Key figureAlex Colville (co-founder, age1)Four movesNvidia × Lilly · Isomorphic · Retro · AnthropicLabelsfact, primary quote, secondhand account and unverified are marked separately throughout
Why does it feel as if another AI company announces a move into biotech almost every week? Nvidia and Eli Lilly are building a joint lab in the San Francisco Bay Area; Isomorphic Labs, spun out of DeepMind, is developing drugs; Sam Altman has personally invested in the longevity company Retro; and Anthropic keeps expanding into the life sciences. This piece follows one question: why they are coming, what they can genuinely help with, and where they will get stuck. We start with one of the people closest to the question, Alex Colville, co-founder of the longevity venture fund age1, and then set out this desk's framework for reading the news. Research date: October 26, 2026; status reflects information searchable as of late October 2026.
Ideas will keep getting cheaper; proof will not.
Nvidia × LillyJoint lab, up to $1B over five years0 outputs disclosed
Isomorphic LabsAbout $2.7B raised in total0 clinical candidates
RetroOnly one in human Phase I0 efficacy data
AnthropicWorkbench and named customers liveOwn programs preclinical
To judge any headline of the form "Company X moves into biotech," ask one question: is it solving an idea problem or a proof problem? The first is getting cheaper fast because of AI. The second, meaning patients, time, regulators and Phase III trials costing hundreds of millions of dollars, does not get cheaper just because models get better.
Colville's core view compresses into three sentences. AI companies are coming because this is where the world's largest per-product revenues and deepest moats are. AI has genuinely changed "thinking up molecules" and everyday efficiency. But validation is "the last frontier," and that is where AI has the least leverage. In his own words on a podcast, AI for biology is "short term super hyped, long term super under hyped."
One boundary first: on the specific question of why AI companies are moving into biotech, Colville's most systematic public account comes mainly from a relayed article published by PANews on October 26, 2026 (author: "Shensi Quan"). It has no transcript and no direct quotes. What he said himself about AI and longevity on the Free Radicals podcast (November 25, 2025, with a full transcript) is what can be quoted directly. We keep the two apart: relayed claims are marked as relayed, and only transcript passages appear in quotation marks.
Summary
- First, the arithmetic works. The global pharmaceutical market was about $1.74 trillion in 2025; two GLP-1 molecules alone, tirzepatide and semaglutide, sold more than $70 billion combined in 2025. Patents plus heavy regulation form a moat that tech products lack: prices are not quickly competed down to marginal cost. For AI labs whose valuations need astronomical revenue to justify them, this is one of the few new markets big enough.
- What AI genuinely helps with sits in the first half of the chain. Protein structure prediction (the AlphaFold database is used by more than 3 million researchers), molecule generation and screening (a drug whose target and molecule were both AI-led met its primary safety endpoint in Phase IIa and showed an efficacy signal), literature and data integration, regulatory documents and clinical operations: this layer has hard evidence and named customers.
- What it has not helped with is the most expensive stretch. Proving a molecule safe and effective in humans takes most of the cost and time of drug development. An August 2026 review in Nature Reviews Drug Discovery concluded that evidence of AI's clinically relevant impact is "disappointingly limited," and as of 2026 no drug widely recognized as AI-discovered has won FDA approval.
- The four signature moves differ in substance. Nvidia × Lilly is a joint commitment of "up to $1 billion over five years," with the split undisclosed and no product readout. Isomorphic has raised about $2.7 billion and is reportedly in talks at a valuation of at least $40 billion, yet as of end-September 2026 no candidate had entered the clinic. Retro is the only one of the four with a drug in human Phase I, and Phase I answers safety questions only. Anthropic's products and customers are real, but key moves such as a roughly $400 million acquisition and its own wet lab rest mostly on media reports, without a full official announcement.
1 · What he says: Alex Colville's view
Source quality first, then the views: every claim in this section carries a source grade (see the boundary note above).
Who he is
age1 launched on September 19, 2023, with a $35 million first close. It was co-founded by Laura Deming and Alex Colville, who serves as a general partner. Deming had earlier founded The Longevity Fund, a longevity-focused venture fund, in 2011. age1's mission is to shift from "sickcare" to "healthcare," investing at the earliest stages (angel, pre-seed, seed and Series A).
