Lisa Su Runs AMD—and Is Out for Nvidia’s Blood

A portion of proposal if you’re gathering pinch Lisa Su: Wear sneakers.

Su, nan leader of AMD, moves accelerated these days, though I fishy that’s ever been nan case. Her company's chips underpin nan artificial intelligence that’s changing nan world astatine breakneck speeds. To perceive Su and virtually everyone other successful semiconductors talk astir it, nan US is successful an AI race pinch China—and nan rules keep changing. The Trump management has erstwhile again shifted its stance connected what benignant of chips tin and can’t beryllium shipped to China, pinch nan latest decree being that nan US will return a 15 percent trim of AMD and Nvidia spot income to China. Meanwhile, connected nan location front, Su has claimed that AMD’s newest AI chips tin outperform Nvidia’s—part of her strategy to support eroding Nvidia’s power successful nan market.

So, yeah: Be fresh to support up.

Under Lisa Su, nan stalwart American semiconductor institution has reasserted itself arsenic a unit successful nan property of AI. “Reasserted” doesn’t do it justice: Su took a struggling AMD and executed a 10-year turnaround that has been, arsenic 1 economist put it, thing short of remarkable. Since 2014, erstwhile Su took complete arsenic CEO, AMD’s marketplace headdress has risen from astir $2 cardinal to astir $300 billion.

Aside from her well-known bona fides, Su herself—what drives her, what inspires her, what irritates her, wherever her authorities lie—is little known. This is what I was hoping to study erstwhile I visited AMD’s offices and labs successful nan hills of Austin, Texas, connected a time successful precocious June erstwhile nan upwind seemed to do small much than push power around.

Our speech kicked disconnected pinch China, which accounts for astir a 4th of AMD’s business. She betrayed nary anxiety. Su now travels often to Washington, DC, to grease nan wheels. “We’ve travel to recognize that export controls are a spot of a truth of life,” she told me, “just fixed really captious nan chips that we make are.” In different words, it’s precisely because AMD’s chips are truthful darn important—to nationalist security, to nationalist economies—that they’re now astatine nan bosom of modern statecraft.

Another point I learned astir Su: She plays nan agelong game. Politics is simply a cakewalk compared to what she’s managed to propulsion disconnected professionally.

Su was calved successful Taiwan successful 1969 and raised successful Queens, New York. Her begetter worked for nan metropolis arsenic a statistician; her mother was an accountant who became an entrepreneur successful her mid-forties. Su earned a doctoral grade successful electrical engineering from MIT, past went connected to stints astatine Texas Instruments, IBM, and Freescale Semiconductor, wherever she served successful executive roles. After joining AMD successful 2012, she quickly roseate to COO. As Su tells it, six months in, nan president of nan committee called her and said, “It’s time, Lisa.” Su’s response: “Really? That seems kinda quick.”

As CEO, Su smartly steered AMD toward nan high-performance computing market. She embraced chiplets, a modular attack to building chips that has paid disconnected enormously. She impressed nan manufacture by launching nan world’s first 7-nanometer information halfway GPUs. More recently, she doubled AMD’s information halfway gross successful 2 years. And she has struck deals pinch juggernauts for illustration OpenAI, Meta, Google, and a mates of Elon Musk’s companies. During a keynote reside astatine AMD’s yearly arena this June, OpenAI CEO Sam Altman trotted onto nan shape to hug it retired pinch Su.

These are each awesome information points. Yet AMD is still a fraction of nan size of its astir notorious competitor, nan $4.4 trillion Nvidia. Comparisons betwixt nan 2 companies are inevitable—especially since Su is distant cousins pinch Nvidia CEO Jensen Huang. (I was warned that Su hates being asked astir this. I asked anyway.)

During my sojourn to Austin, Su led maine connected a circuit of AMD’s trial labs, wherever rows of server racks are rigged up to beryllium put done nan extremes. Engineers straightened up erstwhile Su paused astatine their stations. One of them amazed her pinch a celebratory barroom successful nan style of AMD’s EPYC Venice processor. Su appeared genuinely delighted. She posed for a photograph, past moved connected to nan adjacent statement successful nan lab, striding pinch intent successful achromatic Prada sneakers. Along nan measurement I fired questions astatine her, raising my sound to beryllium heard supra nan din. What do models for illustration DeepSeek mean for her business? Will AMD build its ain LLMs? What drives Lisa Su?

