intel’s consecutive years of process and product-cadence misses, which forcibly opened a window.

All these conditions truly existed. But they were only conditions.

The real question remains unanswered: why do some people, facing such conditions, bet the remaining resources on technology that may only mature years later, while others—including several of AMD’s leaders before her—cut budgets and adjusted directions round after round until the organization no longer had any project that could cross the competitive threshold?

This book tries to answer that question. It unfolds along three lines.

The first is personal. Is difficulty an obstacle, or an opportunity that shapes a person? Why does someone form a response pattern of “going toward it” rather than “going around it”? This part is hidden in her family: her father’s multiplication tables and career as a statistician, her mother’s company built from scratch with an accounting background, and the sentence she said only to her daughter—one day you will come and take over my company.

The second is corporate. Is true leadership the pursuit of immediate interest, or the bearing of long-term responsibility? This is not a question that can be answered by moral posture. Long-termism has a bill that keeps coming due: R&D people must be paid, tape-outs cost money, customers will not wait forever, and investors may lose patience first. Every quarter from 2014 to 2017, AMD answered the trade-off between “surviving today” and “living well three years from now.” And the ways it bought time—including the 2016 technology-licensing deal with China—remain a controversy without consensus to this day.

The third is civilizational. When technology makes life increasingly convenient, are people still willing to train judgment, bear risk, and solve truly difficult problems? Lisa Su is not the only answer. Placing Jensen Huang, Musk, and Bezos against the same era background, one sees something in common: they are all willing to invest capital, reputation, and years of time in things that have not yet been proven. This spirit allows technological civilization to advance. But it does not itself guarantee that the direction is correct. Stronger chips can be used to discover drugs, or to expand surveillance; artificial intelligence can help people learn, or make people abandon judgment.

The three lines ultimately converge into one question: what exactly did Lisa Su defeat?

On the surface, she defeated AMD’s financial crisis and opponents far stronger than herself. At a deeper level, she challenged short-termism, inertial thinking, and the almost instinctive human retreat in the face of enormous difficulty. But if the answer stops here, this book becomes a management bestseller.

So there is a third level. Her success itself raises a new question: as artificial intelligence replaces more and more labor, and even part of everyday thinking, are human beings still willing, like her, to actively solve the most difficult problems? Will a society in which people no longer need to labor for survival produce freer people, or people more easily addicted to instant stimulation?

I have no confidence in answering this question. What this book can do is present one person’s choices as clearly as possible—what she inherited, what she changed, what she bet on, and what price she paid. Then let the reader judge for themselves: facing the same difficulty, why do some people go around it, while others are willing to go toward it?

This is not a question of talent, nor something that can be summarized by an inspirational slogan. It comes from whether a family allows exploration while still requiring responsibility, from whether education trains people to face facts honestly, from whether an organization rewards those who report bad news, and from whether a person, through successive small choices, turns courage into habit.

Now, let us turn the clock back to Tainan in 1969.

 

Part I · A Signboard and a Remote-Control Car

Facing the same difficulty, why do some people choose to go around it, while others are willing to go toward it? The answer is rarely only talent. It usually arises from how a person is educated, how they are allowed to make mistakes, and the internal habits formed after successive choices.

Chapter 1 · The Old House Beside Tainan’s Shuixian Temple

On November 7, 1969, Lisa Su was born in Tainan, Taiwan.

She left the city at the age of three, so her childhood memories do not contain many images of Tainan. Years later, when reporters asked where she came from, she spoke of Queens, New York, of MIT, of IBM. Tainan is more like a background, a starting point she almost does not remember, yet that has always been there.

But this background is worth clarifying, because it is often told as a wrong story.

The popular version goes like this: Tainan produced a prominent family that ran a paper business, dominated the market for more than ten years, and the signboard in front of the old house is still there; the daughter of the prominent family was clever from childhood and destined to become a semiconductor leader. This story reads smoothly, but it turns the place of birth into the foreshadowing of destiny, as if a piece of land would automatically hand achievement to a certain child.

What the real Su family gave their daughter was more specific and more plain than the two words “prominent family.”

The Su family’s business was paper distribution. Distribution is not manufacturing: one need not own a factory or invent products; one relies on channels, credit, and long-term relationships—delivering on time, keeping accounts clear, pricing fairly, so that people are willing to deal with you for decades. This is the typical way of life of Taiwanese small and medium enterprises, and one of the most solid forms of commercial training of that era. What it left behind was not a fortune that allowed one to lie down, but a way of looking at business: what truly determines survival is often not how clever a particular technology is, but whether someone is willing to trust you over the long term.

Interestingly, more than forty years later, when Lisa Su was asked how she brought AMD out of desperate straits, the sentence she repeatedly emphasized was “semiconductors are a team sport.” She needed TSMC, needed Samsung and SK Hynix, needed OEM customers, needed developers and cloud service providers, and needed the hundreds of engineers inside her own company to be willing to stay. There may be an observable link between the family’s commercial background and the collaborative philosophy she later emphasized; but these experiences alone cannot establish a direct causal relationship.

The maternal Luo family was also rooted in Tainan, living near Shuixian Temple, a large clan with many members and a thriving second generation. Regarding this family, outside attention has almost entirely focused on one point: one daughter of the Luo family later became Jensen Huang’s mother, so there is a distant kinship between Lisa Su and Jensen Huang. We will handle this topic specially later. Here it is only necessary to remember one thing: Lisa Su left this family network at the age of three.

She left Tainan very early and received her education in the United States. As for how much the Taiwanese family network actually played a role in her growth, existing materials are insufficient to make an assertion.

Putting these two things together, what the Su family left her becomes clear: on one side the credit concepts that make a business work, on the other side the reality that there is no shortcut. The former taught her how to cooperate with people over the long term; the latter forced her to learn how to solve problems on her own. Everything she later did in her life was essentially between these two points—running toward the most difficult problems (because no one else would run for her), while firmly believing that solving them required a group of people.

 

A Beginning That Does Not Need to Be Mythologized

Regarding her childhood, the most frequently cited story is that she took apart a remote-control car at the age of ten.

This story is often told as a sign of a prodigy: look, she loved taking things apart from childhood, just like Musk. Told this way, the story is actually wasted. Because the real emphasis of the story is not “taking apart,” but “putting it back together,” and her father’s reaction.

She took apart her brother’s exquisite remote-control car into a pile of parts. This is something many children do; the usual outcome is a scolding. She was not scolded. Her father not only did not blame her, but later bought her more complicated toys.

This reaction contains two things, and both are indispensable.

The first is tolerance. If a family only allows success and does not allow things to be damaged, children quickly learn one thing: exploration is risky, so it is better not to touch. As adults, such children become good executors, but rarely become people willing to make decisions amid uncertainty. The second is requirement—the car was eventually put back together. Taking it apart is curiosity; putting it back is the desire to complete. The parents did not stop at “letting her play,” nor did they do it for her.

Freedom without responsibility may not be enough to cultivate an engineering spirit. Responsibility without freedom can suppress the desire to explore. The account suggests that the family sought to balance freedom and responsibility. The account illustrates two concrete responses: not scolding her and encouraging further exploration. The details are currently unverifiable.

 

Multiplication Tables at the Dinner Table

A deeper influence than taking apart the car came from a dinner table.

Her father, Su Chunhuai, was a mathematician. He graduated from the mathematics department of National Taiwan Normal University, later obtained a master’s degree in mathematics from National Tsing Hua University (Hsinchu), received a full scholarship to Columbia University in 1967 to pursue a doctorate, and afterward worked as a statistician for the New York City government.

Note this occupation: statistician. This is work that requires a person to continually use data to judge reality—which district’s crime rate is rising, which budget has a higher input-output ratio, whether a certain policy is effective or only feels effective. For someone who faces the allocation of public resources every day, opinions are worthless; evidence is what matters. The father brought this home.

He used multiplication tables to “test” his daughter at the dinner table.

The word is accurate: test, not teach. Teaching is explanation; testing is setting questions. The child must first calculate on her own; if she gets it wrong, she goes back to find the reason. This daily action transmits a method of knowing the world, not a few multiplication answers: facing a problem, first use your own mind; do not wait for others to give the conclusion; first look at the data; do not judge by feeling.

Anyone who has studied statistics will admit that this training shapes not only arithmetic ability, but the character of still being willing to advance judgment when information is insufficient. The decisions the later Lisa Su had to make were none of them made only after the data were complete. Whether to abandon the low-end market, whether to concentrate resources on one architecture, whether to continue burning money when cash flow was tightest—these questions never had standard answers. She could make decisions because she had been trained from childhood to “calculate first, then decide.”

 

The Sentence Said Only to the Daughter

If the father gave her method, the mother gave her direction.

Mother Luo Shuya came from an accounting background. She arrived in the United States speaking English poorly, yet relied on her professional ability to build a company from scratch.

About the mother, it was Lisa Su herself who spoke of it when she returned to Taiwan in 2025. She said her earliest role model was not any tech giant or business leader, but her mother.

The words her mother often said to her were not “you must study hard.”

They were—“One day you will come and take over my company.”

This sentence is worth pausing on for a few seconds. Because its nature is completely different.

The goal of “you must study hard” is personal achievement; it points to grades, degrees, a good job. The goal of “you will take over my company” is bearing responsibility; it points to other people’s livelihoods, a real existing organization, a group of people who depend on it for a living. The former requires you to become excellent; the latter requires you to become reliable.

And this sentence has a detail: she said it only to the daughter, not to the son.

The same family, the same resources, two children given different expectations. In those years, a mother who had built a business from scratch willing to hand the enterprise to her daughter was itself a choice. It told this girl two things: first, you can become the person who decides; second, the price of deciding is that you must be responsible for the results, not only for your own performance.

Later the media liked to cite another detail: the career options her mother gave her were doctor, pianist, or engineer, and she chose electrical engineering. This choice is often told as “she preferred the hardest road from childhood.” This interpretation may be too heavy. A teenage girl’s choice of major involves interest, ability, the practical considerations of an immigrant family, and the opportunities of the era. But whatever the reason, the result pointed in the same direction: she was placed in a trade where she had to solve real problems with her own hands.

The contrast between her mother’s expectation and her later choice to bear corporate responsibility can be meaningful, but one cannot assert a direct causal relationship on that basis. When the AMD board asked her to become CEO, she could have chosen a more respectable company and a more stable résumé. She went to the one that was bleeding. This is the same action as her mother once shouldering a company.

 

Bronx High School of Science

Around 1972, when Lisa Su was three, the whole family moved from Tainan to New York and settled in Queens.

Queens in that era was a microcosm of the immigrant world: dozens of languages, dozens of livelihoods, children in the same building coming from completely different places. This is a place that easily teaches people reality, and a place that gives no special treatment. The respect you can obtain here comes from your ability, not from your origin.

She later entered Bronx High School of Science and graduated in 1986. This is one of the top public magnet schools in the United States, having produced a number of Nobel laureates. Regarding this school, the most valuable point is not its reputation, but what it means: its admission screening does not look at family background, only at examination.

For this girl, the school’s selection system provided a path to compete for opportunity through academic performance.

What is worth noting is what she would learn in such an environment. Intelligence is not scarce here; the whole class is intelligent. The efforts of immigrant families do not automatically convert into advantage, because every family around is also making efforts. In such an environment, the conclusion a person most easily draws is “I am not as special as I imagined,” and the habit most easily formed is “break the problem apart and finish it bit by bit.”

Bronx Science did not teach her confidence (she has rarely shown overt confidence throughout her life); what it taught her was something more durable: complex problems can be decomposed, and difficult things can be trained.

Years later, an IBM executive would use one sentence to name this habit.

 

A Character That Still Had No Name

At this point, Chapter 1 can make a summary.

What we see is a very ordinary starting point: an immigrant family, parents of middle income, a school district that was neither second-rate nor first-rate, then testing into a good public high school. No single link is sufficient to explain the later achievement.

But some things were taking shape.

She learned to calculate first herself (father’s multiplication tables). She learned to be responsible for results (mother’s succession requirement). She learned that exploration does not bring punishment (the disassembled remote-control car). She also learned that after exploration one must put it back (the car must be restored). She further learned that difficulty can be decomposed (the daily life of Bronx Science).

Added together, these are not yet “character.” They are more like a batch of parts that have not yet been assembled into one thing.

Assembling that thing required an external action—someone putting a sufficiently difficult problem in front of her and telling her: do not go around it, run toward it.

That thing happened after 1994, at IBM.

 

 

Part II · A Distant Family Connection, Two Different Paths

Chapter 2 · A Distant Family Connection, Two Different Paths

Now we must handle the section of this book that is most easily told badly.

Lisa Su and Jensen Huang are relatives. The way this matter has spread on the Chinese internet is roughly as follows: both were born in Tainan, both immigrated to the United States as young children, one is CEO of AMD, one is CEO of Nvidia; even better, there is a genealogical relationship between them—Lisa Su’s maternal grandfather Luo Bomu was the older brother of Jensen Huang’s mother Luo Caixiu. Thus Lisa Su’s mother is Jensen Huang’s cousin, Lisa Su is Jensen Huang’s “cousin-niece,” and by Chinese terms of address should call him “cousin-uncle”; by Western genealogy, they are “second cousins once removed.”

Then the usual sentence is added: so they are actually “one family.”

This sentence sounds warm, and also cheap. Because it requires almost no evidence to put a shell of family epic on the competition between the two companies—the same Tainan family producing the helmsmen of two chip kingdoms, facing off in the artificial-intelligence era. This story is too smooth, smooth enough not to seem true.

And that conclusion does not follow—at least not in the way the popular narrative suggests.

 

A Relationship That Truly Exists but Has No Content

First confirm the factual part: this kinship is real. Luo Bomu and Luo Caixiu were brother and sister, differing by about eighteen years. The Luo family near Shuixian Temple was a large clan; the second generation had at least twelve people. So from the family tree, Lisa Su and Jensen Huang are indeed connected.

But a family tree only shows that your names appear on the same chart; it does not show that you know each other.

Lisa Su immigrated with her parents at the age of three and settled in Queens, New York. On Jensen Huang’s side, the family first went to Thailand, then eventually to Oregon in the United States. One family was on the East Coast, the other on the West Coast toward the north, separated by more than three thousand kilometers and an entire continent. The second generation of the Luo family had more than a dozen people, scattered in different places, each starting families and making a living. The two families never grew up together, nor had a shared family life.

Lisa Su herself has said this very clearly. She has publicly stated that this relationship is “really very distant,” that the two did not grow up together, and that they only met at industry events late in their careers. The gist of her original words is: no family dinners, just an interesting coincidence.

Jensen Huang’s response has also been frank. He has confirmed that it is a distant maternal relative and has handled the matter with his usual humor—“we are family,” “we are both in this industry,” “we buy AMD’s stuff,” “we should keep it in the family,” and “I’m proud of her.”

A detail worth noting is that the two later did appear together at some events, such as the groundbreaking of TSMC’s U.S. plant. But that kind of co-appearance is industry-related, not family-related. What has occurred between them is two companies competing for the same batch of customers in the same market—the weight of competition far exceeds the weight of kinship.

 

Why This Matter Deserves a Chapter

Since the relationship is so distant, why write about it?

Because this coincidence has extremely strong explanatory temptation, and resisting this temptation is precisely the prerequisite for understanding these two people.

If one admits “they are one family,” people will continue thinking downward: is there something special in the Luo bloodline? Is there some gene in the place of Tainan? Did the family education of a prominent clan, the strictness of a Minnan-speaking household, and the striving of immigrant families together produce two chip giants?

Each of these inferences sounds reasonable; added together, none holds.

First, family genes cannot explain specific achievements. The decisions made by Lisa Su and Jensen Huang are completely different in nature: one took over a nearly dead company, chose to rewrite the architecture, concentrate resources, and survive by external manufacturing; the other, from founding the company in 1993, has always been betting on an architecture the market had not yet accepted, and spent twenty years building a software ecosystem. These judgments did not grow out of bloodlines; they grew out of their respective situations, teams, industry positions, and era windows.

Second, shared culture cannot explain different choices. The two do share certain backgrounds: born in Tainan, immigrated as young children, families that valued education, Minnan-speaking household atmosphere. But thousands of people share the same background and have become accountants, doctors, engineers, restaurant owners. Culture can explain a tendency; it cannot explain a specific success.

Third, and most important: attributing the two people’s achievements to the family erases what each truly relied on. Lisa Su relied on thirteen years of technical training at IBM and a group of engineers willing to follow her to finish Zen; Jensen Huang relied on the CUDA ecosystem and a group of developers willing to rewrite code for him. None of these can be given by family.

 

What Is Truly Worth Observing Is Two Immigrant Families

Then what is the only part of this relationship that has explanatory power?

It is not whether they knew each other, but that two immigrant families, in a specific historical period, used similar methods to cultivate two engineering-type leaders.

These two families have some things in common.

The first is treating education as a clear responsibility rather than a vague expectation. Earlier we wrote that Lisa Su’s mother said to her “one day you will come and take over my company.” There is similar narrative on Jensen Huang’s side—families having concrete requirements for their children, not merely “hoping you amount to something.” Such requirements are usually not gentle: they bind learning and responsibility together.

Another is long-term thinking shaped by limited safety nets. Immigrant families usually have no ancestral property to inherit and no network of connections to fall back on. This forces them to think about longer-cycle questions: can a career be done for twenty years, will a skill be replaced. This way of thinking is passed to the children.

A further theme is the combination of family support and personal responsibility. Interestingly, both families simultaneously gave their children two things that look contradictory: one is a rear that can be relied upon (family support, allowing trial and error), the other is a front that does not permit escape (you must solve it yourself). The set of actions mentioned earlier when writing about taking apart the remote-control car—not scolding her, while buying a harder toy—is precisely the microscopic form of this combination.

If one must extract something from this distant kinship, it is this: what they shared was not the power of bloodline, but the same set of work ethics shaped by the immigrant situation.

And ethics can only explain the starting point; they cannot explain the endpoint.

 

Walking Different Trajectories

Starting from similar points of origin, the two people walked quite different paths. These several differences will reappear repeatedly in later chapters.