A popular claim that needs correcting: "Stanford PhD in genetics" is not supported by primary sources. The consistent primary account is that he did his PhD in aging biology in Thomas Rando's lab at Stanford, with research focused on genetic screening. The accurate phrasing is "Stanford PhD in aging biology, with research involving genetic screens."
- Firm
- age1 (co-founded with Laura Deming, launched 2023)
- Prior capital
- Previously managed more than $100 million in investment and philanthropic capital at organizations including the Amaranth Foundation (per the official announcement)
- Check size
- $500,000 to $4 million (age1 website)
- Source grades
- Primary transcript Free Radicals podcast, 2025-11-25; Secondhand, relayed PANews article, 2026-10-26, no transcript; Figure unverified figures in the relayed account with no independent source
Why they come, reason one: revenue scale
ViewSecondhand, relayed
According to the PANews account, Colville opens with a number: semaglutide and tirzepatide together earn more than OpenAI and Anthropic combined, and are still growing fast. The valuations of OpenAI, Anthropic and Google are now so high that they need equally staggering revenue to support them; look around, and the business with the highest revenue per product in the world is drugs. It would be strange not to look.
Fact check: the comparison holds at 2025 scale. Tirzepatide sold $36.5 billion in 2025 and semaglutide about $34.5 billion to $36.1 billion (tallies differ by about $1.6 billion), more than $70 billion combined, far above the two AI companies' actual 2025 revenue. On 2026 annualized revenue the conclusion could change, so the point is time-sensitive.
Why they come, reason two: the moat
ViewSecondhand, relayed
As relayed, he argues that drugs are patent-protected and the industry heavily regulated, so they do not drift toward commoditization with prices driven ever lower like most tech products. He cites an observation Marc Andreessen often makes: prices rise steadily in regulated industries and are pushed down relentlessly in unregulated ones. So in his view biotech is "highly resistant to disruption by AI": no model company will wipe it out within a year, because the hard problems are many and they take as long as they take, however much intelligence is thrown at them.
Reading the signal: Anthropic "it's on"
ViewSecondhand, relayed
As relayed, he says Anthropic published a market map listing life sciences among the areas of greatest impact, and that its team has said publicly this could be the company's second-largest, or at least a top-five, market; Anthropic recently released a new model for the life sciences and mentioned preclinical programs in aging and rare disease. His reaction, as relayed, was a single line: when Anthropic says publicly that this is a priority, "it's on."
Check: Anthropic does officially describe accelerating scientific progress as core to its public-benefit mission (see Part 2), but the specific claims about a "market map" and "second-largest market" could not be verified through official Anthropic channels.
What AI can do now, one: generate molecules
ViewSecondhand, relayed
As relayed, he holds that for some drug classes designing molecules is already far faster and cheaper than before, with no need for people and labs to go back and forth at every step, though "no people, no experiments" is still far away. He cites Nabla Bio as able to generate antibodies and small molecules zero-shot; for now, most work still targets known targets to make new versions of existing drugs, only very quickly, and the next step will extend to entirely new targets.
Note: Nabla Bio's capabilities were not independently checked here; this is Colville's example as relayed, not a verified fact.
What AI can do now, two: everyday work 30% to 95% faster
ViewSecondhand, relayedFigure unverified
As relayed, he says everyday tasks such as writing regulatory documents, researching competitors, finding partners and getting up to speed on a new target are now 30% to 95% faster; eighteen months to two years ago, founding a drug company was a heavy lift, and now it has been almost "democratized."
Note: the 30% to 95% range is extremely wide and has no published method; it is a personal estimate. But "much faster document work" has independent customer corroboration: NovoScribe, built by Novo Nordisk with Anthropic, cut clinical study report generation from months to minutes (partner's figure; see Part 4).
Cold water on the spot: questioning Retro × OpenAI's "50x reprogramming"
ViewPrimary transcript
This is the primary material that best shows how he judges claims. In the Free Radicals transcript (around 00:44 to 01:00) he questions, point by point, the announcement that Retro and OpenAI used GPT-4b micro to design Yamanaka factor variants with roughly 50x better reprogramming efficiency. First, what was shown were markers, not actual efficiency. Second, engineered OSKM variants with roughly 50x improvements have existed in the literature since 2020: "it makes me question, did the AI just read that paper and make a fancy variant that does the same thing, or is it truly de novo?" Third, even if reprogramming efficiency in a dish improves, delivery remains a huge bottleneck, and this has not cured aging. He also said he is "eating popcorn" and genuinely hopes they deliver the data.