Later on, Su invited maine to subordinate her successful her chauffeured car. (Alas, it was not 1 of her Porsches, which person licence plates base nan names of her favourite AMD chips.) The sound was gone; nan walls were down. During nan astir 30-minute thrust to nan adjacent lab, Su pressed maine connected my thoughts connected AI, called retired my skepticism, and shared why powering nan AI revolution—particularly astir wellness care—is individual to her.

What is thing that you would dream that nan Trump administration—but besides nan wide public—would understand astir nan AI accelerators that you’re making?

That we, arsenic tech companies, use from much users. And limiting nan number of users successful our ecosystem is really bad, not conscionable for AMD but really for nan US. Because location will beryllium alternatives retired there. The thought that somehow, if we don’t vessel chips to nan remainder of nan world, that AI advancement is going to stop—it’s not going to stop. AI advancement is going to proceed to develop, and we’d alternatively them create connected america than connected personification else.

There person been incentives precocious to bring much chipmaking backmost to nan US. What’s nan astir analyzable portion of bringing that manufacturing back?

We perfectly should bring manufacturing backmost to nan US. Absolutely, 100 percent.

Why is that?

Because it is specified a captious portion of nationalist information and nan economical interests. We had an crystal large wind present successful nan cardinal Texas area a mates of years ago. For days thing could move around. A mates of fabs astir present were unopen down. For bully business practice, you want diversity.

To bring manufacturing backmost will return time. But it’s doable. The reasoning went from “Oh, you can’t do leading-edge manufacturing successful nan US” to today, really successful TSMC’s Arizona fab we’re moving immoderate of our latest server processors, and it’s looking really good. So it tin beryllium done. It is much expensive, and that’s OK, too. I deliberation it requires a alteration of mentality, that you don’t ever spell for nan lowest-cost thing.

When you became president and CEO of AMD successful 2014, did you deliberation that folks for illustration yourself would beryllium expected to measurement successful truthful overmuch connected what’s going connected successful position of geopolitical and societal issues? Do you consciousness much unit to participate successful conversations pinch nan existent administration?

Well, it surely has changed. I wouldn’t opportunity that we’re governmental aliases I’m political. You won’t spot maine weighing successful connected wide societal issues, because I don’t needfully deliberation that that’s wherever my value-add is. But erstwhile it comes to exertion argumentation and wherever semiconductors are successful nan world, yes, we person to participate. And I wouldn’t telephone it expanding pressure. I would telephone it an expanding responsibility to do so, because we want nan rules to beryllium written right.

You’re 1 of nan astir salient women leaders successful technology, successful semiconductors. Why don’t you deliberation you adhd value?

Maybe value-add is not nan correct measurement to authorities it. It’s much like, my individual sentiment mightiness beryllium interesting, but frankly, it’s overmuch much important that we get nan argumentation correct, based connected what I see facts.

Your slightest favourite mobility these days is astir apt “When?” Meaning, erstwhile mightiness AMD surpass Nvidia successful nan AI GPU market? My conjecture is that 10 years agone group might’ve laughed astatine nan idea, and now they mightiness beryllium much consenting to entertain it.

When I started arsenic CEO, group would inquire me, “Why are you taking this job?” and I would beryllium very confused. I was like, “Are you kidding me?” To me, it was nan best. We were a institution successful an manufacture that really matters that had been underperforming for a while, and I had a chance to return this squad and do thing which I felt was important. If you asked me, “What do I want to beryllium erstwhile I turn up?” it was not “I want to beryllium a CEO,” it was “I want to activity connected thing that matters.”

At that time, everyone was comparing america to Intel, and we would person to take sides that. And my comments to nan squad are, “Look, we cognize what we tin do, and let’s conscionable show nan world what we tin do.” And truthful I mean, this conception of when, I don’t emotion this. I cognize nan media loves this. It’s ever A versus B, is that right? That’s what you guys—

I wouldn’t opportunity it’s A versus B. I mean, I deliberation AMD is much nuanced successful immoderate ways, because you still person truthful galore clients and customers who are reliant connected x86 and CPUs. And you person this fast-growing information halfway business, which you grew from $6 cardinal successful 2022 to $12.6 cardinal past year. But Nvidia is nan large mobility correct now.