Another difference lies in what each leader was betting on. Jensen Huang’s core bet is architectural direction—betting on parallel computing and GPUs, and persisting on this path for twenty years, during which there were long years when the market did not recognize it. Lisa Su’s core bet is organizational direction—betting on which markets the company should abandon, which teams should be retained, and concentrating resources on a few product lines. The former is “I believe the world will become like this”; the latter is “I believe this company can still be reorganized in this way.”

A further difference lies in their public communication styles.

Jensen Huang is a rare “namer.” He continually gives names to a world that has not yet fully arrived: accelerated computing, AI factories, physical AI. He is good at turning a complex technical turn into a grand story that everyone can understand. His public image is expansive, infectious, carrying a certain preaching quality.

Lisa Su is the opposite. Her public expression is restrained, precise, technically solid, and rarely depicts distant visions. The way she speaks is closer to an engineer reporting progress: how much performance this generation of product improves, when this generation will launch, where the next node on the roadmap is. She almost never says “we are going to change the world”; what she says is “we are going to deliver.”

This difference is often simplified into the opposition between “visionary leadership” and “execution-type leadership.” This simplification is wrong. Nvidia’s execution capability is first-rate in the industry—the CUDA ecosystem was not built by speeches; it was raised bit by bit through twenty years of engineering investment and developer operations; AMD’s revival likewise required bold judgment. Looking only at execution and not at judgment cannot explain why one dared to rewrite the entire CPU architecture or dare to abandon the low-end market.

A more accurate statement is: when facing the public, each amplified the part that needed to be amplified given their strategic position.

A challenger needs to lower other people’s risk of change. Lisa Su had to make cloud service providers, OEMs, and developers believe that choosing AMD would not make them regret it, because we will deliver continuously, will exist for the long term, and will support open standards. Her restraint is not introversion of personality, but a communication strategy determined by position.

 

They Did Not Moralize Each Other into Enemies

Perhaps the most worth learning from in this relationship is the way both sides handle competition.

When Lisa Su talks about this kinship, she uses “distant, interesting coincidence,” never emphasizes family, and never describes Jensen Huang as an opponent who must be defeated. On Jensen Huang’s side, there are reports that he once made a joke in an internal setting to the effect that “Lisa Su’s blood made me have to clean toilets at school when I was little”—this statement is indicated as humorous banter and should not be written as factual family narrative in a biography.

More worth recording is the respect he has publicly expressed: “we are family,” “I’m proud of her.” And the small matter that has appeared in many reports: he has said “we buy AMD’s stuff.”

This is not posture, but a higher-efficiency way of competing. Moralizing the opponent into an enemy can obtain temporary emotional mobilization, but it distorts one’s own judgment—you begin to underestimate the other side, and begin to interpret every action of the other side as malice. Mature competition requires accurate assessment of the opponent: where he is strong, how his moat was built, which part can be caught up with, and which part should not be touched in the short term.

Jensen Huang’s sentence “I’m proud of her,” and the candid acknowledgment behind that sentence, is itself a sign of industrial maturity.

 

 

Part III · Thirteen Years at IBM: How an Engineer Became a Leader

Chapter 3 · Choosing the Hardest Major

In the autumn of 1986, a seventeen-year-old girl graduated from Bronx High School of Science and entered the Massachusetts Institute of Technology.

Her original intention was electrical engineering or computer science. In the first two undergraduate years at MIT there is a batch of foundational courses that students privately call “weed-out courses”—6.001 and 6.002 among them. The design intent of these two courses is clear: use extremely high difficulty and extremely fast pace to make some students themselves realize they should change direction.

She finished these two courses, then chose electrical engineering.

The reason was: she thought it was harder.

This detail is worth looking at slowly. Because it is easily read as a personality label—“she has loved challenges since childhood.” Read this way, an explainable choice becomes an unexplainable talent.

A more reasonable reading is: this is a judgment made after being tested by the weed-out courses. She knew where the difference in difficulty between the two sides lay, knew what price this major would require, and then she chose the side with the greater price. This is not impulse; it is a choice made after mastering sufficient information. She was not betting that she was smart; she was betting that she was willing to pay this bill.

She paid this bill for eight and a half years.

MIT’s system is usually four years of undergraduate; she stayed there eight and a half years, obtaining bachelor’s (1990), master’s (1991), and doctorate (1994). In 1994 she was 24.

 

Starting from Two-Inch Wafers

Many people think a doctorate begins in the laboratory. She did indeed begin in the laboratory, but not from any profound research.

In her freshman year she applied for a position in the Undergraduate Research Opportunities Program (UROP) and was assigned to Professor Hank Smith’s semiconductor laboratory in Building 39; the task was to fabricate two-inch wafers for X-ray lithography research.

She later said herself that this work “was pure hard labor.” But she immediately added: “But as an undergraduate I didn’t know; it was great.”

These two sentences together are the honesty of a person looking back at youth after maturity: the work was indeed hard, but that hardship gave her something more important—she first came into direct contact with the real semiconductor manufacturing process. Not structural diagrams in textbooks, not curves in simulation software, but physical objects that had to be made in the clean room and taken for measurement. If a wafer was made badly, it was bad; the temperature, time, and cleanliness of the process would be directly reflected in the result.

Manufacturing is a teacher that shows no mercy. It does not reward cleverness; it only rewards correctness. This cognition would return more than twenty years later at the moment AMD needed it most: chips are not made by writing slides; they are made by finishing the last five percent.

 

A Direction Not Yet Proven

From undergraduate to doctorate, her research direction was SOI—silicon-on-insulator.

The basic idea of this technology is: not to make transistors directly on a whole piece of silicon, but first to lay a layer of insulating material on the silicon substrate and build the transistors on top of this insulating layer. The benefit of doing so is concrete—transistor efficiency is higher, and speed and power-consumption performance can be better. Today SOI is used both to improve chip performance and to reduce power consumption.

But in the years she was doing research, this direction was not yet mature. The technology had not yet been proven; industry had doubts about whether it could be mass-produced; not many teams were willing to invest.

In other words, when she was doing her doctorate, she chose not a hot track, but a long, possibly never-to-become-product direction.

This matter is often handled in biographies as a romantic choice—young people betting on ideals. But placing it on a more practical level makes it more interesting: a person willing to invest the most precious years in an uncertain direction is a capability, not a personality.

Almost all the key decisions in her life have this structure: betting when the direction has not yet been proven by the market.

 

Her Advisor Gave a Piece of Advice

Before she prepared to leave MIT and go to work at Texas Instruments, her advisor Dimitri Antoniadis gave her one sentence:

“Keep technical depth for as long as possible.”

The surface meaning of this sentence is not to rush into management. But its true implication is deeper and more consistent with what later happened.

The promotion path in the technology industry usually encourages people to leave the technical front line as early as possible—switch to project manager, switch to business, switch to management, because management positions rise faster, and technical positions are easily seen as “the people who do the work.” The cost of this path is: within two or three years a person loses judgment of technical details and thereafter can only make decisions by listening to other people’s reports. And once one can only listen to reports, one cannot distinguish which plan has truly been thought through and which only looks reasonable.

Antoniadis’s advice was actually saying: do not leave the front line before you have formed judgment. Technical depth is your only irreplaceable capital at the conference table.

She later did stay in technology for a long time—IBM research, process, devices, architecture, R&D management, more than ten years. This is the place where she was most different when she took over AMD in 2014. At that time AMD did not lack people who could do financial restructuring; what it lacked was a CEO who could understand what engineers were saying and also judge which plan was truly doable.

This sentence has another layer. She later opposed the saying that “technical people and commercial people must choose one or the other.” In a 2006 interview with MIT Technology Review she said she greatly enjoyed switching back and forth between these two domains. This sentence is not pretty talk; it is the long-term fulfillment of Antoniadis’s advice.

 

Two Summers of Industrial Internships

During her studies she did a summer internship at Analog Devices.

Public records of this experience are few, but it played a clear role in her path: it let her first see the complete chain of a chip going from laboratory to product.

The laboratory cares whether a device can work. Industry cares about completely different things: whether it can be stably and repeatedly made, what the cost is, how high the yield is, whether customers are willing to pay a little more for this performance, and when it can be delivered. A scheme that can be called successful in the laboratory may be abandoned outright in industry because the yield is one percentage point worse.

This gap is the first obstacle many researchers encounter when turning to industry. Some cannot get past it and feel that industry “does not respect technology.” She got past it, and digested it well. Later when she talked about semiconductors being a team sport, about products needing to be delivered rather than looking pretty, about the last five percent being where the secret lies—all of these carry the traces of two summers.

 

The Boundary of Judgment at Age 24

In 1994 she left MIT with a doctorate. This was the result exchanged for all the training before: a hard skill, a research method, a character willing to bet long-term.

But she did not yet know how to solve certain problems; this must be made clear.

She did not know how to allocate resources for a large organization—this requires understanding how budget, politics, talent, and timing constrain each other. She did not know how to explain to a board a decision that would only take effect three years later, because in the laboratory she only needed to explain the next experiment to her advisor. She did not know how to stop a direction once it had been proven wrong and persuade dozens of people to accept that conclusion. She did not even know whether she would like managing people, because before this she had managed only her own experiments.

This is the difference of a step: from solving problems to discovering which problem should be solved.

The place she was about to go happened to have a problem that had stuck the entire industry for more than ten years waiting for her. And the action that pointed the way for her would establish her mode of behavior for the next thirty years.

 

Chapter 4 · Between Copper and Silicon

After completing her doctorate in 1994, Lisa Su joined Texas Instruments and subsequently moved to IBM.

Public information about the Texas Instruments period is very little; there is only one cautious description: she stayed there less than a year. Biographers usually pass over it in one stroke, as if it were only a brief transition. But the decision itself indicates something—she stayed only a few months at a large company and left, and it was in 1994. That year the U.S. semiconductor industry was booming; a stable job was something many people sought. She left of her own accord, going to another larger company, still doing research—but that direction was completely different from the SOI she had studied for her doctorate.

The problem IBM handed her was copper.

 

A Problem That Had Stuck the Industry for More Than Ten Years

To understand why this problem was hard, one must first see what was happening inside a chip.

Chips rely on transistors to process signals, but transistors must be connected by metal lines in order to form circuits from tens of thousands of devices. These connecting layers are called interconnects. Traditionally aluminum was used.

The industry had long known the direction of the answer: switch to copper. Copper’s conductivity is much better than aluminum’s; theoretically it can make signals run faster and with less loss.

But copper has a fatal flaw. During manufacturing, copper atoms diffuse into the surrounding silicon and insulating layers. This is not merely a matter of “a little contamination”—diffusing into the active region of the device changes the electrical characteristics of the semiconductor and causes entire batches of chips to fail. Even more troublesome is that silicon-wafer manufacturing cannot be remedied afterward: once contamination occurs, that wafer is scrapped.

So the full picture of this problem is: a material with better performance could not be used for more than ten years because it could not be kept “well-behaved” in production.

What it required was not only a materials invention, but an entire set of solutions: finding a barrier layer that could block copper atoms, figuring out the temperature and time window for annealing, coordinating adjustments to device design rules, and finally making this process stably repeatable on the mass-production line. If any link was missing, the solution existed only in papers.

 

She Later Admitted: This Was Not Her Specialty

Here is a particularly noteworthy point.

Her own statement is: she merely “migrated to where the problem was.”

This sentence is far more accurate than “she invented copper interconnects,” and better explains her true working method. Semiconductor research does not advance straight along a personal résumé. The training one has received gives one a set of ways of handling problems—how to define variables, how to design control experiments, how to judge trends when data are incomplete, how to narrow the search range when failing. But where the problem appears often does not depend on what one wants to research.

When a truly important problem appears, the engineer must bring existing capabilities over and then make up the parts one does not know.

This is a very plain, and very rare, professional attitude. Most people are accustomed to digging deeper and deeper in their own specialty; the advantage is safety, the disadvantage is that when the industry’s problems shift elsewhere, their depth is exactly not there. She did the opposite: let specialty serve the problem, rather than let the problem accommodate the specialty.

 

1998: Copper Goes to Market

Around 1998 IBM launched copper-interconnect technology and advanced it to mass production. Chips using copper connections could obtain roughly 10% to 20% speed improvement relative to traditional aluminum interconnects.

This number does not look astonishing today. But at the time it meant the technical route of the entire industry was rewritten once. Today the chips in your phone, computer, and server almost all use copper connections. Forty years ago this was something that required persuading the whole world.

One thing must be written clearly: this is not her individual achievement.

This project had materials teams, process teams, equipment teams, device-design teams, and manufacturing personnel. The formulation of the barrier layer was the battlefield of materials scientists; mass-production stability was the battlefield of process engineers; device rules had to be aligned with the design team. If any link was missing, copper could not leave the laboratory.

When MIT Technology Review listed her among the “Innovators Under 35” in 2002, the wording was: she played a key role in integrating copper connections into semiconductor chips—including solving the problem of copper-impurity contamination of devices. This wording is accurate. She was not the sole inventor; she was one of the people in this system who solved key engineering problems.

Why is this point important? Because later biographies almost all turn this chip into a genius story. Told this way, it misunderstands both the semiconductor industry and her. Semiconductors are industry, not personal art. The fact that she repeatedly emphasized “team sport” throughout her life is not public-relations language, but first-hand fact she learned from her first important project.

 

What She Truly Took Away Was That Five Percent

Years later, in an interview she was asked what the most important thing she learned at IBM was.

She told something that sounded very unsuccessful. She was working on a copper-interconnect microprocessor project. The research stage looked almost finished; she very much wanted to try new things, but her superior made her stay and finish delivering the product the last stretch.

She did not understand at the time. Later she discovered that the last 5% required to truly push a product out was roughly the hardest part, and also where most of the “secrets” lay.

What is in this five percent?

Yield ramp—samples that work in the laboratory do not mean every wafer on the production line can work. Design-rule convergence—device design must find the actually feasible intersection among performance, area, reliability, and manufacturability. Reliability verification—the chip must not fail under high temperature, high voltage, and long-term operation. Customer verification—partners must be made to believe the chip meets specifications. Manufacturing variation—differences exist among different wafers in the same batch; the process must be able to accommodate them.

These things have one common characteristic: they do not produce brilliant papers and cannot be made into pretty demos, but if any one fails, all the previous work returns to zero.

She originally thought she should leave. She stayed. This finishing work that makes young engineers impatient was later called by her one of the most valuable pieces of training in her career.

Twenty years later, when she needed to persuade AMD’s customers to trust the company again, she was actually facing the same problem. Customers do not look at roadmap slides; customers look at whether the previous generation of product was delivered on time, whether there was stable yield, and how problems were handled when they occurred. Credibility is accumulated chip by chip.

 

Three Things This Experience Left Behind

If one is to inventory at the end of this chapter what she took away from the copper-interconnect project, there are roughly three things.

Another lesson was cross-disciplinary collaboration. For the first time in a real industrial environment she saw how materials, devices, process, design, and manufacturing constrain one another. This enabled her later at AMD to understand: investing resources in an architectural decision actually simultaneously pulls every link on the entire chain; any shortcoming in any link will drag the whole.

A further lesson was the habit of “moving to where the problem is.” This is her most important mode of behavior, and also the boundary between her and “specialist”-type technical leaders. She never bound herself to a particular domain, but let capability serve the problem.

Around 1998 copper interconnects went to market. At that time she was not yet thirty and, in the research system of a huge company, was one of many contributors.

But her position was changing. Because in the next two years there would be an opportunity that turned her from a technical expert into a person who had to stand at the boundary of two worlds and speak.

 

Chapter 5 · A Ten-Person Team

Around 2000 the IBM copper-interconnect microprocessor project was wrapping up. Lisa Su’s situation was: five or six years after obtaining her doctorate, having done a solid stretch of process and device research, having been involved in a technology project that rewrote the entire industry, two paths now lay before her.

The first was the familiar path. Return to the research department, manage larger technical teams, continue upward in the spectrum of semiconductor process. This path is stable and clear; peers all walk it this way.

The second path she later said very frankly herself—she wanted to try management and the business side.

She obtained a direction: forming a group called “Emerging Products.”

 

A Deliberately Small Organization

The initial scale of this group was ten people, all recruited by her own hand.

Its research directions were low-power and broadband semiconductors, and also biochips. Placed in IBM in 2002, none of these things counted as the company’s main business. IBM’s real center of gravity at the time was servers, mainframes, and the semiconductor research system. Low-power devices were a field that was rising but had not yet become climate; biochips were even farther from IBM’s core business.

In other words, this was not a team granted important resources, but a small team allowed to find its own direction.

She herself described this state as “like running a startup company, yet able to use IBM’s resources.”

This sentence sounds like a benefit; actually it describes a unique management situation—she simultaneously possessed two things that usually do not appear together: a very small organization, and a very large backing.

Small means she had no bureaucratic process to rely on, and no brand halo to borrow. Ten people doing one thing, who is working and who is not, can be seen at a glance. The team had to find customers themselves, define products themselves, and prove the scheme feasible in the technical community themselves.

The combination of these two gave her a very rare training environment: she could practice management amid real commercial risk without having to bear the stake of a startup company in which one failure ends everything.

 

She Found a Word for Herself: Translator

Later when asked about her role during this period, she said one word: translation.

This statement is worth unfolding, because it explains what she has truly been doing for the subsequent more than twenty years.

In a ten-person group, “translation” sounds like a small matter. But in a company like IBM, technical people and operators use two completely different scales to judge the same thing.

The technical people’s scale is physical limits, performance, reliability, process feasibility. Whether a scheme can be made, and after being made whether it can be stably repeated—these are the questions they care about.

The operators’ scale is market size, development cycle, gross margin, return on investment, competitive timing. Whether a scheme is worth doing, and after being done whether it can make money—these are the questions they care about.

Both scales are correct. The problem is they often reach opposite conclusions: the technically most elegant scheme may have no market; the product most needed by the market may be technically hard to make, or made too late, with competitors already occupying the position.

The translator’s work is to make both sides believe the team knows what it is doing.

This sentence sounds gentle and smooth; in practice it is very hard. Translation is not simply replacing terminology with plain language. It must bear two heavier pieces of work.

The first is protecting engineering facts from being distorted by commercial wishes. When the business department hopes a certain product will launch three months early, the technical people know what that means—testing must be compressed, yield has not finished ramping, risk is shifted later. At this time the translator must tell the truth, even if the words are unwelcome on the spot. How many technical projects are ruined by this kind of silence: no one stands up and says “this timeline cannot be done,” so everyone tacitly assumes it can be done, until just before launch they discover it cannot.