The bottleneck comes after the molecule: proof, not ideas
ViewSecondhand, relayed
As relayed, he thinks the vision of "millions of zero-employee biotech companies" is far from reality: turning an optimized molecule into an approved drug requires proving it safe and effective, and that stretch takes the lion's share of drug development cost and time; molecule design, hard and costly as it is, is small by comparison. This has become "the last frontier": pick what is most likely to work, then run the whole process as cheaply as possible. Even if one person designs a molecule with AI, every company still needs millions of dollars afterward, and that is exactly where AI has the least leverage today.
The biggest bottleneck: talent and basic research
ViewPrimary transcript
At 01:19:10 in the transcript, asked whether young people should go into academia or industry, he answers: "Both. Talent is the biggest bottleneck in this field, and basic research is still one of the huge bottlenecks."
Criticism of big pharma: not yet serious enough
ViewPrimary transcript
His age1 "2025 Pharma Aging Report Card" grades 11 global drugmakers. From 01:12:43 in the transcript he says big pharma's buy-in on aging over the past three years has been unprecedented (for example, AbbVie's multiyear, multibillion-dollar partnership with Calico, and Novartis's unit for diseases of aging and regenerative medicine), because after their success with GLP-1 and PD-1 they are looking for "the next market bigger than obesity or cancer"; yet "none scored above a B," because none has shown the all-in commitment that will decide the next century's winners.
The field's balance sheet: small, early, no products
ViewSecondhand, relayedFigure unverified
As relayed, he estimates that all companies targeting aging are worth only about $15 billion combined, with company count and funding about one-fiftieth of oncology's; the field is still in "the earliest innings," so every dollar carries great leverage. No anti-aging commercial product has yet been proven effective in humans.
Note: the $15 billion and one-fiftieth figures have no independent statistical source to check against; cite them as Colville's personal estimates.
Timeline: first proof point within a decade, the "penicillin moment" 10 to 15 years out
ViewSecondhand, relayed
As relayed, he expects the first mass-market products within the next ten years, aimed at delaying or preventing some aspects of aging; he calls this longevity's "ChatGPT moment," though feedback will not be as instant as a first conversation with a chatbot. The "penicillin moment" he is waiting for is perhaps 10 to 15 years away (his grandfather was among the first American soldiers to receive penicillin and lived into his nineties after pneumonia); what he wants is "a statin for systems beyond the cardiovascular."
The nearest stand-in proof point: a drug for dogs
ViewSecondhand, relayed
As relayed, he says Loyal, an age1 portfolio company, has cleared efficacy and safety review for its canine longevity drug with only manufacturing left, expected around the end of next year, while a human version is years away.
Check: broadly consistent with official disclosures, but the wording needs tightening. On February 11, 2026, Loyal announced a $100 million Series C led by age1; its LOY-002 has completed two of the three major sections of FDA expanded conditional approval (XCA), reasonable expectation of effectiveness (RXE) and target animal safety (TAS), with the third (manufacturing) in progress. The official wording is "two of three sections complete"; do not inflate it to "approved except for manufacturing."
Investment framework: bet first on what can be proven
ViewPrimary transcript
In the transcript he says the most convincing near-term step is delay and prevention, including proving in dogs that the field is not snake oil; the long-term unlocks are two parallel tracks, restore (reprogramming) and replace. He personally prefers restoring over repairing item by item, because there are thousands of kinds of damage and fixing them one at a time does not scale, while reprogramming might address many at once. He also says "replacing most cells in the body is a 40-year goal."
Note: the betting order summarized by the PANews author (if you can back only one category, back delay and prevention, which can be proven first; replacement and reprogramming come five years later) is the author's summary of the conversation. It matches the primary framework in direction, but "five years later" has no primary source. Among the second-generation details he mentions, "NewLimit has more than $600 million in cash" could not be checked against any official or reliable financial source. He also stresses that the extracellular matrix accumulates damage too, so turning back the clock inside cells alone may not be enough; several approaches must proceed in parallel.