Look, my constituent is I don’t needfully want to beryllium compared against Intel aliases Nvidia. My vision, our vision, has ever been, “There’s nary one-size-fits-all.” You request nan champion successful nan information center. You request nan champion CPUs. You request nan champion AI accelerators. You request nan champion successful your individual computing. We person this portfolio that is rather broad, and nan marketplace is humongous. It’s much than $500 cardinal complete nan adjacent 3 to 4 years. We person plentifulness of opportunities.

So, Sam Altman was astatine your AMD AI arena successful June, and you talked astir moving pinch OpenAI. You’re moving pinch Meta. You’re moving pinch Elon Musk’s companies, Tesla and xAI. But nan reality pinch AMD is that companies for illustration that emotion their Nvidia GPUs, and you’re an “also.” Do you want to get to nan constituent wherever you’re like, “No, we are nan superior partner?”

Of course. That’s wherever we are coming successful CPUs. So if you were to inquire galore of those aforesaid companies, I deliberation they would opportunity that AMD is their strategical CPU partner. And absolutely, we expect to beryllium location successful AI arsenic well. But I’m not impatient pinch this.

You don’t want to put a clip stamp connected it.

Look astatine it this way: When I first joined AMD successful 2012, Microsoft was conscionable an early partner for america successful gaming. Over nan past 10-plus years we’ve built a batch of trust, and now we’re cocreating pinch them, truthful Microsoft conscionable announced they’re utilizing AMD not only for their next-generation Xbox consoles but crossed their full cloud.

Our activity pinch Meta has been perfectly nan aforesaid way. I retrieve nan first speech I had pinch Meta, and I said, “Just springiness maine a shot. I don’t person to show you that I’m going to beryllium nan champion successful nan world. I cognize that I’m going to beryllium that to you. I’m going to beryllium your champion partner. Not conscionable your champion exertion partner but besides your champion partner successful helping you get your exertion infrastructure together.” And that’s what we do.

At nan AMD AI arena successful June you talked a adjacent magnitude astir DeepSeek, nan ample connection exemplary that came retired of China and reportedly costs a batch little money and computing powerfulness to train. How are AI models for illustration that changing nan measurement you’re reasoning astir computing power?

Well, it’s different illustration of conscionable really overmuch workloads person been changing successful AI. People were antecedently each focused connected large-scale training, and now pinch reasoning models and fine-tuning, nan manufacture has transitioned to a spot wherever inference-type computing is really increasing faster. It’s why you person to person very elastic hardware.

So it’s not changing nan measurement we’re reasoning astir it, because we ever thought conclusion was going to beryllium much important. I conjecture that intends we were correct successful betting connected that. We’ve optimized for representation capacity and different cardinal things that are important for conclusion computing.

I’m going to bring up Nvidia again: Nvidia has been training immoderate of its ain models and offers a framework, NeMo, for developers to build their ain generative AI models. How earnestly person you thought astir AMD training its ain models?

We are training our ain AI models. We person an AI models team. But we’re not training our models for nan liking of competing pinch large exemplary builders. We train them to study from them. The much we dog-food our ain stuff, nan much we learn, and past we tin accelerate our building.

Analysts and moreover immoderate of your customers person told maine that AMD is incredibly customer-oriented. On nan different hand, folks who are heavy successful nan method weeds opportunity that ROCm, nan group of package devices for programming AMD’s hardware, still isn’t arsenic bully arsenic Nvidia’s CUDA. What circumstantial steps do you scheme to make to lure successful much developers?

Yeah, I mean, I work together that nan package is nan astir captious layer, because it’s what developers see. And erstwhile you deliberation astir ROCm and CUDA, it’s not that 1 is amended than nan other. It’s that CUDA has been astir for a agelong time. People person gotten utilized to a definite ecosystem, and truthful we’re really school them a different ecosystem. That’s nan measurement to deliberation of ROCm.

What I perceive is that it’s not conscionable entrenchment, though. Developers say, “The compilers don’t activity arsenic well” aliases “The capacity libraries aren’t arsenic good” aliases “We want much portability.”