The second is acknowledging that technology itself cannot automatically create value. Technical people easily have a default assumption: if something is made better, the market will naturally discover it. This assumption rarely holds in the real world. A product needs scenarios, needs customers, needs a price that can be accepted, needs someone willing to change existing purchasing habits. An exquisite technology that cannot find someone willing to pay is a cost to the company.

 

Protecting a Direction Not Yet Believed

Her judgment was truly tested here for the first time.

When a new direction appears, there are usually three evaluations: no value, value unclear, value clear. The first can be abandoned immediately; the third everyone is willing to invest in. The real difficulty is the second—value unclear, which is exactly the category she has bet on repeatedly throughout her life.

The trouble with this category of direction is that it has no clear evidence that lets you say “this is correct.” What you can usually produce is only partial data, some reasonable inferences, and a few early technical-feasibility verifications. The people who support it and the people who oppose it are both speaking inference, not fact. In this situation organizations tend to postpone decisions—because postponement does not require bearing the cost of a wrong judgment.

Her way of handling it later became the most consistent pattern in her life: shrink the direction to the smallest range that can be verified, first prove it feasible, then fight for larger resources.

This point is extremely critical. She is not the kind of leader who uses vision to persuade resources. Her way is first to make a thing, turning the question “does this hold in principle?” into “this can already be used; do we do more?” The former requires belief; the latter only requires judgment. And the latter is always easier to reach consensus on.

 

Cell: A Technical Splendor

The most important project of this stage was the Cell processor jointly developed by IBM with Sony and Toshiba.

Cell was an extremely ambitious design. It attempted to redefine the processor with a highly parallel structure: one general-purpose core plus multiple co-processors (called SPEs), theoretically able to obtain extremely high performance under specific workloads. One of its target markets was game consoles—Sony later used it in the PlayStation 3.

From a technical point of view, Cell was a rare experiment in industry. Making an architecture this radical required cross-company collaboration, a brand-new programming model, brand-new development tools, and several years of time.

And Cell’s historical value is more complex than its success at the time.

It was splendid technically and heavy in market strategy. The PlayStation 3 was ultimately regarded as a relatively less successful generation among Sony’s successive consoles. One of the reasons lay in this chip: parallel architecture is hard to program; developers must learn a new model and accept more complex optimization work. Sony later gradually adjusted its strategy, shifting from reliance on hardware differentiation to reliance on exclusive game content.

Lisa Su’s own evaluation of Cell was restrained in wording. She said this processor was extremely ambitious at the time and explored in the direction of parallel computing; from a commercial point of view it was of course successful; but if ranked by history, the conclusion might be different.

There is a layer in these words that she later valued highly herself: technical leadership and market success are not the same thing.

An architecture can be invincible on paper, can excite engineers, can appear on magazine covers, but if developers are unwilling to rewrite code for it, if the entire ecosystem does not grow along with it, its advantage cannot be realized. Performance numbers cannot solve the customer’s learning cost.

Cell was not wasted. What she obtained from it would be used twice at AMD.

The first time was understanding of this customer type of game consoles. Having participated in the Cell project meant she clearly knew how console customers do planning: a platform’s lifecycle usually spans five to seven years; the chip must be locked years before the console launches; cost, heat dissipation, volume, and supply stability are more important than peak performance. When AMD later competed for the semi-custom chips of PlayStation and Xbox, what she possessed was not technical capability in the ordinary sense, but familiarity with the decision logic of this type of customer.

 

A Special Task About to Arrive

At a point during her IBM career, she received an unusual assignment. The exact appointment date is currently unverifiable.

The company told her there was a one-year task: to serve as technical assistant to Chairman and CEO Lou Gerstner.

She was still relatively early in her IBM career.

She later said her reaction at the time was confusion: why? Why would he want to talk to me? That was a very strange thing.

Thus a young person who had been on the job five years and had previously mainly done pure engineering research was placed in the seat closest to the company’s power center.

The next year would completely change her understanding of the word “problem.”

 

Chapter 6 · One Year, and a Complete Map

Gerstner’s office was at IBM headquarters. During that assignment, Lisa Su encountered problems beyond the technical questions she had previously handled.

 

The World She Originally Knew

Before this, her world had clear boundaries.

In the laboratory a problem can be limited within materials, devices, or architecture. The variables are ones you choose, the conditions are ones you set, the control experiments are ones you design, and the conclusion is finally judged by data and peers. If wrong, the cost is a few months of time and one public failure.

This world has one enormous advantage: right and wrong have objective standards. Whether copper has diffused into the silicon, measure and you know. Whether the device’s speed has truly increased 20%, run it and you know. No one needs to persuade others to confirm that they are right.

This world also has one enormous limitation: the problem is defined for you by others. All the difficult problems she solved in these seven or eight years were not ones she chose. Copper was chosen by IBM, SOI was the direction circled by her advisor, Cell was a company-level collaborative project. Her role was to find the best answer within given boundaries.

This work requires extremely high intelligence and extremely strong patience, but it does not require one thing: judging which thing should be done when there is no standard answer.

 

The Other Map She Suddenly Saw

In her own words, the greatest gain of that year was seeing with her own eyes how the CEO of a large company thinks.

What specifically does this “how one thinks” refer to?

First, the technically most excellent scheme is not necessarily the scheme that should be done.

In the evaluation system she was familiar with, “better” is “better.” But in Gerstner’s position a scheme must pass several gates at the same time: will it still hold three years later? Will competitors do it at the same time? Does the company have the ability to deliver it? Will customers change purchasing habits for it? Doing this thing means not doing three other things—those three things may also be very valuable.

A scheme may be superior in every technical item and still be rejected. The reason may simply be that the company does not have that organizational capability, or the timing is wrong.

This is not contempt for technology. This is a more cruel thing: technical judgment must land within commercial constraints, otherwise it is only possibility.

Second, some decisions have no correct answer, only trade-offs.

Problems in the laboratory usually have solutions—possibly a unique solution, possibly a set of feasible solutions. But many problems before a CEO have no solution at all, only two choices with different costs.

To transform a loss-making business one must at some point cut some people. To protect the R&D budget one must cut spending elsewhere. To win a customer one must accept lower gross margin. These decisions cannot be calculated by data into a correct answer; one can only judge which cost is more bearable.

This recognition is subversive for an engineer. It means that part of the content of leadership is bearing the weight of making damaging decisions.

Third, the work of explanation is as important as making decisions.

The IBM under Gerstner was in a long and painful transformation—turning a hardware-driven company toward services and solutions. This kind of transformation cannot be completed by one technical breakthrough; it requires tens of thousands of people moving in the same direction for several years.

This is a skill she had completely never practiced before. She never needed to explain her experiments; her advisor could read the data. But from this year onward she began to understand: in an organization, judgment and expressive ability are two sides of the same thing. A judgment that cannot be clearly explained cannot become the common action of dozens of people.

 

She Later Rarely Spoke of the Specific Content of That Year

One thing is worth noting.

Records of that year are few. She has mentioned many times that it changed her way of thinking, yet rarely recounts specific conversations in the office, complete sentences Gerstner said to her, or the specific process of a certain decision.

The only relatively widely circulated segment is her confusion when she first received the appointment: why? Why would he want to talk to me? That was a very strange thing.

Her understanding at the time was that the company wanted to find a person closer to new technology, because Gerstner himself was not of technical background, and this company was a technology company.

There is a very valuable detail in this recollection: she initially did not know why she had obtained this opportunity.

This sentence is worth placing in the middle of this chapter. Because it is one of the keys to understanding her subsequent twenty years.

She has never been a person who “designs her own career path.” She did not have a plan that said “become an executive before age 35.” What she did was: whenever a more difficult position with a broader scope of responsibility appeared, judge whether it was worth trying. Gerstner’s office was like this, Freescale’s business was like this, AMD was like this.

This is a very plain career strategy, but it has one prerequisite—you must be willing to walk in, even if you are unclear what you will face.

 

The Translator Became Something Else

After that year ended she returned to the technical department, but the position was already different.

She began to bear broader business responsibility, gradually shifting from researcher to product and organizational leader. By her later division, her career was roughly divided into two segments: the first twenty years or so on the R&D side, the subsequent ten-plus years on the business side. This turning point was in the year of Gerstner’s office.

And the role of “translator” also grew larger. Previously she had translated the language between one technology and a set of customer needs. Thereafter what she had to translate was a larger range of things: the language between a budget and a direction that could only be verified three years later.

At the same time another thought began to emerge.

She increasingly wanted to be the boss of a company.

The logic of the production of this wish is very clear: in a company like IBM her role was always working within certain boundaries. Budget was set by others, priority order was ranked by others, the long-term projects to be protected were decided by others’ quarterly pressure. No matter how accurate her judgment, it could only function within the grid drawn by others.

And she was increasingly clear about one thing: deciding which thing to do affects the result more than deciding how to do it well.

 

A Large Company’s Resources May Also Be Walls

Here a common misunderstanding needs to be clarified.

IBM’s money, laboratories, and customer channels are real advantages, but priority was not in her hands.

So she began to look at opportunities outside. In 2007 she left IBM. The official narrative of this decision is very concise: she had been at IBM thirteen years, from research all the way to vice president of the semiconductor R&D center, in charge of the strategic direction of IBM silicon technology, joint development alliances, and semiconductor R&D operations.

Thirteen years. She placed her entire twenties and early thirties in this company. When she left, she was no longer an engineer.

 

What She Took Away

Reviewing the thirteen years at IBM, she took three things away from there.

The first is technical judgment. She did enough depth of work: process, devices, architecture, R&D management. Her technical authority was not borrowed by title; it was earned by more than twenty papers and a technology project that rewrote the industry.

The second is cross-boundary language ability. She could understand what materials scientists were saying and also understand what Gerstner was asking. This two-way ability was not common in the technical circle at the time. Most people’s path is choose one or the other: either become a deep expert, or switch to management as early as possible. She experienced both.

The third is understanding of “difficult problems.” An executive at IBM who worked with her gave her a sentence to the effect: do not go around the problem; run toward the most difficult problem. Not walk, run. Because there one learns the most, grows the fastest, and most easily pulls away from others.

Regarding the specific source of this sentence, more careful verification is needed later. But its influence truly exists: the series of choices she later made at AMD—abandoning the low-end market, rewriting the CPU architecture, betting on a not-yet-proven AI-workload direction—are all concrete executions of this sentence.

Run toward the problem, not walk.

There is one more thing she took away, harder to discover than the first three: she learned how to protect her own judgment inside a large organization.

At IBM she could not decide priority, but within her own range she could do one thing: make a small direction good enough to prove it was worth investing more resources. This action later became the most stable behavioral pattern in her life—using delivery, rather than persuasion, to exchange for larger resources.

 

 

Part IV · Succession Planning and the Leadership Test

 

Chapter 7 · The First Time Responsible for Profit and Loss

In 2007 Lisa Su left IBM.

Thirteen years. She placed her entire twenties and early thirties in that company—from process research in the laboratory all the way to vice president of the semiconductor R&D center, in charge of the strategic direction of IBM silicon technology, joint development alliances, and R&D operations.

In this company her judgment was sufficient; her position was not. Budget was set by others, priority order was ranked by others; no matter how clearly she knew which direction was worth doing, she could only function within the grid drawn by others. And beginning this year she would change to another way of living: being responsible for a real number.

 

A Technical Officer Parachuting In

She entered Freescale Semiconductor; the first stop was chief technology officer.

The origin of Freescale is worth mentioning. It was originally Motorola’s semiconductor division, spun out around 2004, and in 2006 underwent a huge leveraged buyout, becoming one of the largest privatization deals in the semiconductor industry at the time. This meant: it carried heavy debt, a miscellaneous product line, and an organizational structure that needed to quickly prove its own value.

The chief-technology-officer position was a very natural landing point for someone who had just come out of IBM. Her technical credentials were sufficient to support this title, and the “translator” ability she had accumulated before exactly corresponded to the core requirement of this post: judging where technical resources should be invested, and making the business department accept this judgment.

But chief technology officer is not a position responsible for results. It can decide the roadmap, but does not necessarily control the people and money that deliver the roadmap.

So she soon took another step forward.

 

One Billion Dollars, and a Chaotic Department

Her later post was senior vice president and general manager, responsible for networking and multimedia business—that is, the embedded communications and application-processor segment. This was a business of roughly one billion dollars in scale.

This was the first true business leadership in her career. This position has three fundamental differences from chief technology officer.

First, she was responsible for revenue, not only for technical direction. If the roadmap was chosen correctly but could not be sold, it was her problem; if product delivery was delayed, it was her problem; if gross margin was squeezed, it was also her problem. Technical judgment could no longer be evaluated independently of commercial results.

Second, she had to decide what not to do. In a department of one-billion-dollar scale there are always dozens of directions that can be invested in. Resources are limited, which means she must personally end some projects and turn some people elsewhere. This is something she never did in the ten years at IBM—she had protected directions, but never killed directions.

Third, she had to face a “chaotic company.” Freescale’s situation at the time, by the general description inside the semiconductor industry then, was: product lines scattered, decisions slow, internal trust in the roadmap low. This was not caused by her, but she had to make choices inside it.

 

What She Did

Public first-hand records of the Freescale period are few. There are no detailed operating data, no disclosed list of her specific trade-offs, and no complete recollections of the parties. So here one must write conservatively.

What can be confirmed is the result: the department she managed was originally in a chaotic state lacking clear priority; she sorted it into an organization that could be operated, planned, and delivered externally. There are reports that she later “successfully pushed this company toward listing”—this expression needs to be unpacked.

Freescale did re-list on the New York Stock Exchange in 2011. This was a company-level major event involving overall financial restructuring, debt management, and market return. Attributing it solely to Lisa Su is inaccurate; but the business she was responsible for was one of the parts that needed to produce results in the restructuring narrative.

The second thing that can be confirmed is that her management style took shape here.

At IBM she had already formed the core action: shrink the direction to a verifiable range, first prove it feasible, then fight for more resources. At Freescale this action was used for the first time on real business—no one will hand one billion dollars to a direction because you speak movingly; you can only take out what was done last year and what can be done this year.

This action thereafter became her fixed pattern, all the way to AMD.

 

A Change She Did Not Perceive at the Time

The two-plus years at Freescale, besides business experience, also gave her something more concrete: what the inside of a semiconductor company looks like when it is about to be unable to hold.

At this stage she saw how resource tightness distorts judgment—which types of projects are cut first, which types of people leave first, how the supply chain tightens, and when customers begin to consider a second supplier. She also saw what meetings become when an organization is on a downward path: what people report is not problems, but reasons.

These observations would be put to direct use at AMD two years later. Because what she saw at AMD and what she saw at Freescale in 2009 were the same kind of illness.

But at the time she had not yet organized these observations into a system. For her Freescale was more like a successful internship—she proved she could do things in real business, and then it was natural to continue upward and manage larger business.

At that time she was planning exactly this way.

 

A Phone Call from Austin

Around 2011 she received a phone call from someone.

And AMD’s situation at the time was complete chaos.

Regarding this dinner, public materials have quite vivid narrative: taking the opportunity of a meeting in Austin he invited her to dinner; before arriving both were anxious—he worried she still had a knot about leaving IBM years earlier, she worried he blamed her for jumping ship. To ease the atmosphere he ordered a very expensive California Cabernet. Once the wine was opened the conversation also opened, and the two were pleasantly surprised to discover that neither held a grudge.

These details are literary, but their source is later circulating narrative, not sentence-by-sentence verifiable first-hand record. This book retains them here while necessarily stating their nature: this is a recollection, not dialogue that can be restored. No serious biography should treat descriptions such as “he leaned forward and said word by word” as historical fact—that belongs to the writing of novels.

What can be determined is the substance of that conversation. He told her: AMD has a huge pile of problems, all of them illness at the root; but it has a group of the most top-tier engineers and intellectual property that others do not have. And this position is too suitable for her.

 

Why This Is a Dangerous Opportunity

Now the nature of this matter must be made clear.

From the perspective of career risk, joining AMD was an extremely unprofitable choice.

Her position at the time was good: the Freescale business she managed was smooth, the semiconductor industry was warming, and she could completely obtain a higher position, a more respectable résumé, and less mental wear at another healthy company. And what kind of place was AMD?

Around 2011 AMD carried roughly $2.5 billion in debt; in the past five years it had achieved revenue growth in only one year; the PC market was being squeezed by tablets and phones; server share had fallen below 1%; share in the personal-computer chip market had fallen from 23% to under 10%; what analysts publicly discussed was not how much it would grow, but which day it would go bankrupt. The company had changed four CEOs in the preceding years.

No rational professional manager would regard such a company as an opportunity.

So in understanding her choice one must switch to another logic.

What she valued was not how much this company was worth now, but the scale of this problem. The behavioral pattern of her life appeared once again here: the basis for judging a position is not how safe it is, but how large a range of judgment it allows. At Freescale she could manage a one-billion-dollar business, but those directions had already been set by others. At AMD everything was collapsing, and what collapse provides is precisely—nothing is unchangeable.

And her specific post at the time was not CEO. In January 2012 she joined AMD, responsible for key business. This itself was the first step of a succession arrangement.

 

Chapter 8 · Two Moves

In January 2012 she entered AMD.

To understand her position this year, one must first understand AMD’s position at the time—and why it needed a new face.

 

A CEO Who Arrived One Year Earlier

She was not the most important personnel change at AMD in 2012.

In August 2011 the AMD board brought in a new CEO: Rory Read. He had previously served as president and chief operating officer at Lenovo, participating in the integration after Lenovo acquired IBM’s personal-computer business, and was known in the industry for operations and cost management.

This choice shows what the board was thinking. A company that lost money every year, whose cash was continuously draining, and whose products could not beat the competition, the first thing to solve was not the product problem, but “do not die before the next product comes out.” So an operations-type person who could control costs and hold cash was needed.

Rory Read’s task can be summarized in one sentence: first keep this company from immediately dying financially.

And Lisa Su’s responsibility after entering AMD was exactly the other matter. What she was responsible for was product and business—that is, the “blood-making” half.

The board was playing two moves. One was stopping the bleeding, one was making blood. When she came in, the company was already walking the first step.