The "halo" argument: mission and shield are the same thing
ViewSecondhand, relayed
This is the most unusual passage in the section. As relayed, Colville says he is glad that nearly every frontier lab founder genuinely cares about biosecurity, pandemics and bioterror threats; a typical big-company CEO in that seat "might not care as much." He does not think the labs always get it right or do enough, but "the caring is real." He is also candid about the reputational layer: the labs know that once they are seen as able to cure cancer, they become very hard to restrict or shut down. He admits that "hiding behind that halo is a little unsettling," but thinks their genuine care is a huge benefit for humanity.
"Definitely both. Short term super hyped. Long term super under hyped."
Colville in the Free Radicals rapid-fire round (transcript 01:19:41), answering "Is AI applied to biology overhyped or underhyped?" In the same passage he says he may live to 150 and gives a 50% chance of reaching longevity escape velocity in his lifetime: these are rapid-fire personal beliefs, not model outputs, and should be quoted with that context.
Limits of this section: we found no long-form piece by Colville systematically arguing about AI and biotech, and no X thread of his devoted to AI (his findable posts mainly promote age1's reports, for example on July 29, 2025, when the Pharma Aging Report Card came out and he said the race would "unlock a market bigger than obesity"). On episode 85 of the Core Memory podcast (August 2026), host Ashlee Vance called him his "reality check on longevity hype," but we found only the episode summary, not a transcript, so we quote no specific views from it.
2 · The four moves, on the record
The four moves are often written up as one story. They are really four different ways in: a supplier-style joint lab, a spun-out drug company, a founder's personal check, and a product-line expansion. Comparison table first, then the record for each.
| Move | Announced / start | Money | Form | Hard progress as of October 2026 |
|---|
| Nvidia × Lilly | 2026-01-12 | Up to $1B jointly over five years; split undisclosed | Co-located lab plus compute stack | No drug or clinical output disclosed |
| Isomorphic Labs | Spun out of DeepMind in 2021 | About $2.7B in external funding; pharma upfronts totaling $82.5M, milestones up to about $2.9B | Own pipeline plus pharma partnerships | 0 candidates in the clinic; reportedly in talks at ≥$40B valuation (not closed) |
| Altman × Retro | Seed funding around 2022, disclosed 2023 | Altman personally about $180M; 2026 round at $1.8B pre-money, amount undisclosed | Independent company, personal investment | RTR242 in Phase I (safety only), no efficacy data |
| Anthropic life sciences | Series of moves from 2025-10 | Coefficient Bio acquisition about $400M (reported) | Sells the workbench plus own preclinical work | Products and named customers live; own drug programs preclinical |
Posture · supplier joint lab
Nvidia × Lilly: an AI co-innovation lab in the Bay Area
- Announced
- January 12, 2026, during the J.P. Morgan Healthcare Conference
- Amount
- Up to $1 billion jointly over five years
- Location
- San Francisco Bay Area (starting in South San Francisco)
- Latest
- No new drugs, clinical results or commercial products found
"Up to" is a ceiling, not money spent, and the split between the two companies is undisclosed. The form is co-location: Lilly's biology and medical experts work alongside Nvidia's model builders and engineers to generate data at scale and build foundation models for biology and chemistry. The first focus is a "continuous learning system," in which an agentic wet lab and a computational dry lab run in a closed loop around the clock with scientists in the loop; later work explores robotics, physical AI and digital twins in clinical development, manufacturing and commercial operations. Lilly brings drug discovery, clinical and manufacturing expertise, lab infrastructure and data; Nvidia brings the BioNeMo platform, the next-generation Vera Rubin architecture, open biology models and DGX Cloud compute. Precursor: in October 2025, Lilly had already announced its own AI supercomputer built on more than 1,000 Nvidia GPUs.
An acerbic but structurally sound reading (financial commentary site AInvest): $1 billion over five years is $200 million a year, a mere "option premium" for companies worth about $5.5 trillion (Nvidia) and $1.1 trillion (Lilly). Nvidia's certain gain is installing its whole BioNeMo and Clara stack inside the world's most valuable drugmaker and collecting compute revenue now, whether or not a drug ever emerges. That is commentary, not a finding of fact, but it points to a checkable structure: the shovel seller gets paid first.
