Cash, chips and tal­ent: In­side Nvidi­a's plan to dom­i­nate biotech's AI rev­o­lu­tion

Over the last decade, one tech gi­ant af­ter an­oth­er has tried to get in­to health­care and failed. Ama­zon’s at­tempt to re­shape costs and de­liv­ery sput­tered in­to noth­ing, IBM’s Wat­son Health end­ed in dis­ar­ray, and Al­pha­bet’s well-fund­ed projects like Ver­i­ly and Cal­i­co have yet to pro­duce any­thing near the im­pact of its core tech busi­ness.

So why has Nvidia suc­ceed­ed?

The chip­mak­er has be­come the dom­i­nant en­ti­ty in bio­phar­ma R&D’s hottest area — us­ing ar­ti­fi­cial in­tel­li­gence to de­sign drugs. It’s built a $1 bil­lion-plus rev­enue stream in health and ap­pears to be win­ning the ar­gu­ment that this may fi­nal­ly be tech’s health­care mo­ment — al­beit not in the way past chal­lengers thought it would be.

Along the way, Nvidia’s leather-jack­et-wear­ing, vi­sion-sling­ing CEO Jensen Huang has turned in­to bio­phar­ma’s high­est-pro­file pitch­man, sell­ing a fu­ture of bi­ol­o­gy dig­i­tized. In ad­di­tion to the cov­et­ed H100 chips at the heart of its busi­ness, it al­so sells soft­ware, let­ting drug com­pa­nies crib off its own AI tal­ent. And through its cor­po­rate ven­ture cap­i­tal arm, it’s in­vest­ed in many of biotech’s newest AI star­tups.

All of which has fur­ther in­ter­twined Nvidia with drug de­vel­op­ers large and small, adding fu­el to Huang’s vi­sion of a grow­ing por­tion of phar­ma R&D work flow­ing through AI — and a grow­ing por­tion of phar­ma R&D bud­gets flow­ing through Nvidia.

Sean Mc­Clain

For every be­liev­er, the com­pa­ny has a share of de­trac­tors. Some biotech vet­er­ans see an op­por­tunist over­sim­pli­fy­ing bi­ol­o­gy to sell more chips. Oth­ers see a vi­sion­ary haul­ing a tech-lag­ging in­dus­try stuck on a 90% fail­ure rate in­to a bet­ter fu­ture.

“Nvidia is ab­solute­ly key to en­sur­ing that AI is ul­ti­mate­ly suc­cess­ful in health­care,” Ab­sci CEO Sean Mc­Clain, whose com­pa­ny us­es Nvidia’s chips and soft­ware, said in an in­ter­view. “Even if we have our mod­els, if Nvidia does not ex­ist, AI is not go­ing to make the im­pact in health­care at the end of the day.”

From video games to pro­teins

Nvidia used to be a video game hard­ware com­pa­ny. It pop­u­lar­ized the idea of the GPU — or graph­ics pro­cess­ing unit — in 1999, which quick­ly turned in­to a com­pet­i­tive race with AMD to de­sign the best chip. What pushed Nvidia to the top was soft­ware. It de­signed a new pro­gram­ming lan­guage, called CU­DA, which turned its chips from video-game ren­der­ers in­to gen­er­al-com­put­ing pow­er­hous­es. CU­DA al­lowed Nvidia’s chips to break com­put­ing tasks in­to small­er chunks and then si­mul­ta­ne­ous­ly work on those jobs. All of a sud­den, GPUs weren’t just for video games.

As Nvidia rolled out CU­DA in 2006, it no­ticed cer­tain sci­ence groups — not gam­ing com­pa­nies — were us­ing 40%, even 50%, of com­pute cy­cles on some of the world’s first su­per­com­put­ing cen­ters. In­stead of ren­der­ing video games, they were sim­u­lat­ing how small mol­e­cules and pro­teins moved and in­ter­act­ed.

Nvidia asked one of those users, Klaus Schul­ten at the Uni­ver­si­ty of Illi­nois, to speak at one of its first de­vel­op­er con­fer­ences in 2010. In front of at­ten­dees who most­ly came from the gam­ing world, Schul­ten guid­ed the au­di­ence through wrig­gling 3D mol­e­c­u­lar an­i­ma­tions, show­ing how these sim­u­la­tions an­swered ques­tions like how the swine flu virus re­sist­ed the an­tivi­ral drug Tam­i­flu. But the ap­pli­ca­tions were lim­it­ed, held back by cost and com­plex­i­ty.

Then, in late 2018, Google Deep­Mind pro­vid­ed a break­through when its mod­el, called Al­phaFold, won a long-run­ning pro­tein struc­ture pre­dic­tion com­pe­ti­tion. In a cou­ple of years, Al­phaFold was pow­er­ful enough to turn hun­dreds of mil­lions of strings of amino acids in­to pre­dict­ed pro­tein struc­tures.