 

What She Saw on the First Day

Here one must carefully distinguish: which views she already had at the time, and which were narrated later.

In public interviews she rarely describes in detail any “epiphany-style judgment” of the technical route when she first entered AMD. The most widely circulated statement is: she believed AMD could not beat Intel not because of poor execution, but because the CPU architecture itself was bad—the previous people’s thinking was “optimize a bit more, cut price a bit more, squeeze out another 5% performance,” while her judgment was that the foundation was crooked; adding construction on top was meaningless.

This judgment does exist; it is the key to understanding all her later decisions. But writing it as a flash of insight on the day she joined is inaccurate.

A restoration more consistent with reality is this: when she entered AMD the company had a group of very capable engineers, a batch of intellectual property not available elsewhere, a new architecture under development (later called Zen), and was also continuing to sell products based on the old architecture. The competitive method of the old product line was to fight for share by continuous optimization and lower price. This path still looked walkable in 2012, only more and more laborious.

It did not require extraordinary insight to see that this path could not be walked. What was hard was admitting it could not be walked, and paying the price for it.

 

Friction Cost: A Company’s Most Expensive Expenditure

If one word is to be used to describe AMD’s technical problem at the time, the most accurate is roughly friction cost.

This is not a financial term, but a real form of loss that exists in engineering organizations: resources consumed by the entire organization on repetitive labor and self-doubt because direction is unclear, priority is unstable, and commitments cannot be fulfilled.

It manifests in several concrete aspects.

First, overlap of product lines. If a resource-tight company simultaneously maintains too many directions, each direction does not receive sufficient investment, and the result is that each direction is done poorly. At the time AMD was exerting force simultaneously in multiple markets, multiple price bands, and multiple product lines; it looked like “comprehensive competition,” but was actually “weak everywhere.”

Third, internal organizational doubt. Engineers are smart people. They know which projects are destined not to be done, and also know which projects have been cut and then picked up again repeatedly. When a company is long-term in a cycle of “announce—cut—re-announce,” the most capable batch of people will leave first.

These three things added together are the true lesion of AMD around 2011. It is not reflected in some number on the quarterly report, but permeates every cross-department meeting, every roadmap review, every customer visit. It will not make the company die within one quarter, but will make it unable to recover for ten years.

And its antidote is not “try harder,” but reduce the number of things being done simultaneously, then actually get the remaining things done.

 

Things She Needed to Confirm with Her Own Eyes After Entering AMD

The first thing she needed to confirm was “which capabilities are real.”

In public statements the former IBM executive who brought her to AMD said very directly: AMD has a huge pile of problems, “all illness at the root,” but it has a group of the most top-tier engineers and intellectual property that others do not have. The first half of this sentence is diagnosis; the second half is an asset list.

A company on the brink usually has one thing in common: its balance sheet has an asset that cannot be found on the report—the knowledge of its engineers of how to make a CPU. Once this kind of capability is disbanded, rebuilding takes ten years. And this was precisely the only truly valuable thing AMD had at the time.

The second thing she needed to confirm was “which path is truly walkable.”

The console business was obvious. Sony and Microsoft’s game machines each generation need a custom chip, and AMD can provide both CPU and GPU—this is its structural advantage relative to other suppliers. The gross margin of this business is not high, but it has definite customers, definite cycles, and definite cash flow.

 

Succession Is Not a One-Time Event

Regarding how she became CEO, a common narrative error needs to be clarified.

The circulating version is this: she joined AMD, worked two years, the company got worse and worse, and in October 2014 the board decided to let her succeed. It looks like a crisis appointment.

The actual situation is closer to a staged succession arrangement.

When she entered AMD in January 2012 she was introduced as an important business leader—this was no secret in semiconductor circles. Over the next two-plus years the company simultaneously experienced several things: at the operations level costs were being cut and structure reorganized; at the product level a new generation of architecture was being advanced; at the management level the original division of labor was gradually adjusted.

In October 2014 she became president and CEO. This arrangement was not a punitive crisis appointment, but a plan already in operation reaching its next step.

This also explains why her actions after taking office were so continuous, rather than like a person who had just parachuted in and needed a year to familiarize: many of the judgments she had to make had already formed between 2012 and 2014.

 

Places Where She Must Thank the Previous Incumbent

There is also an angle that is rarely written.

When Lisa Su took over AMD, the company had already completed several important preparatory works. The manufacturing business had been spun off as GlobalFoundries as early as 2009; the company no longer owned its own advanced wafer fabs—this was both a historical burden and unexpectedly cleared space for later “traveling light and betting advanced products on TSMC.” The reconstruction of the new-generation CPU core had already been started by the engineering team. The company had gone through several rounds of painful cost cutting; although it had wounded vitality, it had indeed pressed down cash consumption.

Whether these things were done well enough can be debated; but their existence means that when she took over she had a usable set of cards.

If in 2014 AMD had faced a completely blank table—no team, no architecture under development, no console business, no cash—then even the strongest leader could only do liquidation.

Her luck was that the thing she had to save still existed; her ability was that after recognizing this she did not waste it on even application of force.

 

Chapter 9 · The Decisive Test

The way Lisa Su had to prove she could take this seat was not by publishing strategy, but by taking on a very concrete question:

Make two customers who are mortal enemies both agree to use chips designed by the same company.

This question was placed before her around 2012. Its name was semi-custom business.

 

Why a Chip Question Would Be Written into Business History

Before writing about this matter, its weight must first be made clear.

Game consoles are different from ordinary computers. After a console launches, it usually sells for five to seven years. This means the chip is locked two or three years before the console launches and cannot be replaced in between. Console makers cannot gamble wrong, because if they gamble wrong the entire platform collapses: insufficient performance will be outdone by competitors, power consumption exceeding standard will cause heat-dissipation and volume problems, costs that cannot be held will directly eat the already-not-high hardware profit, and unstable supply means missing the launch peak season.

The standards by which console customers select chip suppliers are completely different from PC customers. PC customers look at the performance and price of this generation of product; console customers look at “whether you can accompany me through these seven years.”

And the greatest difficulty at the time was: the two console makers are direct competitors, yet the supplier is the same company.

Sony and Microsoft compete for the same batch of users, the same batch of exclusive games, the same batch of shelf positions in the game-console market. They cannot accept their chip schemes being known by the other side, and still less accept the two sides using the same chip—that would equal hardware differentiation being wiped out, and competition remaining only game content.

So the real difficulty of this question is not technology, but information and trust.

 

What She Had to Do

Public materials record very little of this negotiation process. What can be confirmed is its basic structure.

AMD’s structural advantage in this business is: it is the only company that can simultaneously provide high-performance CPU and high-performance GPU. Intel’s graphics capability is weak; Nvidia has no x86. Consoles need to put both into one chip and make an SoC—this way both cost and power consumption can be controlled. This combination advantage is not absolute, but at the time there was almost no alternative.

Her work roughly contained three parts.

The second part was making both customers feel they were getting a custom scheme rather than a generic part. This required isolation at the design level, rigorous arrangements at the intellectual-property level, and strict separation of information for the two customers throughout the development cycle. This is the part that consumes the most trust cost—she needed to simultaneously prove to both customers: the things you tell me will not flow to your opponent.

The third part was giving certainty. At the time AMD’s financial situation was not good. Console makers betting their next-generation platform had to confirm that this supplier could survive seven years later. This matter she could not solve with promises; she could only use other things—for example what this business itself could bring, and the resource guarantee for this business line inside the company.

Regarding whether she personally flew to Japan and Seattle, whether she drew architecture diagrams on whiteboards, and what specific exchanges occurred during the negotiations, public first-hand records are insufficient. This book handles this place conservatively: confirm the nature of the matter and her role, do not describe unverified scenes.

 

Two Results from One Strategy

The result can be confirmed: AMD simultaneously won the custom chips for Sony PlayStation 4 and Microsoft Xbox One.

Both consoles launched in 2013. They were highly similar in architecture—both used AMD’s x86 CPU and Radeon graphics cores—differences mainly coming from frequency, memory configuration, and system design. This also confirms the solution to that difficult problem: not making completely different chips for the two sides, but making both sides able to accept “we are using the same generation of technology platform, but each has its own scheme.”

The significance of this business to AMD has three layers.

At the revenue level it provided a definite, stable, multi-year cash flow.

At the confidence level it proved that AMD’s engineering capability was still recognized by world-class customers—when a company was regarded by Wall Street as near death, the weight of this evidence is hard to overestimate.

At the strategic level it placed AMD in a position that is hard for others to replace. Although Intel and Nvidia are respectively stronger in CPU and GPU, neither can provide a complete SoC solution. This position has been retained later.

 

Cash Flow Is Not Profit, but It Is Life

Here one thing must be made clear.

But in a cash-tight company, “good gross margin” and “able to save life” are not the same thing.

In AMD’s most difficult period in 2015, the semi-custom business provided stable cash contribution and customer relationships. It was classified in the financial reports under the “Enterprise, Embedded and Semi-Custom” segment; in the second quarter of 2015 this segment’s revenue grew 13% sequentially, mainly driven by semi-custom SoC sales; by the fourth quarter of 2015 this segment’s revenue was $488 million, of which semi-custom shipments set an annual record.

These numbers are not pretty, but they are important. They mean: in the years before Zen brought revenue, the company had a definite blood-transfusion line.

Long-termism here first revealed its true shape: it cannot be maintained only by determination; there must be a business that pays the bill for it.

 

Why This Question Was Her Final Exam

Now return to the question at the beginning of this chapter: why did the board hand this question to her?

Because it simultaneously tested three things, and these three things are exactly what a semiconductor-company CEO must possess.

First, technical judgment. Can she understand where the boundary of an SoC is, can she say to the customer “this power-consumption target can be achieved” or “this cannot be done.”

Second, customer relationships. Can she simultaneously obtain the trust of two competing customers, and under the premise of information isolation make both feel valued.

None of these three things belongs to pure technical work. They all occur at the boundary of technology, commerce, and organization—exactly the set of skills she trained in the thirteen years at IBM, from Gerstner’s office and the Emerging Products group.

She passed this exam. Two years later, in October 2014, she became AMD president and CEO.

 

Chapter 10 · AMD’s First Woman CEO

On October 8, 2014, Lisa Su’s employment agreement with AMD took effect.

The appointment announcement wrote president and chief executive officer. When the media reported it they added a sentence: she is the first female leader in AMD’s forty-five-year history.

This identity is real, and also meaningful. But what this chapter must handle is a colder question: symbolic meaning and operating reality are not the same thing.

A company will not automatically improve because of the gender of its leader.

 

The Books She Took Over

First look at AMD’s actual situation when she took over.

Market capitalization: less than $2 billion. Before this the stock price had long hovered between $2 and $3; in July 2015 it fell further to $1.61. The label Wall Street gave AMD at the time was “uninvestable”; what analysts discussed was not how much it would rise, but which day it would go bankrupt.

Debt: about $2.2 billion.

Market: PC-processor share had fallen from a peak of 23% to below 10%; server share had fallen below 1%, almost equal to disappearing from this market. The meaning of these two numbers is more serious than it looks—server business is high-margin business; its disappearance means AMD lost the ability to supply its own blood.

Industry environment: traditional PC sales fell from more than 360 million units in 2011 to under 280 million units in 2015. Even more fatal was that the real growth occurred in phones and tablets, and this revolution AMD completely missed. Qualcomm, Apple’s in-house development, and MediaTek had long divided it up; no position was left.

Organization: it had changed four CEOs in the preceding years. This number itself explains the problem—every change was accompanied by route adjustment, team reorganization, and reset of commitments. By 2014 the internal trust in long-term plans was necessarily low.

These four sets of numbers together constitute a very special situation: what she took over was not only a company with a bad financial situation, but a company that no longer believed it could turn around.

 

The First Thing She Did Was Not Issue Orders

Most people taking over such a mess have the first reaction of cutting. Cut costs, sell assets, protect cash flow, hold on one more day if possible, drag R&D to later. The several previous incumbents at AMD basically did this.

Her first step was not this.

Her first step was diagnosis.

This action sounds plain, but it was actually a public negation of the previous route. The previous people’s thinking can be summarized as: “Our products are weaker than Intel’s, so we must optimize, must cut price, must squeeze out another 5% performance, must hold share.” This set of thinking was still running in 2014. It is not absurd—under extremely limited resources, improving existing products is the least-effort way of survival.

Her conclusion was: this path cannot be walked, not because of poor execution, but because the CPU architecture itself has a problem.

 

“The Foundation Is Crooked”

This judgment needs to be unfolded, because it is the starting point for understanding all of AMD’s decisions in the subsequent ten years.

A CPU architecture refers to a set of foundational design: how instructions are fetched, decoded, executed, and written back; how caches are layered; how the pipeline is arranged; how multi-core cooperates. It determines what this chip can theoretically achieve. Every subsequent improvement—raising frequency, adjusting cache, optimizing scheduling—can only be carried out below the ceiling given by the architecture.

If the architecture itself is backward, then “optimize another 5%” is meaningful, but it cannot exchange for a true competitive position. You can do the best within the given framework, then discover that the opponent has changed to a better framework, and all your local improvements immediately return to zero.

This is the meaning of “the foundation is crooked”: adding floors to a building whose foundation is crooked, no matter how carefully added, cannot solve the problem of the building itself.

And admitting this has an extremely high cost.

Because tearing down and rebuilding means: three to four years of time, several billion dollars of R&D investment, while the company simply does not have that much money on the books, nor that long a patience that can be borrowed. It means that during these three to four years AMD must continue to fight the market with backward old products—continue to take a beating, continue to lose share, continue to be mocked by analysts. It means management must admit to the board, customers, and employees: what is being sold now is no good; we are remaking it; please wait until 2017.

This requires a rarely seen organizational capability: enduring long-term publicly acknowledged poor performance in exchange for a product that does not yet exist.

 

Her Solution: Not Rebuild Everything, but Rebuild the Core

Here a common misunderstanding needs to be corrected.

When she said tear down and start over, it did not mean dismantling the entire company and remaking it. This is impossible.

What she did was limit the scope of reconstruction to the layer that truly decides success or failure: the CPU core architecture. Everything around this core—graphics, console semi-custom, chipsets, channel relationships, existing customers—must be retained, because they are either making money or maintaining the company’s existence.

What is rebuilt is the engine, not the entire vehicle.

This distinction is extremely important. It turns “tear down and start over” from a romantic slogan into an executable resource arrangement: one line rebuilds, the remaining lines protect cash flow. Which line is responsible for the future, which line is responsible for living, must be clearly distinguished, otherwise both sides will fail.

 

She Took Three Actions

Regarding the strategy after she took office, it can be summarized as three concentrations. The timing of these three actions need not all be precise to a certain month, but the direction is clear.

The first is market concentration.

This decision will immediately appear in the financials as revenue decline. Actively abandoning revenue is one of the earliest, and most unwelcome, decisions she made.

The second is architecture concentration.

The nature of this decision is: use today’s loss to buy a possibility three years later.

The third is partner concentration.

Advanced manufacturing must rely on external wafer fabs. AMD had already completed the spin-off of the manufacturing business in 2009, becoming a fabless design company. What needed to be solved next was making the architecture route coordinate with process nodes—betting the most competitive products on partners that can provide leading process.

This action is later often simplified as “Lisa Su turned to TSMC.” This statement is inaccurate. The situation when she took over was: the company no longer owned its own advanced wafer fabs, and was constrained by existing wafer-procurement arrangements. What she had to do was not “pioneer” outsourced manufacturing, but gradually bet the most advanced products on partners that could truly provide leading process, and accept the accompanying cost and contractual price.

These three actions together constitute a very plain strategy: reduce the number of things being done simultaneously, and actually get the remaining things done.

 

The Moment of Responsibility Transfer

Return to the beginning of this chapter.

Regarding the moment she became CEO, the easiest version to write is “appointed in crisis, with high hopes.” But a more accurate description is: that was not a moment of victory, but the formal handover of a responsibility.

At the time she had no good news she could announce. Market capitalization under $2 billion, stock price at a low, debt pressing, server business almost empty, the only thing on the product roadmap worth looking forward to still three years away. She could not even promise “next year will improve,” because the most honest judgment was: before Zen comes out, the company will look even worse.

Taking office at such a time requires not confidence, but an ability to coexist with bad news for a long time.

From this perspective, the identity of “the first female leader in AMD’s forty-five-year history,” besides its historical significance, also has a more practical layer of meaning: in an industry dominated by men, that believes in short-term performance, and that takes rankings and share as heroes, she was the person willing to publicly announce “we will look bad for three years.”

 

 

Part V · $1.61

 

Chapter 11 · $1.61

On July 16, 2015, AMD announced second-quarter results.

The wording of the financial report is standard corporate style, but the numbers inside are the appearance of a company approaching the endpoint.

 

A Financial Report That Needs to Be Read Three Times

First list the key numbers.

Revenue $942 million. The same period the previous year was $1.441 billion. One-third less.

Gross margin 25%. The same period the previous year was 35%. Down ten percentage points.

Operating loss $137 million, net loss $181 million.

The core Computing and Graphics segment had only $379 million in revenue, with a single-quarter operating loss of $147 million.

And the most critical number may be this one: R&D expense $235 million.

The proportion of R&D expense to that quarter’s revenue was close to one-quarter.

In other words, this company was losing money on one side and investing close to one-quarter of its revenue into R&D on the other.

 

Cash Boundary

Modern companies rarely die of losses; they usually die of cash rupture.

Loss is an accounting concept; a loss-making company can still live as long as cash is still flowing—customers pay, suppliers deliver, employees are paid, the daily cycle continues. What truly kills a company is the interruption of the cash-flow cycle: no money left on the books, suppliers demand payment in advance, customers begin to worry about delivery, employees begin to look for the next job, and the cycle rapidly deteriorates.

So to look at AMD’s situation at the time, one cannot look only at the income statement; one must look at the balance sheet.

By the end of 2015, AMD’s cash, cash equivalents, and marketable securities totaled about $785 million. The same period the previous year was about $1 billion.

Total debt about $2.26 billion.

Full-year data: revenue $3.99 billion (down 28% from $5.51 billion in 2014); full-year net loss $660 million; gross margin fell from 33% to 27%.

Putting these two sets of numbers together, the picture is clear: nearly $800 million cash on hand, nearly $2.3 billion owed, annual loss of more than $600 million.