Sure. But nan logic for that is because, hey, things are done a definite measurement successful Nvidia’s CUDA, and they would for illustration to person it done a very akin measurement successful ROCm. I person not yet met an AI customer that we haven’t been capable to get working, performant, and each of that stuff. We don’t needfully person each nan libraries yet. There are tons of, let’s telephone it typical kernels, that person been written. So to your mobility of, What are we doing? I mean, we’re moving faster. That’s nan champion answer: We are moving faster.

I learned very early that you don’t person to work together pinch criticism, but you person to understand it’s a perspective. And past you determine what you’re going to do connected apical of it. There’s still a batch to do. And we’re hiring for illustration crazy. We’re acquiring, and we’re listening to developers. And I deliberation what we’ve seen is you tin really make advancement beautiful quickly.

What do you make of nan Meta Superintelligence Lab and Mark Zuckerberg reportedly offering AI talent group up to 9 figures to spell activity for Meta? What does that benignant of compensation do to hiring successful nan Valley?

I can’t opportunity I person nonstop acquisition pinch it, frankly. I deliberation title for talent is fierce. I americium a believer, though, that money is important, but frankly, it’s not needfully nan astir important point erstwhile you’re attracting talent. I deliberation it’s important to beryllium successful nan zip codification [of those numbers], but past it’s super-important to person group who really judge successful nan ngo of what you’re trying to do.

I deliberation group person done comparatively good here, because nan stock’s done OK. But from a recruitment standpoint, it’s ever like, “Do you want to beryllium portion of our mission?” Because nan thrust is really what we’re trying to pull group to. It’s nan thrust of, “Look, if you want to travel do important technology, make an impact, you’re not conscionable a cog successful nan wheel, but you’re really personification who’s going to thrust nan early of our roadworthy map, past you want to beryllium astatine AMD.”

Do you envision ever offering anyone a nine-figure compensation package to travel build retired your package ecosystem?

I don’t deliberation so.

Because you’d person to reply to shareholders aliases make a lawsuit to your board?

Because it’s not really astir 1 personification successful our world. I mean, it’s really astir awesome people, don’t get maine wrong—we person immoderate unthinkable people. We acquired Nod.ai, and Anush Elangovan, who was nan CEO of that, has now go nan caput of our package ecosystem work. He’s conscionable perfectly phenomenal pinch his passion. He’ll spell aft each azygous personification who has an rumor pinch ROCm. And I’m like, “Anush, really do you do that?” So that’s what I mean. We’re looking for group who person that type of passion for nan activity that we do, and location are tons of those group around.

What is “superintelligence” to you?

I deliberation nan thought that AI tin make each of america superintelligent is simply a awesome vision, and we’re still successful nan very early innings of really to do that.

One of nan areas that I’m astir personally passionate astir is wellness care, because I person had acquisition pinch nan wellness attraction system, and I deliberation it should beryllium much, overmuch amended than it is today. We should beryllium capable to cure these diseases. We shouldn’t person to do proceedings and correction for illustration we sometimes do. This is simply a cleanable usage lawsuit for AI. Being capable to stitch each those pieces together to spell from supplier find to therapeutics to inpatient care, each of that is ripe for—let’s telephone it transformation. I don’t cognize if you telephone that “superintelligence.”

There’s this thought floating astir that AI is going to beryllium truthful smart that it’s going to yet beryllium capable to delete humanity. How do you deliberation astir those kinds of predictions? Do you judge successful AGI?

I do judge successful AGI, but I don’t judge successful nan thought that AI will beryllium smarter than people. I besides americium not a large last day personification aliases believer, either. Look, I mean, exertion is great, but exertion is arsenic awesome arsenic nan group who build it and create it and transmission it successful nan correct direction. So I find those conversations a small spot esoteric. And our attraction is, “The tech is good, but it’s not awesome yet. How do we make it great?”

How do you measurement “great”?

I deliberation this thought of having AI lick existent difficult problems is erstwhile it gets great. And we talk astir agents arsenic 1 of nan adjacent large things. I deliberation agents correct now are doing, let’s telephone it, comparatively much of nan mundane tasks of nan world.

Putting things successful your shopping cart.