“The Al­phaFold mo­ment con­vinced us this was pos­si­ble,” Kim­ber­ly Pow­ell, Nvidia’s vice pres­i­dent of health­care, told End­points News in an in­ter­view in San Jose.

Kim­ber­ly Pow­ell

The chip­mak­er had al­ready spent years in bi­ol­o­gy’s weeds, in ar­eas like mol­e­c­u­lar dy­nam­ics, ge­net­ic se­quenc­ing, and cryo-elec­tron mi­croscopy. By turn­ing amino acid se­quences in­to high­ly ac­cu­rate 3D struc­tures, Al­phaFold led Pow­ell to imag­ine what else could be pos­si­ble.

“Once you can rep­re­sent some­thing in a com­put­er, you can start build­ing mod­els,” Pow­ell said. “Once you have mod­els, you can start learn­ing more about their in­ter­ac­tions. This, to me, is a new dawn of com­put­er-aid­ed drug dis­cov­ery, mov­ing in­to many more ar­eas than just the mol­e­c­u­lar dy­nam­ics sim­u­la­tions.”

In fall 2022, Nvidia launched BioNeMo, a soft­ware ser­vice of­fer­ing ac­cess to dozens of AI mod­els, in­clud­ing Al­phaFold. Oth­er mod­els can sim­u­late mol­e­c­u­lar dock­ing with Diff­Dock, gen­er­ate small mol­e­cules with MolMIM or pro­teins with RFd­if­fu­sion, or pre­dict how strong­ly a mol­e­cule binds to a pro­tein with Neu­ralPLex­er. Over 100 bio­phar­ma cus­tomers are now us­ing BioNeMo, and Pow­ell sees this as just the start.

Hard­ware re­mains the heart of Nvidia’s busi­ness. It’s cur­rent­ly build­ing su­per­com­put­ers for the No­vo Nordisk Foun­da­tion in Den­mark and for Am­gen in Ice­land — tap­ping in­to the lat­ter’s mas­sive ge­net­ic trove through its de­CODE ge­net­ics sub­sidiary. But its soft­ware, tai­lored to the drug in­dus­try to put that com­pute in­to prac­tice, is be­com­ing a huge fo­cus.

Philipp Lorenz

“They know the in­dus­try so well that I nev­er talk to them about chips, which is kin­da weird,” said Philipp Lorenz, chief tech­nol­o­gy of­fi­cer of Base­camp Re­search, a Lon­don-based start­up gen­er­at­ing a mas­sive data­base of pro­tein se­quences from all types of or­gan­isms.

Part of Nvidia’s soft­ware play is based on the fact it has some of the top AI tal­ent that phar­ma can’t hire. A mix of pay, cul­ture, and re­sources has lured the best en­gi­neers to rich­ly fund­ed AI star­tups like Ope­nAI and An­throp­ic or to lega­cy tech gi­ants like Google and Meta, said Dy­lan Reid, a ven­ture cap­i­tal­ist at Zetta Ven­ture Part­ners. Drug­mak­ers are “not get­ting close” in try­ing to at­tract top AI tal­ent now, Reid said.

Nvidia fills that gap, with its own en­gi­neers tak­ing the lead on op­ti­miz­ing, train­ing and fine-tun­ing mod­els for drug com­pa­nies.

“These phar­ma com­pa­nies, they’re not Ope­nAIs,” Pow­ell said. “That’s not the kind of peo­ple that they have on staff.”

In its lat­est of­fer­ing, Nvidia in­tro­duced what it calls mi­croser­vices last month. These are AI mod­els ready to use out of the vir­tu­al box, with Nvidia charg­ing $4,500 per GPU per year or $1 per GPU per hour to use these ser­vices. Nvidia says a phar­ma com­pa­ny can start us­ing these mod­els with­in min­utes, re­quir­ing no AI ex­per­tise of its own. It lets drug com­pa­nies do what they know best, while Nvidia pro­vides its core strength — com­pute pow­er and en­gi­neer­ing know-how.

“Jensen is con­tin­u­ous­ly think­ing about the fu­ture and will make in­vest­ments in the present even if it costs him bot­tom line to make the fu­ture bet­ter,” Vi­jay Pande, a biotech VC at An­dreessen Horowitz, said in an in­ter­view. “That’s why he is where he is now, and why Nvidia is where it is now.”

Huang opened this year’s Nvidia de­vel­op­er con­fer­ence with a two-hour talk in front of over 10,000 at­ten­dees packed in­to the SAP Cen­ter in San Jose. “Un­be­liev­able new ca­pa­bil­i­ties will be in­vent­ed,” Huang said in pitch­ing AI as an in­dus­tri­al rev­o­lu­tion span­ning vir­tu­al­ly every in­dus­try.

And he has com­pared the pos­si­ble fu­ture of drug dis­cov­ery to the evo­lu­tion of de­sign­ing chips, which went from a high-er­ror rate, com­pli­cat­ed process, to in sil­i­co de­sign with vir­tu­al­ly no er­rors.