 

Every Cut Is a Gamble

In such a situation every management decision becomes a variant of the same question: survive today, or live well three years later?

Talking only about vision without managing cash may mean being unable to hold until Zen is completed.

Leadership here is not choosing a pretty answer, but arranging the rhythm within survival constraints.

Rhythm specifically refers to: which expenditures can be cut, and to what degree; which must be protected, and with what reasons explained to the board; which businesses can be sold, and what impact the sale will have on the future; which collaborations can be negotiated, and what price must be paid for the negotiation.

Most of these decisions are not public, and none can be proven “correct.” They can only be evaluated afterward.

 

The Role of Semi-Custom

In this financial report there is one bright spot.

The Enterprise, Embedded and Semi-Custom segment’s revenue grew 13% sequentially in the second quarter, mainly driven by semi-custom SoC sales. By the fourth quarter the segment’s revenue was $488 million; full-year semi-custom shipments set a record.

These numbers have a characteristic: they are not pretty, but they are definite.

The gross margin of the semi-custom business is not high; console makers are extremely sensitive to cost; the chips themselves are sold close to cost. But the volume is large, the customers are stable, and the cycle spans multiple years.

So in AMD’s most difficult years it bore the role not of “growth engine,” but of blood-transfusion vessel.

This is not a romantic role. It cannot change the company’s technical backwardness, nor reverse the stock price. Its function is to keep cash moving on the books during the three years of waiting for Zen. The company did not have to talk about a humiliating financing or sale in 2016 precisely because every quarter this business was paying the bill.

 

The Words She Said in the Financial Report

This time her words are worth looking at separately.

In the second-quarter 2015 financial-report statement her wording was roughly this: the sequential growth of the Enterprise, Embedded and Semi-Custom segment and channel business was insufficient to offset the impact on the PC-processor business from demand lower than expected. The company continues to execute the long-term strategy while responding to the current market environment.

In the fourth-quarter statement she said a bit more: 2015 was financially very difficult, but key R&D investment and focus on innovation placed the company in a better position to deliver better products in 2016, improve financial performance, and gain share.

These two paragraphs have one common characteristic: no grand words, no timetable promising improvement, not even attempting to package bad news as good news.

At a time when the stock price had fallen to $1.61 and analysts were publicly discussing bankruptcy, this language choice is intentional. She does not need to sell confidence to the market. What she needs is to make four groups of people simultaneously believe the company has a tomorrow: customers believe the roadmap will be fulfilled, suppliers believe bills will be paid, employees believe staying has meaning, investors believe R&D investment is not burning money into a black hole.

Near the cash boundary, restraint itself is a management tool. The organization needs to know the problem is serious, while still believing the next step can be completed—these two things must hold simultaneously.

 

Who Was Consumed in These Three Years

This chapter also needs to write a part that is usually not written into corporate biographies.

In those three years the company did many things in order to survive that it later did not want to talk about much.

It cut people. Reducing operating expenses means fewer positions; this is something that happens in every structural reorganization, and also the concrete content represented by “restructuring charges” in every financial report.

It terminated projects. Reducing the number of things being done simultaneously necessarily means cutting some directions, and on the cut directions there are usually people who have invested several years.

It cut expenditures. From travel to equipment, from recruitment to outsourcing, every item has a concrete cost.

It signed agreements with long-term consequences. Among them the technology-licensing arrangement with Chinese partners—that money was especially important in 2016, and therefore buried the controversy that later continued to expand.

Without adding these things, the book numbers of this chapter are only numbers. And they originally mean that a batch of people used their own positions to exchange time for the company.

 

Chapter 12 · What She Inherited, What She Changed

This chapter must do something that is rarely done seriously in biographies: divide the things on her into three shares.

One share is what others had already done and she took over; one share is the judgments she herself made; one share is the consequences she must bear. These three things are often mixed together in writing, with the result that she is either written as a savior or written as a lucky executor.

After separating them clearly, one discovers a more interesting thing: what she truly did at AMD was not creating something, but choosing what to protect.

 

The First Share: What She Inherited

First talk about the things that already existed when she took over. This part must be written clearly, otherwise the credibility of the entire book will be discounted.

The architecture of external manufacturing. AMD’s manufacturing business had been spun off as GlobalFoundries as early as 2009. That is to say, when she took office in 2014 the company no longer owned its own advanced wafer fabs.

This point is very important, because the vast majority of popular narratives about her contain an erroneous statement: “Lisa Su sold AMD’s wafer fabs and turned to TSMC.” She did not sell the wafer fabs. This matter occurred three years before she joined AMD, completed by the management of the time.

The CPU core under reconstruction. Zen was not a project started from zero after 2014. The engineering team and architecture leads had already begun reconstructing the CPU core—senior architect Jim Keller returned to AMD at this stage and participated in this work. What Lisa Su took over was an architecture reconstruction that had already been started, but had not yet been proven, and had not yet obtained sufficient resource guarantee.

x86 intellectual property and customer relationships. AMD holds the x86 architecture license, possesses a complete microprocessor intellectual-property system, and retains customer channels in the PC market—although share was shrinking.

Console business. Sony and Microsoft’s semi-custom chips were already shipping. The gross margin of this business is not high, but the cycle is long and the cash is definite.

Heavy debt and historical burden. About $2.2 billion in debt, consecutive years of losses, and an organization that had experienced multiple reorganizations and leadership changes and lacked trust in long-term commitments.

The shape of these six things added together is: a company with real technical assets, with definite cash flow, but without clear priority, and carrying historical debt.

 

The Second Share: What She Changed

Now one can say what she truly did.

First thing: shrink the scope of reconstruction to one place.

The statement “tear down and start over” is easily misleading. She did not, and could not, dismantle the entire company and remake it. What she did was: concentrate reconstruction on the layer that truly decides success or failure—the CPU core architecture—while letting all other business lines remain, because they are either making money or maintaining the company’s existence.

Second thing: make clear “what must be protected.”

This is the most essential difference between a CEO and a CTO. The technology officer decides “which technical route is superior”; the CEO decides “which few things absolutely cannot fail.”

In her version those few things were roughly: the reconstruction cycle of Zen, the continuity of the console business, the long-term investment in the data center. None of these three produce short-term returns—the price of protecting them is letting the company continue to look bad for a not-short period of time.

Third thing: admit that some things are already impossible to achieve.

A layer not often mentioned: after taking office she also had to clearly state which goals should be abandoned. Low-end market share was abandoned, the method of price war was not fought, some existing product lines were no longer invested in.

The decision to actively abandon revenue is usually the most unwelcome management behavior. It will be complained about by customers, resisted by the sales team, and used by analysts to question management’s ability. But without this step, “concentration” is only a slogan.

Fourth thing: straighten the organization’s attitude toward time.

This one is the least conspicuous, and may be the most important.

Short-termism is usually not understood as greed; more often it manifests as fear. Because of worry that the next quarter will not look good, every project is shaved a bit of budget—R&D a bit less, verification a bit faster, sample tape-out pushed a bit, testing shortened a bit. Each action looked at alone is reasonable; the result added together is: no project is able to cross the competitive threshold.

One of the things she had to do was make the organization believe: “We are looking bad, but this is part of the plan, not a signal of failure.”

This sentence spoken to shareholders is useless; it must be spoken internally—spoken to those engineers who, before Zen is made, every day face a backward product, a suppressed share, and an ugly quarterly report.

 

The Third Share: The Consequences She Must Bear

The first two shares look like bonus items; the third is not.

She must bear three years of looking bad. This is not rhetoric. In the years from her taking office until Zen launched, the company’s financial performance continued to be very poor, the stock price long stayed at a low, and external questioning of management never stopped. Anyone who chooses this path must accept a reality: during these three or four years, all evidence that can publicly prove she did the right thing does not exist.

She must bear some unclean trade-offs. Among them is that 2016 technology-licensing deal with Chinese partners. The several hundred million dollars it brought were very important that year and also gave the company an additional path to survive during that period. At the same time it also buried the controversy about technology outflow and national security that later continued to expand. A commercial decision made in order to survive may only be fully evaluated ten years later.

She must bear uncertainty. The directions she bet on—data center, AI workloads, open ecosystem—were none of them judgments that could be verified at the time. For them to hold requires external conditions to cooperate: process partners delivering according to plan, customers willing to change purchasing habits, competitors making mistakes, the market moving in a certain direction.

She cannot control these conditions. What she can do is, when the window opens, make the company already prepared.

 

Then What Did She Truly Do Right

If the three shares above are combined, her contribution can be compressed into one sentence:

On a piece of real assets she made the correct trade-offs—protected the few things that must succeed, abandoned some things that looked important but were not worth it, and bore the looking-bad of the intervening years.

This sentence does not sound as moving as “turning the tide,” but it is closer to fact. It explains something easily overlooked: why AMD was able to revive, while many chip companies that faced similar difficulties were not.

The answer is not “they lacked a Lisa Su.” The answer is: many companies in difficulty chose even application of force, so that no single item could cross the threshold; even if someone proposed the correct trade-offs, they were usually replaced by the board within eighteen months because of performance pressure, so the trade-offs were abandoned halfway.

 

An Honesty That Must Be Written into the Book

This chapter must also leave one honest sentence.

The analysis framework above—“inherit, change, bear”—is a division made by the author for understanding this history. It is not her own statement, nor the company’s official strategy document of the year. What she can publicly say is usually only fragments of it: we will focus on a few domains, we must protect R&D, we must do the right thing for the long term.

Writing this framework too neatly will instead lose truth. Real management is not neat: some decisions are argued out in meetings, some are forced by cash flow, some are only retroactively recognized as strategy after they are made.

And the reason it is worth organizing this way is precisely because the result has already occurred—we are looking back afterward. The risk of doing so is easily producing an illusion that every step was calculated. The handling method of this book is: state the conclusion clearly, while making clear that it is the result of looking back.

 

Chapter 13 · Three Concentrations

Earlier it was said that the strategy after she took office can be summarized as three actions: market concentration, architecture concentration, partner concentration.

What this chapter must talk about is the hardest of these three actions—not choosing direction, but making the directions align with each other.

 

Concentration Is Not “Choose the Good, Abandon the Bad”

First make clear what the word “concentration” actually refers to in management.

It is easily understood as a priority ranking: give the best resources to the most important business. This understanding is not wrong, but it misses the truly difficult part.

In a resource-tight company the problem is usually not “not knowing which direction is better.” The problem is precisely that every direction looks still okay.

The low-end PC market has thin gross margin but large shipment volume; existing products are backward but still have customers; graphics business is fiercely competitive but is AMD’s traditional strength; console business has low gross margin but stable cash flow; servers have almost no share, but if successful the return is highest.

None of these five things is obviously something that should be abandoned. Each can speak a set of reasonable reasons, and each has people responsible for it, each has a set of internal support relationships.

This is the truly hard place of concentration: it is not technical judgment, but political judgment.

So “three concentrations” are not three strategic choices, but three groups of conflicts that must be borne.

 

Market Concentration: The Price of Abandoning Revenue

The content of the first concentration is investing resources in domains where computing performance truly matters—high-performance computing, data center, gaming.

Translated into actual action it is: low-end market share is no longer protected.

What does not protecting share mean? It means the sales team will see the market share they are responsible for falling; it means channels will complain there are no models that can fight a price war; it means a certain revenue item in the quarterly report will obviously shrink.

Actively abandoning revenue when finances are tightest is the most unwelcome kind among this type of decision. Because the benefit it brings is three years later; the cost is next quarter.

And it has another hidden cost: abandoning the low end will weaken the company’s presence in that market. When a brand does not appear in a certain price band for a long time, the memory of channels and users will fade, and it will be harder to return later.

So this action is not “cleverly focusing,” but using a batch of clear losses to exchange for a vague possibility.

 

Architecture Concentration: Not Cutting the Most Money-Burning Line

The second concentration is protecting the reconstruction cycle of Zen. This is the only purely technical judgment among the three actions, and also the only one whose “correctness” can be verified afterward.

The hard place is: before Zen is made, it is a pure expenditure.

It does not produce revenue, does not improve the quarterly report, does not need to be delivered to customers; it only needs money, people, time, and an undisturbed environment. And these three things are exactly what a cash-tight company lacks most.

Here a concrete risk needs to be written clearly: the cost curve of chip projects is not linear.

To protect Zen one must, in the stretch where it burns the most money, is most uncertain, and is farthest from results, block the pressure to cut. This is the most critical resource defense after she took office.

 

Partner Concentration: The Most Misunderstood Step

The third concentration is manufacturing partners. This is the part this chapter must focus on handling, because popular narratives about it are wrong the most.

First make clear the situation she inherited.

In 2009 AMD spun off the manufacturing business as GlobalFoundries, becoming a fabless design company. That is to say, when she took over in 2014 the company no longer had its own advanced wafer fabs.

So the real problem is: where should her most advanced products be manufactured, and what price must be paid for it.

Then make clear what she actually did.

Her approach was not “pioneering outsourced manufacturing,” nor “suddenly turning to TSMC,” but: gradually configure the most competitive products to partners that can provide leading process, and accept the accompanying contract costs and supply-arrangement adjustments.

The difficulty of this action is not technical judgment, but cost. Adjusting manufacturing partners means handling existing procurement commitments, possibly generating fees, and relying on more than one partner simultaneously for a considerable period of time.

Finally talk about what she obtained.

Chiplet is not her invention. The architectural idea comes from the engineering team; industry also has multiple explorations. What she did was select it as company strategy and bear the risk of failure.

 

The Relationship Among the Three Concentrations

Now the three actions can be looked at together. There exists an easily ignored mutual dependence among them:

Abandoning the low-end market is to free resources to protect Zen; protecting Zen needs a manufacturing partner that can keep up on process; and the choice of manufacturing partner in turn constrains the design space of Zen.

In other words, these three things are not three parallel lines, but one chain. If any link on the chain is judged wrong, the value of the other two links will be discounted.

This also explains why looking at this history purely from the angle of “she made a wise decision” is distorted. The three actions had very low visibility at the time; their benefits only successively appeared after 2017, while their costs had already been fully cashed in 2015 to 2016.

 

What She Truly Protected on This Chain

Return to the core question of this book.

If the three concentrations are translated into one sentence, what she wanted to protect was not a certain chip, but the company’s ability to make long-term decisions.

Once a company falls into a state of “every quarter must look good,” it loses the ability to make long-cycle investments. And in the semiconductor industry any truly competitive product requires three to four years. This means there is a structural conflict between short-term performance pressure and the competitive threshold—not that this term of management is not smart enough, but that this industry itself requires you to do some things that look stupid in the current period.

The three concentrations she did are essentially drawing a protection range for this ability: on these three lines, decisions are not allowed to be changed for short-term reasons.

Whether this range was drawn correctly can be debated; but it must exist, otherwise any long-term investment cannot survive the pressure of the first performance decline.

 

Chapter 14 · The Cost of Long-Term Thinking

Long-termism in books is usually a positive word. In this chapter it is a bill, and it comes due quarterly.

 

What Is on the Bill

If one is to draw an expenditure table for her situation between 2014 and 2017, it would have these items.

R&D personnel must be paid. Chip design is labor-intensive work; the scale of a core architecture team is in the hundreds, and they cannot be unpaid for three years. These people will not lower living requirements because the company “is making long-term investment.”

Customers will not wait forever. OEMs and cloud service providers make product plans years in advance. When they say to AMD “we will wait for you,” they are actually only waiting a reasonable time window. If Zen is delayed too long, customers will turn to Intel, and this turn may be permanent.

Investors may lose patience first. This is the most realistic one. Long-termism needs shareholders to wait together, and among shareholders there are always people who are assessed quarterly themselves. When the stock price is long-term low, the call for the board to replace the CEO will not disappear by itself.

The supply chain must be prepaid. Capacity must be locked in advance, wafers must be ordered in advance; these all require cash to go out first and products to come back later.

 

What Ways She Used to Buy Time

Since the bill cannot be reduced, one can only find money. The tools available to her are several.

The first is the console business. Sony and Microsoft’s semi-custom chips continue to ship, providing stable but seasonally fluctuating cash contribution. This is the cleanest sum: no additional commitment needed, no explanation needed, just deliver on schedule.

The second is financing and debt management. The company needs to maintain cash levels through credit arrangements and debt operations, while avoiding triggering risk conditions in debt covenants. This kind of arrangement does not produce products, but produces survival space.

The third is cutting non-core projects. This is the financial-side manifestation of the earlier “three concentrations”: cut the unimportant lines, and the money is freed.

The fourth is asset and cooperation arrangements. Including technology licensing, joint ventures, and other forms. This category of method can bring relatively fast cash, and is also most likely to leave long-term consequences.

 

That Deal, and Why It Deserves a Separate Chapter

In 2016, during AMD’s most financially difficult period, it established a joint-venture and technology-licensing arrangement with Chinese partners, involving server-processor development based on early Zen technology, bringing the company about $293 million in licensing and related revenue.

What weight was this money at the time? Putting it together with contemporaneous numbers: the company’s full-year net loss was about $497 million, year-end cash around $800 million. $293 million is not icing on the cake; it is a full year of breathing space.

So the controversy about this deal from the beginning was not on the same level.

Critics’ judgment is long-term. They believe AMD was one of the few companies at the time that truly pushed open a crack in the x86 door; the direct consequence of the technology license was cultivating a competitor that later occupied an important position in China’s server-chip industry; and the long-term value of technology spillover far exceeds those several hundred million dollars of licensing fees.

Defenders’ judgment is short-term. They believe management conducted a limited, single-generation license within the legal framework, the most sensitive underlying layers were not all handed over, and reported to the U.S. side; for a patient lying in intensive care this breath was very important; without this money the company might not have been able to hold until Zen launched.

This matter will be handled separately in a later chapter of this book, including separation of the four levels of technical boundary, legal issues, national-security issues, and moral issues. Here only mark its position: it belongs to part of the survival-period trade-offs, not an event that can be simply handled with moral posture.

 

How She Herself Responded to the Stage of “Not Seeing Results”

This part needs caution, because it most easily slides into inspirational narrative.

In 2020, at a public event at Stanford Business School, she spoke a passage directly related to those three years. She said that the truly important things in her career were none of them known from the beginning whether they could be done; she also said she has a preference for “solving difficult problems,” because that is exactly where leadership truly takes effect.

