Thrive Capital 谈 Neuralink、OpenAI、Isomorphic Labs 与 OpenEvidence

Thrive Capital's healthcare lead argues the next decade of biotech will be defined less by drug discovery models than by rebuilding the entire value chain around them — manufacturing, clinical trials, commercialization, and care delivery. Dr. Nworah Ayogu, a trained internist who helped launch Amazon Health and joined Thrive about two and a half years ago, spends most of this 70-minute conversation on the structural economics of American healthcare and the specific companies he thinks will capture it. The episode's title points at Neuralink, and Ayogu does call it "a more interesting company than SpaceX" because "the brain is vaster than space" — but that is a few minutes inside a much broader discussion of Isomorphic Labs, OpenEvidence, Fortel, Cadence, Oscar Health, and OpenAI's quiet emergence as what he calls one of the largest healthcare providers in America.

The through-line is Ayogu's claim that the frontier of science is moving faster than at any prior point, and that the binding constraint keeps shifting downstream. He walks through AI drug discovery as a sequence of problems — target biology, molecular design, off-target effects, then in vitro and in vivo validation — and argues that Isomorphic's advantage is being "bilingual": world-class AI researchers alongside chemists and biologists at the level of Lilly, Pfizer, the Broad, and the Wyss. From there he moves to the macro case: $5 trillion in annual US healthcare spend heading toward $9 trillion by the end of the decade, 20% of all jobs, the largest line item in the federal budget at six times defense — and, in his sharpest framing, roughly $40 trillion spent over the last decade with life expectancy unmoved.

What follows is a structured extraction of the episode's arguments, numbers, company mechanics, and quotable moments, organized by theme rather than by the order they arose in conversation.

AI drug discovery is a chain of problems, and the bottleneck keeps moving

Ayogu's central technical argument is that "AI for drug discovery" is not one problem but a stack of them, and that solving the first exposes the next. The first is biological risk: understanding what a target does when you hit it, and what happens elsewhere in the body. The second is molecular design — actually building a molecule that binds, complicated by three-dimensional shape changes, hidden pockets, and the industry's "undruggable" targets. The third, and the one he says is now the frontier, is predicting downstream effects: whether the molecule does what you want in a cell, in an animal, and eventually in a person.

His illustration of why this is hard is the research adage that "we've cured cancer in mice 50 times" — the best way to survive cancer is to be a mouse. Animal models are an imperfect proxy for humans, and the only true test is another human, which is why the models need data that does not yet exist at scale. He frames the current moment as two parallel data problems: data to verify the experiments you are running, and a new vertical of data to make the models good enough to predict at all.

On Isomorphic Labs specifically, Ayogu pushes back on the critique that protein folding is only one problem. AlphaFold came out in 2018, he notes, and the team has been working on the adjacent steps — in vitro testing, in vivo testing, clinical trial design — for years. He cites Demis Hassabis's stated goal of moving a process with 90%-plus failure rates and 10-to-15-year timelines toward 88–90% success rates, and argues that even a partial improvement (say, 5% to 60% success) radically changes what a company is willing to attempt. He credits ARPA-H with work on clinical trial efficiency, and warns that getting very good at preclinical work while clinical remains the bottleneck just relocates the problem.

"We often say that in a lot of ways what we're seeing now is that our problems haven't gotten smaller, the tools we have have gotten bigger."

The value-chain argument extends past discovery. Ayogu predicts the same supply-chain cascade AI produced in compute, memory, and energy will play out in biotech across manufacturing, development, trials, and commercialization — and that if new drugs arrive every three years instead of every ten, manufacturing capacity has to absorb that demand.

Isomorphic's edge is being bilingual, and Thrive's approach was persistence

Thrive's philosophy, in Ayogu's telling, is backing companies that "speak two languages fluently." Isomorphic is his cleanest example: AI researchers at the level of DeepMind, Anthropic, and OpenAI, sitting alongside chemists and biologists he rates as equal to those at Lilly, Pfizer, the Broad Institute, and the Wyss Institute. He describes an internal culture — "the Isomorphic way" — built on crosstalk between the two disciplines rather than a handoff.