“Bi­ol­o­gy is a bil­lion times hard­er than physics-based in sil­i­co de­sign of a chip, which is say­ing a lot. But I think it’s in­evitable that we’re go­ing to get there,” Re­cur­sion CEO Chris Gib­son said in an in­ter­view. “It’s just a ques­tion of who, what, when, and how.”

Chris Gib­son

These kinds of lofty hopes have cre­at­ed plen­ty of doubters, es­pe­cial­ly among in­dus­try vet­er­ans who have seen sim­i­lar Sil­i­con Val­ley dreams fiz­zle be­fore.

Kei­th Horn­berg­er, a vet­er­an med­i­c­i­nal chemist, de­scribed Huang’s view as “wild­ly op­ti­mistic” and “hope­less­ly naive.” Even ar­dent back­ers of AI in bi­ol­o­gy have called for the field to tamp down the rhetoric, such as Schrödinger’s CEO warn­ing of “dan­ger­ous” lev­els of AI buzz last sum­mer, or in­sitro CEO Daphne Koller call­ing the cur­rent hype “po­ten­tial­ly de­struc­tive.”

Ul­ti­mate­ly, the on­ly met­ric that tru­ly should mat­ter is how many re­al med­i­cines are de­vel­oped — a mile­stone the field hasn’t yet cleared. Re­cur­sion ex­pects its first Phase 2 read­outs lat­er this year, with its CEO com­par­ing the mo­ment to SpaceX fir­ing off its first rock­ets.

“Peo­ple are launch­ing things,” Gib­son said. “There are go­ing to be fail­ures, but ul­ti­mate­ly, we’re all fig­ur­ing out how to put the pieces to­geth­er, and when it starts work­ing in 12 months or 36 months or what­ev­er, a few years, it’s go­ing to be an en­tire­ly dif­fer­ent in­dus­try.”

The gold rush

All the while, Nvidia keeps grow­ing deep­er roots. An­drew Gos­tine, CEO of a hos­pi­tal au­toma­tion start­up called Ar­tisight, said his com­pa­ny has tried oth­er chip­mak­ers, even as Nvidia has in­vest­ed in Ar­tisight, but found the ef­fort to op­ti­mize mod­els and fit them in­to the start­up’s ecosys­tem a “huge pain.”

An­drew Gos­tine

“The amount of mon­ey I would have to spend to buy AI tal­ent to do that op­ti­miza­tion, Nvidia gives it to you for free — if you buy their su­per-ex­pen­sive GPUs,” he said while pre­sent­ing at last month’s Nvidia de­vel­op­er con­fer­ence.

Ar­tisight is one of more than 20 com­pa­nies that Nvidia’s two-year-old cor­po­rate VC arm has in­vest­ed in — an­oth­er way the chip­mak­er has deep­ened its health­care pres­ence. The group doesn’t have a set amount of mon­ey to put to work each year, or many di­rec­tives be­yond find­ing in­ter­est­ing star­tups work­ing on big goals and us­ing Nvidia’s chips.

A few months be­fore its San Jose spec­ta­cle, Nvidia joint­ly held a sep­a­rate, much small­er event at the JP Mor­gan health­care con­fer­ence. On a rainy Jan­u­ary night, biotech ex­ec­u­tives and in­vestors packed a room to hear Huang’s vi­sion of rein­vent­ing one of the most fail­ure-prone busi­ness­es on earth. He spoke of a mirac­u­lous AI rev­o­lu­tion poised to re­shape drug R&D.

“We deeply be­lieve that this is go­ing to be the fu­ture of the way that drugs will be dis­cov­ered and de­signed,” Huang said.

Ear­ly ev­i­dence is trick­ling in that Huang may be right — or right enough to keep win­ning cus­tomers. Am­gen, an ear­ly user of Nvidia’s soft­ware, says it has boost­ed the like­li­hood of an an­ti­body pro­gram reach­ing the clin­ic from 50% to 90%, while short­en­ing the typ­i­cal re­search time­line from two years to nine months. Genen­tech re­cent­ly launched a study test­ing a lung dis­ease drug can­di­date in in­flam­ma­to­ry bow­el dis­ease, based on un­ex­pect­ed con­nec­tions be­tween the two re­vealed by ma­chine learn­ing. Star­tups like Gen­er­ate:Bio­med­i­cines and Iambic are now in the clin­ic with drugs tak­ing ad­van­tage of some of these new­er AI mod­els.

In a blunt sym­bol of where the mon­ey is in biotech right now, the Jan­u­ary JP Mor­gan event was host­ed at the San Fran­cis­co Mint. The 150-year-old build­ing was con­struct­ed to serve the boom­ing needs of the Cal­i­for­nia Gold Rush, turn­ing gold nuggets in­to cash. Nvidia is ea­ger­ly bring­ing phar­ma in­to the next fren­zy.

“Please, if you have a hard time with com­pu­ta­tion or ar­ti­fi­cial in­tel­li­gence, send us an email,” Huang told the JPM crowd. “We’re here for you.”