The materials also record a relatively widely circulated sentence to the effect: what leadership is truly doing is organizing the team to do something no one thinks can be done.

These words can serve as side evidence of her mindset, but cannot be taken as the work record of those three years. Expressions in public interviews are usually post-hoc organization, not diaries of the time. The real process is more likely: not so much composed self-cognition, only continuous trade-offs that must be made on the spot.

 

The Other Half of Her Personal Life

This chapter also needs to write a part easily overlooked: what she paid in those years was not only professional cost.

In public information her lifestyle is described quite formulaically: sleep about five to six hours; likes to practice boxing with a coach, the media summarizing this as “punching the sandbag in the morning, punching Intel in the daytime”; often drinks a certain Starbucks tea.

These details are easily processed in transmission as personal-charm packaging, but they more likely reflect another thing: in a stage of long-term absence of positive feedback, a person needs a set of rhythms that can be self-sustained. Boxing is not a hobby, it is regulation; a fixed beverage is also not taste, it is ritual.

There is one more thing worth placing here: her husband Daniel Lin is a PhD in economics, teaches at American University, teaching microeconomics, macroeconomics and other courses, research directions involving entertainment-industry organization and competition theory. The two now live in Austin, Texas, while AMD headquarters is in Santa Clara, California. She extremely protects privacy; her husband rarely appears on her social media.

The significance of this segment is not gossip, but that it is the other half of those years. A person who every day must handle bad news at the company needs a living space that is not defined by that company’s quarterly report.

 

 

Part VI · When Bad News Arrives

 

Chapter 15 · When Bad News Arrives

By 2016 AMD’s books were still bleeding.

Revenue returned to $4.27 billion, slightly better than 2015; full-year net loss about $497 million. The company could talk about Zen, could talk about the future, could talk about the data center—the books were still a sea of red.

That year the company maintained operation on three pillars: cash flow from the console business, the several hundred million dollars brought by that technology license, and financing and debt management.

And that year was also the time when AMD internally most needed to answer one question: when bad news becomes daily, what is a company’s first reaction?

 

The First Reaction Is Culture

What this chapter discusses is not a certain crisis, but a more lasting thing.

If the reaction is to find someone to blame, information will be delayed and beautified. The first engineer who discovers the problem will first think “will this be counted on my head if I say it,” so he will first confirm, first find people to discuss, first find a layer of protection for himself. By the time the message reaches the person who can decide, two weeks may already have passed.

If the reaction is to pretend the problem is not serious, the team will lose the scale of judgment. If the person in charge says of an obviously serious fault “it’s okay, it doesn’t affect the node,” then the next time it is truly serious the team will think it is also “okay.”

If the reaction is anger, engineering resources will be consumed in self-protection. People will begin preparing explanatory materials rather than solving the problem; meetings will become defense courts.

All three paths will make the same result appear: bad news arrives later and later.

And in a chip company, bad news arriving late often means the cost multiplies.

 

A Part That Must Remain Honest

Regarding whether AMD had specific, dramatic engineering crises during this period, the public evidence currently available to this book is insufficient.

The materials contain a very attractive statement: early Zen samples once had serious problems, even unable to boot normally; the team bypassed the defect in one month, and verification progress was instead advanced; she was on a business trip in India at the time, with a ten-hour time difference, and after receiving the call only said one sentence “testing cannot stop.”

These details, if true, would be the best segment of this book. But currently no traceable first-hand source can be found. They come from Chinese-program retellings, lacking support from party recollections, emails, meeting records, or contemporaneous technical reports.

So this chapter does not write these scenes.

But this chapter can still write “bad news”—only shifting the focus from an unverified crisis to the verifiable situation.

 

Verifiable Bad News

In those years the bad news AMD continuously faced was these.

The Computing and Graphics segment had a full-year operating loss of $502 million in 2015. This is an entire core product line, losing money for a whole year.

R&D expense cannot be stopped, and cash is continuously being consumed. Both things holding simultaneously means no quarter can relax.

People in the organization are making judgments. Smart people know which projects are destined not to be done, and also know which directions have been cut and then picked up again. When a company is long-term in a cycle of “announce—cut—re-announce,” the most capable batch of people will leave first—and the timing of their leaving is usually the years when the company needs people most.

 

Where Leadership Exists

If bad news is not a late-night accident, then where does leadership exist?

It exists in not beautifying numbers for consecutive multiple quarters.

The language choice in the financial reports has been written earlier: the wording she used in those two years was very restrained, not promising a timetable for improvement, and not packaging bad news as good news. This matter looks only like public-relations style; actually it simultaneously transmitted the same information to four groups of people: this company knows where it is.

It exists in not repeatedly changing direction for the team.

If a company adjusts priority once a year, then every adjustment will waste part of the accumulation of the past year. In a period of resource tightness, stability itself is a resource.

It exists in defining “looking bad” as part of the plan.

This sentence spoken to shareholders is useless; it must be spoken internally. What engineers face every day is a backward product, a suppressed share, and an ugly quarterly report. If no one tells them “this is intentional, not that we failed,” they will respond with the most reasonable personal choice—leaving.

None of these three things constitutes a moving story. But added together they are the reason this company did not fall apart in 2016.

 

If Those Scenes Are One Day Confirmed

This chapter must finally leave a place.

If subsequent interviews can confirm the early-sample fault, then the details that need to be pursued are these: who first discovered the problem? Which modules were affected? Which verification links could continue to advance? How long did the re-spin take? Who made the resource decision? At the time were there emails, problem-tracking records, or review-meeting minutes?

Only when these questions are answered does the scene qualify to enter the main text.

And even if confirmed, it should not be written as “she created a miracle with one sentence.” Defect repair in chip projects is collective work: the verification team finds the boundary, the design team locates the problem, process and packaging teams cooperate, project management re-schedules the timetable. The leader’s role in this kind of matter usually has two: decide not to stop the parts that can continue, and bear the resource cost required for the correction.

These two roles are already important enough. There is no need to expand it into a personal myth.

 

Chapter 16 · 2017: Being Believed Again

In 2017 Ryzen launched.

Regarding that day, public records contain the scenes of the launch event, the numbers of performance comparisons, the carnival of media headlines. But what this chapter must discuss is not the launch event.

What this chapter must discuss is: what truly changed AMD’s destiny that year was not a launch event, but the organization regaining a record of fulfilling commitments.

 

First Look at the Results

In full-year 2017 AMD revenue was $5.33 billion, up 25% year-over-year; GAAP operating profit $204 million, net profit $43 million.

Compare: 2016 was a net loss of about $497 million, 2015 was a net loss of $660 million.

The number $43 million, placed in AMD’s later scale, is insignificant. But it has a nature that other things cannot replace: it means consecutive years of losses terminated.

Lisa Su in the financial-report statement called 2017 a key turning point, the reason being that the company reshaped the product portfolio, expanded gross margin, and achieved full-year profitability.

 

Why the Sentence “Not the Launch Event” Is Important

A chip company launching a generation of successful product is usually narrated as a dramatic node: a certain year and month, a certain product launches, market wind direction changes.

This narrative misses the true mechanism of product success.

The second layer is pricing power. Once the product is no longer backward in core performance, one need not continue to exchange share with price, and gross margin has the chance to improve. This point is directly reflected in the 2017 statement of “expanding gross margin.”

The third layer, and the layer most easily overlooked: predictability.

Enterprise customers do not buy one generation of product; they buy a route. What server customers must evaluate is not how much faster this chip is, but whether three generations later it will still be there, whether interfaces are compatible, whether the software ecosystem can keep up, whether supply is stable. A customer deciding to put AMD into its own server planning is essentially bearing risk for its own next three years.

So what truly turned AMD around was not how strong that generation of product in 2017 was, but that after 2017 the company fulfilled the next generation according to commitment time after time.

 

How the Record of Fulfillment Is Accumulated

This matter is worth concretizing, because it is not abstract.

Products must launch on schedule. One delay can be explained; two delays and customers begin to discount the evaluation. Three consecutive years of on-time delivery can rebuild the judgment “the date this supplier says can be believed.”

Performance must meet the commitment. Comparison numbers at launch events usually select conditions favorable to oneself; customers will test internally themselves. If real performance is substantially lower than claimed, what is lost is not this generation’s orders, but trust in the next generation.

Supply must keep up. Insufficient capacity will make customers more cautious in the next procurement. If a company is short of stock every time a new product launches, customers will believe it cannot support their own business scale.

Problems must be able to be handled. Defects, compatibility issues, patches—what customers look at is not whether you will have accidents, but how soon you respond after an accident, how you solve it, whether there is concealment.

None of these four things can be obtained by one success. They can only be accumulated by consecutive multiple product cycles. This is also why this book repeatedly emphasizes: AMD’s revival did not occur at a certain moment; it grew bit by bit over many quarters.

 

After 2017 the Road Is Still Long

Here one must resist a temptation: writing 2017 as “the turnaround is complete.”

After 2017 AMD’s position in the personal-computer market did improve, and the server business also began to regain customers. But the recovery of server share is a slow process—enterprise verification cycles are long, software adaptation needs time, procurement decisions involve historical relationships. A company re-entering a market and it becoming the preferred choice again in that market are separated by several years.

At the same time the company was still carrying several things: debt structure, relationship constraints with process partners, weakness in software and development ecosystem, and dependence on the most advanced manufacturing.

So the significance of 2017 is not victory, but restoration of qualification. AMD regained the qualification to participate in long-term competition—able to formulate three-year roadmaps, able to let customers write it into planning, able to let developers be willing to invest time in it.

 

From the Organizational Perspective, What 2017 Left Behind

The one observation this chapter truly wants to leave is this.

In those years the thing AMD internally lacked most was not money, nor technology, but a repeatable way of working.

A company long in difficulty usually exhibits one characteristic: every project is treated as a once-and-for-all gamble. Success is luck, failure is fate. In this state the organization learns nothing, because it has no margin to summarize: why was it delayed last time? Is it a problem of estimation method, or of architecture design, or of verification method?

After 2017 AMD began to have the margin to do this. It was able to form a repeatable product rhythm: design, tape-out, verification, launch, iteration. This rhythm itself is the product of long-termism—it needs stable resources, stable teams, and stable timetables; it cannot grow in an environment that changes every three months.

 

 

Part VII · The $293 Million Question

 

Chapter 17 · Technical Facts

In February 2016, during AMD’s most financially difficult period, it established a joint-venture and technology-licensing arrangement with Chinese partners.

Regarding this matter, discussion in the Chinese-speaking world has almost entirely concentrated on one word: sell. Sell to whom, sell how much, sell what technology.

This chapter adds no judgment; it only does one thing: split it into technical facts. Because on this matter the factual level itself contains a large amount of confusion, and positions are often built on top of the confusion.

 

The Basic Structure of the Deal

Time: around 2016.

Form: not a simple technology sale, but a packaged arrangement that included joint ventures, technology-licensing permits, and R&D cooperation and other clauses of different natures.

Result: public materials show that AMD thereby obtained about $293 million in licensing and related revenue.

Counterparties: the Chinese-partner side involved Tianjin Hygon, and behind it the lineage of the Chinese Academy of Sciences and Sugon.

Here the first thing that needs to be made clear: this is not “selling the company to China,” nor “AMD being acquired.” AMD is a company listed in the United States; its ownership, board, and headquarters did not change because of this. What it sold was a right of use, and it was time-limited and boundary-limited.

 

What Was Licensed

This is the core question, and also the place most often blurred in the controversy.

The rough outline sketched by public information is:

The license was described as single-generation. That is to say, it does not constitute a continuous technology-transfer channel; it is not “every new technology must be handed over,” but a one-time, scope-limited permit.

The most sensitive underlying technology was retained. The expression of this point is relatively consistent across sources, but “which count as sensitive underlying” has no public complete list.

Subsequent iterations were not within the license scope. After related entities were placed on the U.S. export-control list in 2019, she clearly stated there was no subsequent technology transfer.

 

Three Technical Questions That Must Be Separated

Regarding the technical level of this deal, three questions are often mixed together. They must be separated.

The first question: what was licensed?

This is a question about “object.” Which intellectual property, which designs, which tool chains, which process-related materials the license scope includes determines the weight of this matter. Public information is insufficient to give a complete answer.

The second question: what did the licensee make?

This is a question about “result.” Around 2018 the Hygon side launched server-processor products based on that technical route. Thereafter some developers tested and observed: the related products are highly similar at the software level to AMD’s server chips; porting AMD’s kernel code over requires changing only a very small part.

The observation “only needs to change a very small part” is the key technical evidence for understanding the whole matter. Its meaning is: the two products are not “referencing the same design idea,” but share a considerable degree of homologous structure at the code level. This shows the license was not formal; it truly transmitted a technical foundation capable of establishing a product line.

The third question: how fast was this result?

This is a question about “distance.” From license to product launch only a short time intervened. This speed itself shows: what the licensor provided was not “direction,” but something already close to engineering.

Looking at the three questions separately yields a description more accurate than “selling technology”: this is a scope-limited, generation-limited, but content-real technology license that enabled the licensee to possess the capability to launch server-grade products in a relatively short time.

 

Which Statements Do Not Hold

At the level of technical facts several common statements also need to be cleaned up.

“AMD handed the entire Zen architecture to China”—does not hold. What was licensed was the technical route already existing at the time, not the complete new-generation architecture she subsequently advanced; and it was described as single-generation.

“Nothing was transferred, only an empty shell was bought”—also does not hold. The high similarity at the code level shows technology was indeed transmitted; the licensee’s products have a substantive foundation.

“This was AMD’s only life-saving money; without it it could not survive”—the expression is too strong. The weight of that revenue was indeed large, but the company simultaneously had other cash sources. A more accurate statement is: it significantly improved the company’s survival space in that year-plus.

“This was business she negotiated only after taking office”—needs verification. This kind of cross-border joint-venture arrangement usually requires a longer negotiation cycle; the start time of the deal and the public time are often not the same thing. This chapter does not adopt the narrative of “she decided it after taking office.”

 

Why Technical Facts Are Important

A concrete example: if the license scope is “continuous transfer of the entire architecture,” then it is completely different in moral evaluation from “a one-time, generation-limited permit.” The nature of the former approaches transfer of technology-output capability; the latter is closer to exchanging an existing asset for cash—the long-term consequences of these two things differ enormously.

So any judgment about this matter must first return to the technical boundary. This is also precisely the writing principle of this book: technical facts cannot be replaced by position.

 

Chapter 18 · Commercial Interests

After the technical boundary is made clear, it is then the turn of that money.

This chapter only handles one question: in the situation of 2016, what exactly did this $293 million mean.

Looking at It Inside the Financial Report

In 2016 AMD’s situation was this: full-year net loss about $497 million; year-end cash and marketable securities around $800 million; debt scale on the order of $2.2 billion.

Now put the $293 million in.

It is not icing on the cake; it approaches two-thirds of that year’s net loss. In other words, without this revenue that year’s loss would have been closer to doubling, cash consumption would have been faster, and the company would have needed to find more money from elsewhere before Zen launched—and “finding more money from elsewhere” in a company whose stock price is long-term depressed and that analysts publicly discuss as potentially bankrupt is a very expensive thing.

This is the core of the commercial defense, and it is real: surviving is the prerequisite condition of all decisions at the time. A company that can no longer continue operating cannot protect any technology, any position, any long-term plan.

The Strongest Version of the Commercial Defense

If one is to make the strongest defense for this deal, it would roughly be written like this:

First, it was legal. The deal was completed within the legal and regulatory framework of the time; the technical-license scope involved was limited; management reported as required.

Second, it was limited. The license was described as single-generation, one-time; the most sensitive underlying layers were not all handed over; subsequent technology iterations were not within the license scope.

Third, what it exchanged for was time. This money gave the company an additional cash-flow channel in the most dangerous window and fought for space for the completion of the new-generation architecture.

Fourth, it exchanged for market. China (including Hong Kong) has long been an important revenue source for AMD, in recent years accounting for roughly one-fifth of its revenue. Establishing relationships with local partners has practical value for a company that needs scale.

Fifth, the alternatives were worse. If there had been no this money at the time, the company might have needed to accept more humiliating financing conditions, sell more valuable assets, or go directly toward liquidation.

None of these five is about long-term consequences. They all answer the same question: if not done at the time, what would happen.

The Strongest Version of the Critics

The critics’ argument can likewise constitute a complete version, and it does not rely on moral accusation:

First, the time scales are different. Licensing fees are a one-time several hundred million dollars, while the consequences of technology spillover may last twenty years. Exchanging one quarter’s cash for a long-term pattern change of an industry is hard to calculate as worthwhile financially.

Second, the nature of the licensee. The other side of this deal is not an ordinary commercial partner; behind it connects to China’s national-level scientific-research system and technology institutions regarded by the United States as strategic competitors. This nature was not caused by AMD, but it cannot be avoided when evaluating the consequences of this deal.

Third, the depth of the license. From the degree of technical homology of subsequent products, this was not a formal license. It enabled the licensee to possess the capability to launch server-grade products in a very short time—this “very short time” itself is evidence that what was transmitted was sufficiently close to engineerable.

Fourth, AMD’s position is special. In the x86 ecosystem, companies that truly hold a complete architecture license and possess design capability are extremely few. Therefore this license is not one among many optional schemes, but opening a crack in a door that was originally very narrow.

Fifth, commercial defense cannot replace consequence assessment. A decision that is rational under the situation of the time cannot be inferred to be wise in the long term. These two things must be evaluated separately.

Note: the critics’ argument also does not rely on the premise “she is a bad person.” It relies on time scale—when the evaluation cycle is stretched beyond ten years, an originally reasonable commercial decision may present a completely different nature.

Why the Two Judgments Cannot Be Reconciled

Looking at the two sides’ arguments together, one discovers they are not giving different answers to the same question.

The defense side answers: “In February 2016, what should AMD do?”

The criticism side answers: “Looking back in 2026, what did this deal cause?”

These are two different questions. The constraint conditions of the first question are cash, debt, and time window; the second question has no constraint conditions; it can invoke all subsequently occurring facts.

Therefore, using the results of 2026 to evaluate the decision of 2016 is methodologically invalid. But conversely, using only the situation of 2016 to evaluate this decision is also invalid—because the conclusion it reaches is “the company of the time had no choice,” and if this sentence holds, it means any legal decision made by any enterprise under survival pressure should not be held accountable. This inference obviously cannot be accepted.