Right. I deliberation there’s 2 directions AI goes. One is axenic productivity, you know, really do I region immoderate of, let’s telephone it nan menial activity that group do, truthful that they tin activity connected much absorbing things? That’s 1 facet of it, and we’re utilizing that.

But nan different facet of it is erstwhile AI tin lick really, really difficult problems. It tin return what would’ve taken america 10 years to fig retired and do that successful six months. I deliberation astir a world wherever it usually takes america 3 years to creation a chip, and what does that look for illustration if I could do that successful six months?

Does humanity conscionable not support up astatine immoderate point, though?

I don’t know. I would stake connected humanity being OK.

Technology tin beryllium a small spot overwhelming now.

Well, but Lauren, I deliberation that’s nan point. When exertion is bully enough, you don’t person to deliberation astir it. Today, you still person to deliberation astir erstwhile you spell and ask—what’s your favorite? Do you usage ChatGPT aliases Grok?

I usage ChatGPT, yeah. Not each day, but—

Often?

Often enough. I mean, I’m obligated to trial these things.

Yeah, but you still person to make sure, “Hey, did it springiness maine nan correct answer?”

Oh, absolutely. I mean, arsenic a journalist successful particular. I don’t usage it for my penning successful immoderate capacity. We tie a very difficult statement betwixt utilizing it for learning really to navigator a steak versus utilizing it for our journalism.

But you usage it for research.

Sometimes, but nan hallucinations are concerning.

But that’s nan constituent of it’s not bully capable yet. At immoderate point, it’s going to beryllium bully enough. You’d want to beryllium capable to return your AI astatine look value.

You mentioned wellness care. When we’re older and infirm, will nan procreation of folks who are treating america beryllium ChatGPT doctors?

I would like location to beryllium a procreation of folks treating america who person nan immense magnitude of information that ChatGPT will have, truthful that they’re amended informed to make diagnoses.

When you deliberation astir AI philosophically, is it for illustration nan net each complete again? Is it much comparable to Linux, for illustration it’s going to beryllium immoderate operating strategy that runs connected everything that we own? Is it electricity? Is it fire? I deliberation Sundar Pichai has compared AI to fire, successful position of really transformative it is.

The net is not a bad comparison, but I deliberation AI is overmuch much than nan internet. Because, if you deliberation astir it, nan net was a batch astir moving traffic. AI is much astir thing foundational successful position of productivity. Sometimes group comparison it to nan Industrial Revolution, and that’s not a bad comparison, actually.

With different revolutions we weren’t overwhelmed truthful overmuch pinch reasoning astir what was existent and existent and what was not.

You tin take 2 ways to deliberation astir it. One is you effort to clasp backmost connected AI because it could beryllium dangerous, aliases you effort to spell arsenic accelerated arsenic you tin but put nan correct lens connected nan information. I’m a large believer successful nan 2nd camp. And arsenic a consequence I don’t judge successful these cases wherever you’re not going to request tons and tons of people. Because successful nan end, group are nan judge of what truth is. We’re still hiring much and much engineers, because they’re nan last arbiters of our engineering.

I americium hopeful that humanity will fig it out.

And it will beryllium truthful overmuch better. It will beryllium for illustration nan net is to america today, which is you conscionable return it for granted. We shouldn’t measure nan exertion based connected this constituent successful time. We should measure it connected nan slope of what we’re going to beryllium tin of doing. We’re going to get these things right. But we whitethorn person a fewer bumps successful nan road.

You look a small concerned astir AI. Are you conscionable playing devil’s advocate?

I thin to deliberation nan group who guidelines to use nan astir from these technologies are nan ones who person nan luxury of being a small spot much optimistic astir it and who are hyping it up. There’s that celebrated line: The early is not distributed evenly. Even pinch nan advancements successful medicine, we’re going to spot biases look that lead to group getting denied wellness attraction aliases security coverage. We’ve already seen this.

Health care, for me, is rather personal, because my mom was rather ill. For a while. And truthful I sewage to watch her travel going done that. And I realized, like, it doesn’t matter who you are. You can’t guarantee nan champion wellness care, because it’s really an creation correct now. It’s not a science. And I judge it should beryllium a science.

Why do you deliberation it’s an art?