The origin story matters for how Thrive won the deal. Ayogu says Josh Kushner spent significant time engaging with Hassabis and the team, making the case that Thrive knows healthcare deeply and is as passionate about the space as they are. Both sides, he says, believe a large share of AI's benefit will accrue in healthcare. Ayogu frames the relationship as one built before there was a raise to compete for, not a competitive process.

Neuralink's real moat is vertical integration, including its own pathology lab

The episode's title topic gets its fullest treatment here, and Ayogu's argument is that Neuralink's differentiation is not the implant alone but the breadth of in-house capability. The company's first patient populations were people with ALS or paralysis — he notes many are veterans injured by IEDs, often in their 30s and 40s, who lost the ability to work or play with their children. Neuralink's device lets them control a computer with their minds and operate motorized arms; a separate trial implanted the device in the vocal cortex so a non-speaking patient could speak. Ayogu's framing: "we may not be able to replace the kind of arms and legs that you've lost, but we can give you that functionality back."

The talent inventory he runs through is the substance of the section. Hardware builders at the level of SpaceX and Tesla; over 1,000 microscopic threads inside the implant to read from neurons; a purpose-built surgical robot to implant threads at that scale with precision and reliability; practicing neurosurgeons, including people he trained with at Harvard Med, co-developing the tools; battery engineers, because users wear the device all day; radiologists, because the company scans brains. And pathologists — Neuralink insources its own pathology lab, which Ayogu says he had never seen a company do. Most organizations send slides out. Neuralink brought world-class pathologists in-house because safety data on brain effects is mission-critical.

"I actually think Neuralink is a more interesting company than SpaceX. The brain is vaster than space."

He extends the point: "We don't make it to space. We don't invent silicon chips without our brain." And: "The brain is still a more efficient engine of compute than every chip that NVIDIA has ever made." On recruiting, he credits president DJ Sehgal as a compelling leader and argues that talent density is self-reinforcing — the best people want to work with the best people, and Neuralink created the category rather than entering it.

Fortel rebuilt the hearing aid around a custom chip and semantic noise processing

Hearing loss affects 1.5 billion people worldwide, and Ayogu expects that number to rise given concert attendance and constant headphone use. The stakes are larger than convenience: hearing loss is the greatest modifiable risk factor for dementia, associated with roughly a one-third increase in likelihood. His explanation is that hearing is not just sound capture but the brain's ability to dampen and select — which is why conventional hearing aids, which amplify everything, fail in restaurants and on crowded streets, driving users to stay home and compounding social isolation.

Fortel's founder Matt DeJong, previously a hardware builder at Butterfly (handheld ultrasound), spent about 15 years thinking about the problem before assembling a team roughly five years ago. Two technical bets define the company. First, a custom chip taped out at TSMC — Ayogu notes it is unheard of to tape out a first chip successfully, but the form factor demands it: small enough to be unobtrusive, all-day battery life, and fast inference for translating incoming speech. Second, semantic noise processing, an idea DeJong drew from pandemic-era Zoom noise cancellation: identify what a sound actually is — a conversation in front of you versus a passing car — and dampen accordingly.

The validation anecdotes are concrete. Beta users included Steve Martin and Henry Kravis, both described as hard judges of quality who became advocates. One beta customer refused to return the device: "over my dead body." And on the third-party rating scale used across hearing aids, where 4.5 out of 5 was the highest score ever recorded, Fortel scored a 6 — breaking the scale. Ayogu says there is currently a waitlist, with a New York office for fittings and rapid expansion planned.

Cadence shows remote monitoring can cut hospitalizations 25% and save Medicare $2.5M weekly

Ayogu's framing for Cadence is arithmetic: if you see your doctor four times a year for an hour each, that is four hours out of 8,760 — the other 8,000-plus hours are unmonitored, and that is where care actually changes. Cadence mails hardware (blood pressure cuff, scale) to patients' homes, runs AI monitoring on the continuous signal, and staffs humans the AI can escalate to. It partners with health systems rather than building outside them, operating across roughly 7,000 sites of care and monitoring over 100,000 patients, integrated into the EHR so physicians see the same data.