What truly needs to be asked is a harder question: when a company under survival pressure sells technology rights, for which consequences should it be responsible?

This question has no standard answer, but it is far more meaningful than “should one scold or not.”

Her Personal Position on This Matter

There is a passage that needs to be written clearly here.

In 2019, after related entities were placed on the U.S. export-control list, her public statement was clear: no subsequent technology transfer; the license was one-time.

The nature of this statement needs to be accurately understood. It is not defending a moral decision, but stating a factual boundary: it has occurred, but it will not continue. This is a technical explanation, not a value judgment.

She has also emphasized on other occasions that semiconductors are a team sport and need long-term cooperation with supply-chain partners such as TSMC, Samsung, and SK Hynix. This sentence is in the same logic as that deal: what she is handling is a highly interdependent global industry, in which “complete self-sufficiency” has never been an option.

This is not exemption, but situation. It shows that the choice she faced was never “whether to deal with China or not,” but “under what framework, at what cost, bearing what risk to deal.”

A Point That Is Dishonest If Not Raised

Finally one needs to say something not very comfortable.

That deal rescued in 2016, but its negotiation and structural design obviously required a relatively long period of advancement, involving cross-border cooperation, multi-party joint-venture arrangements, technical-scope definition, and regulatory communication. Of these works, how much was led by her after taking office, how much continued from the previous incumbent, public materials are insufficient to answer.

This blank must be left honestly, rather than filled with “she was appointed in crisis and decided to save the day.” Because if it is a continuing arrangement, then her role in it is closer to execution of the decision and drawing of the boundary, rather than designer of the deal—the responsibility natures of these two roles are completely different.

Chapter 19 · National Security and Moral Judgment

In June 2019 the U.S. Department of Commerce placed several Chinese entities on the export-control list, among them involving one party related to that deal.

From this moment onward, a business that in 2016 had been measured in cash was placed in another scale for measurement.

 

Four Layers of Questions, Must Be Separated

What this chapter must do is only one thing: split the four questions that have always been mixed together.

The first layer is the legal question: was this deal legal at the time?

This question has a relatively clear way of answering—look at the export-control rules of the time, the legal framework of technology licensing, and whether the company made disclosure and reporting as required. Public records show the deal was completed within the framework of the time; management claimed it had reported to the U.S. side. The controversy at this layer is mainly not “whether it violated rules,” but “whether disclosure was sufficient”—and this point requires first-hand documents to judge; this book currently does not possess them.

The second layer is the commercial question: was it worth doing?

This was already handled in the previous chapter. The core is the conflict of time scales: one-time cash gain versus a period of technology spillover that may last twenty years.

The third layer is the security question: what potential uses does the transferred technology have in military or national-security terms?

This layer is most easily overestimated, and also most easily underestimated.

The manifestation of overestimation is: equating all commercial chip technology directly with military capability. Server processors are general-purpose computing products; their direct military meaning usually does not lie in the chip itself, but in whether it can support large-scale computing infrastructure—and this, in today’s AI era, has precisely become more sensitive than ten years ago.

The manifestation of underestimation is: because it is a civilian product, determining it has no security meaning. The boundary between high-performance computing capability and artificial intelligence, code-breaking, simulation, and data processing has become increasingly blurred in the past ten years.

So the accurate statement is: the security meaning of this type of technology is indirect and grows with time, not immediate and singular. This can neither support the defense of “it is unrelated to security,” nor support the accusation of “it equals arming the opponent.”

The fourth layer is the moral question: even if everything is legal, does the enterprise still bear responsibility beyond the minimum compliance standard?

This layer is one that the first three layers cannot answer.

 

Why Law and Commerce Both Cannot Answer the Fourth Layer

Law answers “can it be done.” Its judgment basis is the text of rules and regulatory intent. A company that complies with the rules then in effect is legally clean.

Commerce answers “is it worth doing.” Its judgment basis is cash flow, market, risk, and return. A deal that is financially worthwhile is commercially reasonable.

While morality asks “should it be done.”

The difference among these three questions does not lie in degree of strictness, but in the coordinate systems they use.

Law and commerce both take the present and the calculable range as boundary: rules are valid within their effective period; deals settle within the contract period. Morality requires you to consider those consequences outside the rules that cannot be displayed on the current-period report—including the impact of your behavior on the industrial pattern ten years later, including the costs you yourself do not directly bear.

This is why “legal” cannot become the endpoint of moral defense. A law allowing something may be because legislators did not foresee a certain situation, may be because rule updates always lag technological change, or may be because the long-term consequences of the matter could not be seen clearly by anyone at the time. The permission of law only indicates “this is not a violation,” not “this is right.”

This is also not requiring enterprises to bear infinite responsibility. It is requiring a more restrained thing: when making a decision to cede rights, place into the judgment those consequences that cannot be written into the contract, cannot be settled in the current quarter, and cannot be borne by oneself.

 

A Difficulty That Refuses to Disappear

But this chapter cannot stop here, because if it only stops at “enterprises should have higher standards,” it becomes a correct empty sentence.

The real difficulty is: in 2016 AMD’s situation did not allow it to slowly weigh the consequences ten years later.

If the company in that year, because of insisting on higher standards, refused this deal, then the most likely result is not “it preserved technological sovereignty,” but that within two years it would go toward liquidation or be forced to accept harsher merger conditions. At that time technology, team, and intellectual property would be disposed of in another way, and this time it would not even have room to negotiate conditions.

In other words, the moral requirement here collides with a concrete dilemma: when a company’s survival itself depends on a deal with long-term cost, does it have the qualification to talk about “higher standards”?

This dilemma cannot be resolved by any clean statement.

What can be said are only two things.

First, situation can explain behavior, but cannot cancel consequences. A deal becoming understandable because it was made in extremity does not equal the consequences it caused therefore disappearing. The industrial pattern formed ten years later is real and will not change because the company of the year was in difficulty.

Second, evaluation needs to be layered, not compromised. Hitting both “the situation of the time” and “the long-term consequences” fifty-fifty and reaching the conclusion “it is both right and wrong” equals saying nothing. A more meaningful way is to give judgments separately on two dimensions: on the survival dimension this decision is understandable; on the industrial and security dimension its consequences need to be accounted for.

These two judgments hold simultaneously and do not offset each other.

 

What She Bears in This Matter

Finally the nature of responsibility must be made clear, because this is the place most easily unbalanced.

What she bears is not the responsibility of “designing a harmful deal.” As stated in the previous chapter, public materials are insufficient to confirm to what extent the negotiation and structural design of this deal were led by her. If its start was earlier than her taking office, then her role in it is closer to boundary drawing and execution.

What she truly bears is another thing: after obtaining the benefit of this money, how to define its boundary.

This point has content in actual operation: whether the license is single-generation, one-time, or a continuous channel; whether the most sensitive underlying layers are retained; after being placed on the list whether subsequent transfer is clearly cut off. The drawing of these boundaries determines whether this deal is “one-time asset realization” or “long-term technology-output channel”—the consequences of the two differ enormously.

She defined its boundary within a one-time, non-continuing framework. This is a judgment that can be evaluated, and its nature is closer to the containment of risk rather than the manufacture of risk.

And this is precisely the work least easily seen in this kind of matter: it does not produce brilliant narrative, nor appear in launch events or financial-report headlines. It only makes a thing not become worse.

 

 

Part VIII · Beyond the Breakthrough: The Next Competition

 

Chapter 20 · February 2022

February 2022 is a time point worth writing separately.

In this month AMD’s market capitalization first surpassed Intel’s. In the same month the company completed the acquisition of Xilinx for about $49 billion. And roughly around this time she succeeded as chairman of AMD.

Three things crowded into the same month, looking like a watershed of an era.

But what this chapter must handle is precisely the distortion brought by the word “watershed.”

 

What Market Capitalization Surpassing Intel Means

First make clear the nature of this number.

Market capitalization is market expectation, not company strength. It is obtained by stock price multiplied by shares outstanding, and stock price reflects investors’ judgment of future cash flow—a view, not a report card.

So the accurate meaning of the sentence “market capitalization surpassing Intel” is: at that time point the market believed AMD’s future was more worth betting on than Intel’s future.

The establishment of this judgment has several layers of basis.

First, the reversal at the product level had already occurred. Earlier it was written that after 2017 AMD, relying on the new core architecture, re-entered competition. Around 2019 the company obtained a relatively leading position on process nodes—this is a very rare thing in AMD’s history.

Second, the server business continued to recover. This is the slowest and also most important part. Recovery of share in the enterprise market requires multiple years, but once recovered it brings high margin and long-term customer relationships.

Third, Intel’s mistakes opened a window. Process advancement blocked, product cadence repeatedly, giving the originally suppressed side space to breathe and counterattack.

Fourth, and the layer most obvious looking back today: AI. The market began to reprice computing-power demand, and AMD was in the beneficiary position of this narrative.

Of these four, only the first two are directly related to AMD’s own behavior. The latter two are external conditions. Mixing them together turns a story of “the company did the right things and also caught the window” into a story of “the company won the opponent by its own ability.”

 

The Xilinx Acquisition

The other matter of the same month is the acquisition of Xilinx.

The weight of this deal needs to be explained. Xilinx is not an ordinary small company; it is one of the major manufacturers in the field of programmable logic devices (FPGA) and adaptive computing, with long-term customers in communications, industry, automotive, aerospace, and data center. The price of about $49 billion was one of the largest mergers in the semiconductor industry at the time.

What this acquisition brought AMD.

Structural reinforcement of the product line. Previously AMD’s composition was mainly CPU and GPU, plus semi-custom business. What Xilinx brought is a third computing form: programmable logic. This has clear application scenarios in the data center—certain workloads are more efficient with FPGA than with general-purpose processors.

Scale and customers. Xilinx’s user group does not completely overlap with AMD’s original PC and server customers; the acquisition brought new industry channels.

Difficulty of integration. This point was not said much in the reports of the year, but it is real: a company known for engineering culture going to integrate another company known for engineering culture but with a completely different customer structure. Organization, product roadmap, sales method, talent retention—every item needs time.

So the evaluation of this acquisition cannot look only at price and strategic significance; it must also look at whether it has truly been digested. This requires several years to answer.

 

The Matter of Succeeding as Chairman

Regarding her succeeding as chairman, cautious handling is needed.

Chairman and CEO are two different positions. The CEO is responsible for execution—formulating strategy and leading the organization to make it; the chairman is responsible for presiding over the board, supervising management including the CEO, and being responsible to shareholders. Under normal circumstances these two positions are held by different people, because there is a checks-and-balances relationship between them.

When the same person holds both, it means the company’s governance structure has changed: the board chose to let her simultaneously master execution power and the leading power of the supervisory procedure. This is not rare in a rapidly changing industry that requires high concentration, but it does weaken the design of separation of powers.

Why she could obtain this position. The answer is not complicated: the company’s strategic direction obtained the board’s continuous support, and her position in it had already become irreplaceable. This is both confirmation of the work of the past eight years and a concentration of risk.

 

A Causality Easily Told Wrong

Now return to the question at the beginning of this chapter: is February 2022 a watershed?

If told by timeline, it is indeed a landmark node. But if told by causality, it is not the starting point of anything.

Market capitalization surpassing Intel is the lagged result of a series of decisions from 2014 to 2021. Those decisions were made in 2014 (abandon low end, protect architecture reconstruction), 2016 (maintain R&D when cash flow was tightest), 2017 (deliver the first generation of new architecture according to commitment). And 2022 is only the moment these decisions were recognized by the market.

The Xilinx acquisition is the result of negotiations from 2019 to 2021. Its significance will only appear in later years.

The chairman position is confirmation of an already-completed execution period.

In other words: nothing happened in February 2022; it only made what had already happened visible.

This is precisely a methodological problem this book repeatedly emphasizes: a company’s most important decisions are usually made when it is least favored and also has the least positive feedback. And its recognition by the outside often waits many years.

 

But Achievement Also Changes the Problem

There is also a layer that must be written in here.

Success manufactures new dangers.

A company that has won respect in the identity of challenger, after becoming a large company, will face a set of brand-new problems.

The scale of the organization. Projects increase, layers increase, cross-department coordination costs rise. The kind of focus that once let the company survive is hard to maintain after scale expands—because every business can speak a set of reasons why it should be invested in.

The expectation of capital. Rising market capitalization means market expectation also rises. A company’s stock price contains future growth commitments; this will make “long-term investment that takes three years to take effect” difficult again—only this time the pressure comes from must continue to grow, not from must survive.

Judgments that were correct in the past may need to be modified in the new situation. This is not contradiction; this is the normal state of leadership: the problem keeps changing.

 

Chapter 21 · Making the World Willing to Pay for a Second Choice

At the end of 2023 AMD launched data-center AI accelerator cards, formally declaring war on Nvidia.

By 2026 this war has already been fought for nearly three years. And its nature is different from all the campaigns AMD has previously experienced.

 

This Time the Opponent Is Different

First make clear where the difference lies.

The battles AMD previously won, the opponent was Intel. That was product-to-product competition: whose core performance is higher, whose process is more advanced, whose power consumption is lower, whose price is more attractive. In this kind of competition, as long as you remake the foundation once and make a product that can truly fight, customers will reconsider you—because no one is willing to procure server chips from only one supplier.

Nvidia’s moat is not on the chip, but on the software.

CUDA has accumulated since around 2006 and has already walked twenty years to today. It is not a driver; it is an entire system: programming model, acceleration libraries, development tools, debugging means, community documentation, and most important—millions of developers already accustomed to it.

This means a very concrete consequence: customers migrating to AMD is not changing a piece of hardware, but changing a set of working methods.

Code must be rewritten, re-optimized, teams re-trained, and the risk of performance uncertainty borne. For a cloud service provider racing time to deploy AI, this cost may be much higher than the hardware price difference.

This is the essence of this war: not “whose chip is stronger,” but “who can make the world willing to pay the migration cost.”

 

AMD’s Situation: Starting from “Second Choice”

Her path was chosen very clearly.

Not pursuing immediately replacing Nvidia, but first becoming a “credible second choice.”

This goal sounds conservative, but it conforms to the reality of a challenger. The so-called credible second choice contains several concrete conditions:

Performance sufficient. Not requiring comprehensive lead, but no obvious shortcoming on mainstream workloads.

Software can run. Mainstream frameworks such as PyTorch, TensorFlow, JAX must be usable; mainstream models must be deployable; common optimization means must be effective.

Supply predictable. Cloud service providers deploy thousands of accelerators at a time; if supply is unstable, any performance advantage is meaningless.

Economics worth migrating. This may be the most direct one—customers are willing to bear the migration cost on the premise that the cost-performance gap is large enough.

The key of this goal is: it does not require you to win; it only requires you to be counted in. And once counted in, one enters the candidate list of procurement decisions and has the opportunity for continuous learning and improvement.

 

The ROCm Path

AMD’s weapon against CUDA is ROCm—an open-source software stack.

The word “open” needs to be concretized, otherwise it is only a value declaration.

ROCm’s actual claim is: developers can view, modify, and adapt this set of software without having to rely on a single supplier to provide all tools. For customers this means theoretically lower lock-in risk, more optimization space, and possibly better cost control.

But open also has its costs, and these costs were very real in the early years.

First, maturity. The maturity of a software stack needs time and also needs feedback from a large number of real users. Things CUDA has accumulated for twenty years cannot be made up by ROCm in a few years.

Second, fragmentation risk. Open means more participants, and also means more compatibility combinations need to be maintained. When all possible configurations must be supported, the engineering burden becomes complex.

Third, developer habit. This is the hardest layer to change. Developers are not calculators that rationally choose the optimal tool; they will use what they are familiar with, and will also follow peers. Once habit is formed, changing it requires a generation or a paradigm shift.

 

A Statement Worth Recording

In public discussion in 2026, the AMD side gave a judgment of the “CUDA moat” worth recording.

The gist is: customers already rarely mention CUDA anymore. One of the reasons is that artificial intelligence itself is becoming an aid to migration—using AI tools can convert and optimize code originally written for CUDA onto ROCm. AMD has also launched corresponding platforms to teach AI programming assistants how to adapt and optimize existing code for ROCm.

This judgment needs to be treated cautiously.

It is the company’s claim, not a market conclusion that has been verified. If migration truly becomes easy, that is a major benefit for AMD, so it has reason to express it this way. And independent verification needs to look at two things: how high the real migration cost is, and what the actual performance is after migration. Neither of these can be proven by launch events.

So the writing method of this chapter is: record this statement, mark that it is a claim, point out the method of verifying it.

 

Rack-Level Systems: Change of the Competitive Unit

In 2025 to 2026 both AMD and Nvidia are doing the same thing: turning the unit of competition from “chip” to “entire rack.”

AMD in July 2026 launched the Helios rack-level solution: one rack connecting 72 GPUs, integrating EPYC processors, Pensando networking, and ROCm software.

The significance of this shift is not only product form.

First, it acknowledges a technical reality: the bottleneck of AI is already not on the single chip. Memory capacity, memory bandwidth, interconnect between chips, data transmission within the rack, power supply and heat dissipation—these system-level factors often determine actual performance more than the peak computing power of a single chip.

Second, it changes the threshold of competition. Selling chips only requires the chip to be well made; selling racks requires simultaneously doing chips, networking, packaging, power supply, heat dissipation, and software correctly. This requires higher comprehensive engineering capability of a company.

Third, and the most interesting point: it makes AMD approach the opponent in posture.

AMD’s previous selling point was “open” and “choice,” the implicit contrast being Nvidia’s vertical integration. But the rack-level system itself is a highly integrated product—it needs someone to do unified design across chip, network, and software boundaries.

AMD’s solution is: integration is necessary, but integration should be built on open standards, and the design can be contributed to the industry. It links Helios with the Open Compute Project (OCP), emphasizes co-design with the community and with customers, and plans to contribute the Helios architecture to OCP.

This is a very subtle position: it does not oppose integration; it advocates that the rules of integration should be open. Whether this claim can be accepted by customers depends on one thing—whether systems of open standards can avoid being pulled open in efficiency by vertically integrated systems. If the gap is too large, any discourse about choice will lose persuasiveness.