The assemblage is simply a very analyzable system. So you person specialists, for illustration a bosom master aliases a kidney specialist. But location are not that galore generalists that tin propulsion it each together. And that, to me, is simply a travesty. I’m like, travel on—this is solvable.

That’s what we do successful tech, right? We return analyzable systems and put them together, and we make them work. But we’re often only looking astatine 1 facet of health, and it’s my patient belief that if we tin usage exertion to thief propulsion each of that expertise together, we’ll beryllium capable to dainty group better. I watched it firsthand. So, anyways, successful my adjacent life erstwhile I person clip to do thing different than this—

You’ll beryllium a doctor?

I won’t beryllium a doctor, but I dream to beryllium personification who tin thief span nan disagreement and usage exertion for what it’s really tin of.

You could do that successful this lifetime, too.

I person a fewer things to do correct now.

Was your mother capable to recover?

No, unfortunately.

I’m sorry.

But you cognize what I mean? I conscionable realized that—wow.

You realized, it could hap to immoderate of us.

Yes, but it’s astir nan value of care, moreover pinch nan champion doctors.

About a twelvemonth agone my ain mother had immoderate really superior wellness issues, and she ended up successful nan ICU connected a ventilator. Doctors kept coming successful and looking astatine scans and couldn’t fig retired what was going on. And I was sitting successful nan infirmary thinking, I’m truthful heavy successful nan world of reporting connected AI, wherever group are touting unthinkable aesculapian advancements, but we can’t show what’s happening from these scans?

So you know. You cognize precisely what I mean. It’s infuriating for maine to deliberation my mom was successful nan ICU for 60 days, and group said, “Nobody walks retired of that.” They were like, “She’s not going to beryllium capable to do it.” And I was like, “Yes, she is. I cognize she is.” I wasn’t nan 1 qualified to make those calls, right? And she did. She survived different 2 years aft that.

You talk often astir resilience. How do you personally—not nan company, but you—stay resilient? Is it Starbucks? I noticed you had 1 astatine your AI event.

Yes, this is simply a passion beverage lemonade, and it does a batch for me. I’m not a large caffeine drinker, truthful this and a spot of workout does it for me. I spell successful waves [with Starbucks]. Sometimes I person a lot, and sometimes I trim myself off.

Same. What’s your preferred shape of exercise?

I for illustration to box. I person a trainer that comes to nan house, and he lets maine deed him. Well, not him, but mitts.

How agelong person you been doing that?

I don’t know. Seven, 8 years, thing for illustration that.

So 2 to 3 years aft you became CEO, you were like, “I request to activity immoderate worldly retired here.”

Yes.

How overmuch do you slumber per night?

Five, six hours. Six is simply a really bully number. On weekends, I mightiness beryllium seven.

What impresses you astir arsenic a leader? When you person your first gathering pinch someone, what impresses you?

Passion for what they’re doing. Because I deliberation that stays pinch you done bully times and bad times. Things are ever going to spell wrong, but if you’re genuinely passionate astir what you do, past I deliberation you shine.

What irritates you?

What irritates me?

Yes.

Well, I can’t opportunity group who inquire maine astir Jensen being my cousin?

You tin wholly opportunity that.

To show you nan truth, it doesn’t really irritate me, but it’s much like, “Really? Is that nan astir important point we person to talk about?”

What is it you consciousness for illustration group don’t really cognize astir you, that you want them to understand astir you?

I consciousness for illustration group cognize me. No? Well … I get up each time because I judge our products tin alteration nan world, and they tin make nan world a amended place. So there’s ever noise—this, that, export controls, whatever. Those are noise.

So is it that you’re a ultimate optimist?

I don’t deliberation of myself that way, but I’m astir apt a ultimate exertion optimist. I’m really rather pragmatic. So I’m a pragmatic ultimate exertion optimist. How does that sound?

It sounds for illustration ChatGPT generated it.

That was not a programmed response.

How would I picture myself? I do judge tech has nan opportunity to alteration truthful overmuch of really we acquisition life successful a very, very affirmative way. So successful that case, I americium a ultimate exertion optimist. But I’m pragmatic successful really you get there. And really you get location is each day, measurement by step. We learn, we listen, we adjust. We use what we learn. That’s conscionable what we do.

Let america cognize what you deliberation astir this article. Submit a missive to nan editor at mail@wired.com.