The published outcomes he cites: 25% reduction in hospitalizations, over 80% improvement in blood pressure control (which prevents strokes), and roughly a full point reduction in A1C. On cost, he says Cadence saves Medicare $2.5 million every week. The mechanism he emphasizes is timing — calling a patient an hour after a high reading often fails because they have moved on to bingo, Trader Joe's, or grandchildren, so the agent intervenes at the moment of measurement, with chart context to distinguish a chronically high reading from an acute one. On AI-clinician integration, he says Cadence's agents show 98% concordance with what the prescriber would have done, and that AI has driven a 70% increase in reaching patients.

OpenEvidence hit two-thirds physician adoption in three years, and the reason is product quality

Ayogu's thesis on clinical software is that the industry misdiagnosed physician behavior. The standard line is that doctors dislike technology; his counter is that doctors all own iPhones and Macs and simply hate the bad enterprise tools they are forced to use. OpenEvidence was the first clinical tool he heard physicians praise unprompted — "with joy in their eyes." He tried it, found it differentiated, then discovered founder Dan had gone to college with him, emailed, and drove to his house to meet.

The adoption comparison is the section's headline number: Epic took 40 years to reach its scale and has under 50% of physicians using its product, while OpenEvidence reached two-thirds of American physicians in three years. Ayogu calls it an N-of-1 product — a clinician sees why it is different immediately.

OpenAI is now one of the largest healthcare providers in America

Ayogu's claim here is deliberately provocative: with 300 million people asking ChatGPT healthcare questions every week, OpenAI is functionally one of the largest healthcare providers in the world. He argues healthcare, not coding, is the hero use case for most Americans — "find me someone who does not have a story" of putting a complex medical journey into ChatGPT. He cites a partnership with Boston Children's and Harvard hospitals in which roughly 300 unsolved pediatric cases yielded new diagnoses for about 20.

His account of bringing OpenAI into healthcare is a story about changing minds. About two and a half years before the episode, he organized a series of meetings between OpenAI leadership and top biotech, payer, and health system CEOs. Reasoning models were not yet public, but he had access and began feeding the first halves of New England Journal of Medicine case studies — the House MD-level difficult cases — into the model weekly, cutting them off before the diagnosis. He says the model got them right every week, where he and colleagues might get two in ten. When he showed health system leaders, he was treated "like a heretic" until they saw it live. One biotech CEO told him the tools might be good for customer service but would never be useful in drug discovery; two years later, Ayogu says, that same CEO has announced billions in spend in the space.

The downstream argument is about system capacity. If patients can handle more of their own care, the system can redirect resources — he notes a three-to-four-month wait for an orthopedic surgeon in LA, and argues that if surgeons only saw cases requiring surgery, waits could compress to a month, patients would be happier, and surgeons would earn more by operating.

"As a nation, I think we need a 10x nurse a lot more than we need a 10x engineer."

$40 trillion over a decade bought zero life expectancy gains

This is the episode's most quotable indictment. Ayogu lays out the macro numbers: $5 trillion in annual US healthcare spend, projected to reach $9 trillion by the end of the decade; healthcare is the number one employer at 20% of all jobs, and projected to account for 30% of new jobs; it is the largest line item in the federal budget at six times defense. He says he would happily accept higher spend if outcomes improved — "if someone told me, hey, we're gonna spend a trillion more on healthcare this year, but everyone gets to live to 100, I'll take that trade any day."

The historical arc he draws: life expectancy went from 50 to 70 between 1900 and 1950, driven by sanitation, water, reduced infant and child mortality, antibiotics, and vaccines. From roughly 1990 to 2010, it went from 70 to 80, half from reduced smoking and half from statins and blood pressure control reducing strokes. Then, over the last decade, roughly $40 trillion spent with life expectancy flat.

"The last 10 years, we've spent about $40 trillion on healthcare. We have not moved our life expectancy in age. And if that is not an indictment of the failure of the healthcare system, I don't know what is. $40 trillion for no change."

He notes the distributional story: Hawaii at 80 versus West Virginia at 72; a roughly 15-year gap between the richest and poorest 1% of men. Cancer mortality fell substantially over the same period — breast cancer was once near-fatal — but was offset by rising overdoses. His prediction is that the next decade should be the one where the US goes from 80 to 100, if the system does its job.