 

About Those Performance Numbers

Both AMD’s and Nvidia’s launch events are full of comparison numbers. This chapter must make clear their nature.

AMD side: the company claims Helios under specific models, specific sequence lengths, and specific interaction conditions can achieve higher certain throughput indicators or better cost-performance relative to Nvidia’s rack solution. These numbers come from AMD laboratory calculations, modeling, or estimates, and used specific test conditions; the company clearly marked these premises in the materials.

Nvidia side: likewise has a large number of statements about performance, energy efficiency, and deployment scale; these are also the company’s own claims.

The common nature of these numbers is:

They are all company claims, not independent verification. Either side at launch will choose indicators, configurations, and comparison objects favorable to itself.

They all depend on premises. Different models, sequence lengths, batch sizes, concurrency levels, and software versions will all significantly change the result. A solution that leads under low concurrency may lag under high concurrency.

They are none of them revenue. Announcing customer cooperation, displaying deployment cases, disclosing roadmaps—the distance between these three things and “forming scale revenue” may be measured in years.

The handling method of this book is: record both sides’ claims, mark their nature and conditions, do not take either side’s numbers as conclusion.

 

Chapter 22 · Two Industrial Orders

The previous chapter handled product and ecosystem. This chapter must handle another layer of change.

The competition between AMD and Nvidia is already not only a product contest between two companies. It has become a contest between two industrial orders—one advocating that a unified platform can bring efficiency, the other advocating that open standards can retain choice.

And behind these two claims stands an older question.

 

First Make Clear This Is Not a Morality Play

Summarizing the two routes as “open versus closed” easily becomes a morality play: open is good, closed is bad; freedom is virtuous, monopoly is evil.

This framework neither conforms to fact nor helps judgment.

Nvidia is not a closed company. Its GPUs can run any framework, its model libraries are open to everyone, its development tools also have free versions. What it truly does is a more concrete thing: making the entire technology stack as much as possible written in the same set of architectural language. From the developer writing the first line of code, to the model running on tens of thousands of accelerators, to the power and cooling of the machine room, all try to use the standards it defines.

The definition Nvidia gave itself in its 2026 annual report is “vertically integrated, horizontally open.” Although this statement is brief, it accurately describes its posture: tightly interlocking vertically, extensively open horizontally.

AMD is also not a purely open company. The rack-level system it launches is itself a highly integrated product, needing someone to do unified design across chip, network, and software boundaries. The openness AMD advocates is that the rules of integration should be open—contributing the design to industry standards so that customers and partners can choose and replace within them.

So the real difference is not open versus closed, but: who decides the interfaces.

 

The Respective Rationality of the Two Models

The reason for vertical integration is efficiency.

When chips, interconnect, network, software, and system are designed by the same team, they can optimize targetedly against each other. A concrete example: if the software team knows how the memory hierarchy of the next-generation chip will change, they can adjust scheduling strategy in advance; if the network team knows the communication pattern between GPUs, they can design switch topology targetedly. This cross-layer optimization is hard to achieve in loosely combined systems.

Jensen Huang has a very straightforward expression of this efficiency, to the effect: rather than optimizing seventeen different technology stacks, optimize one set and let this one set run on the entire cluster; when the software updates, the existing AI factories as a whole improve performance.

The reason given by open standards is choice.

When interfaces are public and participants are diverse, customers need not completely hand themselves over to one supplier. This has direct value for the buyer: bargaining power, supply security, and influence on the technical route.

There is also a more practical consideration: any single supplier has the possibility of making mistakes. If one company’s route choice is wrong, the customers who depend on it will be damaged together; if there are multiple feasible schemes on the market, customers can hedge this risk.

So the value of openness does not lie in the abstract concept of “freedom,” but in that it transfers part of the risk off the customer.

 

Both Face the Same Problem

Here is a symmetry that is rarely pointed out.

The risk of vertical integration is that judgment errors are amplified.

If a company simultaneously defines chips, network, software, and system, then every judgment error of it will be conducted along the entire stack. Customers not only bought its chips, but also bought its architectural assumptions. If this assumption is later proven wrong—for example a certain computing paradigm does not become mainstream, or a certain memory technology cannot be walked—the loss borne by customers will be amplified manifold.

The risk of an open ecosystem is that efficiency is averaged.

When participants are numerous and interfaces need to be compatible with multiple combinations, optimization can often only be done to “still okay for everyone,” and cannot be done to “extremely optimized for a certain workload.” In a market where every bit of energy efficiency directly corresponds to electricity bills and revenue, this gap is real.

So neither route is free. The difference is where the cost is placed: vertical integration places the cost into dependence; the open ecosystem places the cost into maintenance complexity and performance ceiling.

 

The Real Stake of This Competition

Now one can ask that more fundamental question: what is this competition truly contending for?

Not market share, not profit margin, not even technical leadership.

What is contended for is the power structure of AI infrastructure.

Specifically, three things:

First, who defines the standards. Who decides how chips communicate with each other, how racks are organized, how software is written, how models are deployed. This is not a technical detail; this is the skeleton of industrial order.

Second, the way value is distributed. In a highly integrated system the distribution of value is dominated by the integrator; in an open system the distribution is decided by multi-party game. This will decide who can make money in this industry.

Third, the direction of evolution. Where technology goes next, who judges, who trials and errors, who bears the trial-and-error cost. When only a few participants have the ability to bear trial and error, the exploration direction of the entire industry will narrow.

This is why the importance of this competition exceeds the scope of two companies. It concerns what shape computing-power infrastructure will grow into in the next ten years, and who has the right to speak in this shape.

 

Her Position on This Matter

Lisa Su’s position in this matter can be summarized in one very plain sentence:

She is not trying to abolish integration; she is trying to give integration an opponent.

This position is not radical, nor does it require her to be a moral model. Its rationality comes from a more practical judgment: when a market has only one feasible technical route, this market is no longer a market.

And her position also has a particularity that needs to be pointed out: AMD is both a challenger and a holder of vested interest. It hopes open standards can open the market, but in its own product line it likewise needs integration in order to compete. So its “open” claim is not a pure value choice; it also contains positional consideration.

Chapter 23 · Life Beyond the Balance Sheet

The previous chapters talked about industrial pattern, technical route, and power structure. This chapter must return to a more daily level.

Because behind these strategies she is still a person living in Austin, Texas, practicing boxing every day, sleeping very little.

A Space Not Defined by the Quarterly Report

First a concrete number: she now lives in Austin, Texas, while AMD headquarters is in Santa Clara, California.

This is not a small matter. Managing a company whose market capitalization is measured in hundreds of billions of dollars usually means living near headquarters, able to enter the conference room at any time. Choosing to live more than two thousand kilometers away is actively drawing a boundary for oneself.

The function of this boundary is often more important on executives than outsiders think.

What Austin provides is precisely such a place. It gives work a physical endpoint—get off the plane, go home, fly back the next day.

This is not free and easy; it is a necessary condition for maintaining long-term working ability. Earlier it was written that in the years 2014 to 2017 she had no good news she could publicly announce; the company needed her to bear looking bad for consecutive multiple quarters. In this situation, if a person does not leave a space for herself that is not disturbed, it is hard to get through three years.

Boxing and That Cup of Tea

Regarding her life details, public records repeatedly appear these several items.

She likes to practice boxing with a coach. The media summarize this in one sentence—“punching the sandbag in the morning, punching Intel in the daytime.” This statement has circulated widely because it has visual quality.

She sleeps about five to six hours.

She often drinks a certain Starbucks tea.

These three items are used in transmission as personal charm: an energetic, combative strong-woman image. But more worth noting is their function.

Boxing is not a hobby; it is regulation. A person who every day must handle bad news needs a channel that can release accumulated tension. Exercise usually bears this function on this type of person—it provides a clear beginning and end, has a measurable object of confrontation, and does not require speaking.

A fixed beverage is also not taste; it is ritual. In work whose itinerary constantly changes, small repeated things can provide a stable anchor. This kind of habit is often not a display of personality, but a byproduct of high-pressure state.

Her Husband

Regarding her marriage, public information is little, but several things are clear.

Her husband is Daniel Lin.

He is not a figure in the tech circle. He is a PhD in economics—doctorate from George Mason University, bachelor’s from UCLA—teaches at American University, teaching microeconomics, macroeconomics and other courses, research directions including organization and competition theory of the entertainment industry.

The two now live in Austin, often attend public events together, including the White House state dinner and related dinners of Time magazine.

She extremely protects privacy. Her husband rarely appears on her social media. When he can be seen in public occasions, it is usually in the role of accompanying attendee, not the object of introduction.

Here is one thing worth thinking about.

A person who has reached the highest position in the semiconductor industry, the choice of spouse is a university professor. This is not typical in the Silicon Valley executive group—in that circle spouses are usually investors, entrepreneurs, or people likewise employed in tech companies.

This combination brings two different rhythms of life. The rhythm of a chip company is calculated by quarter and product cycle—acceleration, delay, competition, delivery. The rhythm of academic life is calculated by semester, paper, and long-term research. There is almost no intersection between the two.

This belongs to the author’s observation, not her own statement. She has never publicly explained this point, and this chapter should not explain it for her.

Regarding Children, the Part That Needs to Be Written Clearly

Here a problem that exists in the materials but has insufficient evidence must be handled.

The materials mention: public information has almost no details about children; there are early family photos showing her husband holding a little girl; Chinese reports say the child takes the mother’s surname. But there is no authoritative public confirmation.

The handling method of this book is not to write it.

There are three reasons.

First, evidence is insufficient. The basis is inference from family photos and statements of Chinese media; there is no verifiable public confirmation. Writing them into the main text is using an unverified inference to fill a person’s private life.

Second, this is the private domain. Her family members are not public figures and have not chosen to become objects of reporting. Even if the information is accurate, there is no reason to publicly expose a minor for the completeness of a biography.

Third, it is unrelated to the theme of this book. What this book pursues is how she made choices in the face of difficulty, and what these choices mean. Whether children exist or not does not change any analysis already made.

So in the main text this book only retains one sentence that can be confirmed: she extremely protects privacy. This is not an information gap; this is an intentional boundary.

A Low-Key Person, Why Must Be Public

At this point this chapter must handle the final contradiction.

She is a low-key person, but her position requires her to be public.

This tension runs through her public image. Contrasted with Jensen Huang, this difference is particularly obvious. Jensen Huang is an outstanding public speaker; he gives names to a world that has not yet arrived, turns technical turns into stories everyone can understand. She is the opposite: wording restrained, rarely depicting distant visions, speaking manner closer to an engineer reporting progress—how much performance this generation of product improves, when it launches, where the next node on the roadmap is.

But her low-key is not a natural state, but a choice.

The market leader needs to expand the world’s imagination, because its growth depends on the entire world’s capital expenditure turning toward its direction. The challenger needs to lower other people’s risk of change, making customers believe that choosing it will not make them regret.

Position decides the way of public expression; personality only limits the scale on this way.

What she can do is, in the occasions where she must be public, give sufficient information, say the commitments accurately, then not say more. This is both personality and a strategy: when a company has repeatedly changed direction in the past and commitments could not be fulfilled, the most powerful language of a new leader is the record that has been fulfilled, not a pretty vision.

Final Chapter · As Technology Grows More Powerful, Do We Grow Wiser?

Lisa Su’s story usually ends with a set of numbers: stock price from $1.61 to market capitalization of several hundred billion dollars, from server share under 1% to re-entering competition, from being sentenced to death by Wall Street to sitting at the main table of the White House state dinner. Such an ending suits commercial legend. But it cannot answer the question this book raised at the beginning.

What Did She Defeat

Return to the original question: what exactly did Lisa Su defeat?

The first layer, what she defeated is a company’s product crisis and organizational disorder. This part has been written a lot earlier: abandoning the low-end market, protecting architecture reconstruction, bearing three years of looking bad, getting the few things done. These are achievements that can be verified—products launched, finances improved, customers returned.

The second layer, what she challenged is short-termism. The significance of this layer is more universal than the first. In the semiconductor industry any truly competitive product requires three to four years, while the performance pressure of every listed company is calculated quarterly. There is a structural conflict between the two. Her choice—within a formal, institutionally recognized range, refusing to change long-term decisions for short-term reasons—is not a personality advantage, but a system design: she drew a protected area for long-term investment that is not allowed to be eroded by quarterly pressure.

The third layer, she represents a mode of action in the face of difficulty. Run toward the problem, not around it. This line from her taking apart the remote-control car at age ten, to choosing electrical engineering at MIT, to taking over a company that was about to die, has actually never changed.

But any personal victory has boundaries.

She relies on the team—Zen is the work of hundreds of engineers, not one person’s epiphany. She relies on the industrial ecosystem—TSMC’s process, customers’ procurement decisions, developers’ choices. She relies on the era window—Intel’s mistakes, the explosion of AI capital expenditure. She has also made choices worth controversy, some of whose consequences will only be seen clearly much later.

A real person has both courage and is limited by conditions; both bears responsibility and cannot foresee all consequences. Placing these three layers of achievement and these three layers of limitation together is the complete presentation of her.

A Larger Picture

Now pull the lens back.

Placing Lisa Su together with Jensen Huang, Musk, and Bezos against the same era background, one sees something in common.

They are all willing to invest capital, reputation, and years of time in things that have not yet been proven.

Musk in 2002 bet almost his entire fortune on a rocket company that almost no one believed could succeed, and later made the first liquid-fuel rocket developed by a private enterprise to enter Earth orbit. Bezos used private capital, spent twenty-six years, invested about $28 billion to do Blue Origin; the gist of what he himself has said is: because I do not know whether this thing can succeed, the risk is really too high, so I am willing, before seeing hope, to carry it alone.

These people no longer lack money. Even if they earn more money, quality of life will not improve further. So what they pursue is obviously another thing: breaking through some boundary previously considered impossible, proving it is possible.

This spirit has a certain similarity to the pursuit of excellence spoken of by the ancient Greeks. It is not an optional personal hobby, but a mechanism by which civilization maintains vitality. If this spirit is removed, society will begin to live dependent on existing achievements and ultimately lose the ability to create the future.

Technological civilization needs this force. This point must be made clear first.

But Excellence Does Not Automatically Equal Good

If this chapter only writes to here, it becomes a praise of outstanding people.

And the real problem is behind.

Excellence itself does not automatically equal good.

Stronger chips can be used to discover new drugs, or to expand surveillance. Artificial intelligence can help people learn, or make people abandon judgment. Enterprises can push innovation through global cooperation, and may also transfer security and ethical costs to those who cannot participate in decision-making. The capability of a technology and its uses are two different questions.

The earlier $293 million deal is a concrete form of this problem. It protected a company’s survival in extremity, and also left traces in the industrial pattern ten years later. These two judgments hold simultaneously and will not offset each other.

So technological civilization cannot have only the one item of “ceaseless self-strengthening.” The wisdom of Chinese tradition has never been one sentence, but a set.

“Ceaseless self-strengthening”—subjectivity in the face of difficulty. This is the footnote of Lisa Su’s life: do not go around the problem, run toward it.

“Virtue carries all things”—moral responsibility above capability. The greater the capability, the higher the requirement to carry responsibility. This item is exactly the other half of “ceaseless self-strengthening”: it does not deny advancement; it requires the advancer to bear the consequences of advancement.

“Think of danger in times of peace”—self-restraint in favorable circumstances. The most dangerous moment is not necessarily when resources are scarce; it may also be when external applause is loudest.

“Extremes reverse, fullness declines”—this item is often read as fatalism. It is not. It is a warning against complacency, not a negation of success. The yin-yang mutual containment in the Taiji diagram is not saying “good days will always end,” but “when you are at the extreme of prosperity, the seed of change is already inside.” The value of this cognition is that it requires people to remain alert even at the best time.

These four items together constitute something more complete than “success studies.” Success studies only teach people how to win; “ceaseless self-strengthening” plus “virtue carries all things” asks what to do after winning.

That Question That Refuses to Disappear

Finally return to the question that was raised in the preface of this book but has not been directly handled until now.

When technology makes life increasingly convenient, are people still willing to train judgment, bear risk, and solve truly difficult problems?

From the technical angle this conception is possible. But from the human angle it brings an unsettling question: if a person no longer needs to labor for survival, will he lose something?

The worry is concrete.

If a society no longer encourages people to bear responsibility, face risk, and challenge limits, then even if it possesses the most advanced technology, it may not be able to produce again people willing to challenge the impossible.

This worry is not a negation of abundance, but a suspicion of “what abundance itself can bring.”

It needs to be stated: this belongs to the author’s thinking, not a scientific conclusion. Regarding the impact of artificial intelligence on labor and human capability, the research community has no consensus, and this book should not give a seemingly definite judgment. Raising this question is because it truly exists, not because it has an answer.

What Is Lisa Su’s Answer

If one must find on her a kind of response to this question, it may be this.

Her value does not lie in providing a set of success formulas that everyone can copy. Such formulas usually do not hold: her family, era, industry window, team, and luck cannot be replicated.

Her value lies in letting people see one thing: difficulty sometimes does destroy people, but people, under the support of correct education, system, and sense of responsibility, can also use difficulty to complete self-shaping.

Every qualifier in this sentence is intentional.

“Correct education”—not preaching, but letting the child calculate first herself, allowing her to take things apart, while requiring her to put them back.

“System”—not personal will, but whether an organization rewards people who report bad news, whether it allows long-term investment not to be eroded by short-term performance, whether it makes people believe that the ugly effort of today has meaning. The three concentrations she did are essentially creating such a system inside the company.

“Sense of responsibility”—not a moral slogan, but the substance of the sentence her mother once said: “One day you will come and take over my company.” Not “you must be excellent,” but “you must be responsible.”

None of these three things is talent; all can be designed, given, and required. This may be the only answer this book can give that is not quite an answer.

The Last Sentence

Artificial intelligence may replace more and more labor, and even part of everyday thinking.

At that time, what human beings most need to protect may not be the speed of doing a certain problem, but three things: choosing problems worth solving, bearing the consequences brought by the choice, and still being willing to try honestly when the world says it is impossible.

Whether technology can make civilization stronger ultimately does not depend on how much computing power there is, but on whether the people who master this computing power are also willing to make themselves more judgmental, more self-disciplined, and more responsible.

This question Lisa Su has not answered. She has finished her part; the remaining part is the matter of every person who reads to here.