N-of-one medicine works, and the price tag is the next bottleneck

The canonical example Ayogu gives is baby KJ, an infant with a rare genetic disease that prevented ammonia processing, fatal in more than half of cases, with liver transplant as the only conventional option. At Penn, clinicians designed a custom gene-editing treatment delivered when the baby was seven months old. It worked. The child is now two, walking and talking. The treatment cost roughly $1.5 million to develop and deliver for one patient.

That figure is the crux of his argument about healthcare's hardest question: what is the price of a life, asked as a system on a regular basis, because every cancer treatment, transplant, and gene therapy is life or death. He says he would pay it — raise taxes, cut elsewhere — but acknowledges the aggregate affordability problem and argues the answer is innovating across the supply chain so the cost is not $1.5 million per patient. The frontier science is arriving; the manufacturing, payment, and infrastructure to make it repeatable and cheaper is not yet built.

Lilly won GLP-1s by innovating across the value chain, not just the molecule

Ayogu's explanation for Lilly's dominance is a template for the trillion-dollar healthcare company. Novo was first to market, and Lilly's drug had better weight loss — but he argues the decisive factors were manufacturing and commercialization. When GLP-1s were so popular that shortages made them unobtainable, Lilly had leaned into manufacturing capability and got supply on board roughly a year before Novo unblocked theirs. On the commercial side, Lilly Direct let the company partner and lean into care delivery in ways a traditional pharma company would not. His broader point: the same way AI made it easier to build software companies with more surface area out of the gate, biotech companies are now doing science plus manufacturing, or science plus care delivery — which makes them more defensible and gives them a much bigger TAM.

He applies the same lens to Keytruda: Merck's single drug generated 2025 revenue larger than all the major labs combined. And to the Merck-Moderna cancer vaccine, where a patient's tumor is removed, sequenced, and used to build a personalized mRNA vaccine given alongside Keytruda as an immune boost. In phase three, he says, patients who received it were 50% less likely to have a recurrence within five years — and because the immune system has memory, later recurrences get caught while small.

Oscar Health's 95% drawdown and recovery, and where incumbents go next

Oscar was Thrive's first-ever investment, co-founded by Josh Kushner when he co-founded the firm. Ayogu admits he told Kushner and Mario Schlosser at the time that it was a bad idea — the last new health insurer was UnitedHealthcare in 1977, pre-internet, and building a new insurer takes too much scale and capital. He says he was glad to be proven wrong. Oscar fell 95% from its peak as of December 2022 and is now back above its IPO price, up roughly 100% year over year, during a period when health insurers broadly had a hard 18 months. His explanation is the tech stack Oscar built a decade ago — its own claims processing infrastructure — which let it move fast when AI arrived.

On the exit environment, Ayogu expects both outcomes: standalone companies with enough surface area to IPO, and incumbents that move rather than stand flat-footed. He cites Lilly launching Tune Labs with NVIDIA, Oracle moving fast, and McKesson — a distributor — buying large oncology care delivery assets, which he calls a new kind of lane-crossing. The mechanism is that when it is easier to build, companies move faster into adjacent spaces.

What to watch

The unresolved tension in the episode is between Ayogu's optimism about the science and his own account of the system's failures. He predicts the next decade takes US life expectancy from 80 to 100, but the same conversation documents $40 trillion spent for zero movement, a 15-year gap between rich and poor men, and a $1.5 million price tag for a single cure. His answer is that the bottleneck keeps shifting — discovery, then trials, then manufacturing, then payment — and that companies which build across the whole chain will be the ones that reach trillion-dollar scale. The specific markers to track: whether Isomorphic's 88–90% success-rate ambition produces measurable clinical-trial compression; whether Fortel's waitlist clears and the company scales beyond New York; whether Cadence's published outcomes hold as it grows past 100,000 monitored patients; whether the biotech CEO who dismissed AI in drug discovery follows through on the billions in announced spend; and whether OpenAI's 300 million weekly healthcare queries translate into durable system-level capacity rather than just consumer self-service.