Posture · spinout, own pipeline
Isomorphic Labs: DeepMind's drug-company spinout
- Origin
- Spun out of Google DeepMind in November 2021, part of Alphabet; founder and CEO Demis Hassabis
- Funding
- First external round $600M (2025-03); Series B $2.1B (2026-05); about $2.7B in total
- Clinical status
- As of 2026-09-30: no candidate in human trials, no IND announcement
- Valuation rumor
- Reportedly in talks at ≥$40B valuation (2026-10, early stage, not closed)
Pharma deals: in January 2024 it signed with both Lilly and Novartis. Lilly paid $45 million upfront with up to $1.7 billion in milestones (excluding royalties) across several small-molecule targets; Novartis paid $37.5 million upfront with up to $1.2 billion across 3 targets; together, nearly $3 billion. In February 2025 Novartis expanded the deal by up to 3 more programs; on January 20, 2026, a multi-target partnership with Johnson & Johnson was added, terms undisclosed. Reminder: milestones are paid only as candidates clear each stage, and most are never paid; nearly $3 billion is "everything hits" ceiling arithmetic, not revenue.
Pipeline: the internal pipeline (oncology, immunology) has not even named a candidate; Hassabis's earlier goal of "clinic by end-2025" has slipped to "end-2026." Its drug design engine IsoDDE published a technical report in February 2026 claiming more than twice AlphaFold 3's accuracy on hard protein-ligand cases: a self-reported benchmark, not a drug, more checkable than a press release but still not peer-reviewed clinical evidence. Seeking a $40 billion valuation with zero clinical data is the cell most worth watching here.
Posture · personal check
Sam Altman's personal investment in Retro Biosciences
- Amount
- About $180M of initial funding from Altman personally, not OpenAI money
- Goal
- Stated goal: add ten years to healthy lifespan
- Latest
- Lead drug RTR242 in human Phase I (the only one of the four)
- Valuation
- First close of a new round in 2026-05 at $1.8B pre-money, amount undisclosed
Three mechanisms in parallel: boosting autophagy (restoring lysosomal acidity and restarting the cell's recycling system), cellular reprogramming (Yamanaka factors) and plasma-inspired therapies. Lead drug RTR242 is an oral small molecule that restores lysosomal function and autophagy, aimed first at Alzheimer's disease. In December 2025 it dosed its first healthy volunteer in Australia, entering Phase I; the company says it took 15 months from choosing the indication to first human dose. August 2026 reports: Phase I added two higher-dose cohorts and grew from 76 to 108 participants, with no major safety signal so far according to the company; there is no Phase II timeline and no efficacy data. Phase I in healthy people can answer only safety, tolerability and dosing questions.
There is another layer with OpenAI: a custom OpenAI model, GPT-4b micro, designed SOX2 and KLF4 variants, and in early 2025 the company claimed more than 50x higher expression of stem-cell reprogramming markers. Outside scientists could not verify it at the time (unpublished, model not released); later reports said the results were reproduced across donors, cell types and delivery methods. Colville's objections are in Part 1: the same event, read very differently by the press release and by an expert in the field. Separately, the Financial Times reported in January 2025 that Retro was raising a $1 billion Series A with Altman participating; whether that round closed at that size has not been confirmed by the company (unverified).
Posture · product-line expansion
Anthropic expands into the life sciences
- Start (official)
- Claude for Life Sciences launched 2025-10-20
- 2026-01 expansion (official)
- Added connectors including Medidata, ClinicalTrials.gov, ChEMBL and Owkin, plus clinical operations features
- 2026-06-30
- Claude Science (beta): a workbench connecting 60+ scientific databases and tools
- Acquisition (reported)
- Coefficient Bio acquired 2026-04, about $400M in stock; no official announcement
The official text says "accelerating the pace of scientific progress is core to Anthropic's public benefit mission," with the aim of supporting work from early discovery through translation and commercialization. Claude for Life Sciences runs on Claude Sonnet 4.5 (Protocol QA benchmark 0.83 versus a human baseline of 0.79, both company figures), with connectors to Benchling, BioRender, PubMed, Wiley Scholar Gateway, Synapse.org and 10x Genomics, and free API credits for researchers through its AI for Science program; named customers on the page include Sanofi, Novo Nordisk, the Broad Institute, Genmab and Schrödinger. Its head of life sciences, Eric Kauderer-Abrams, called it a "threshold moment."
The reported side: Coefficient Bio had fewer than 10 people, mostly computational biologists from Genentech's Prescient Design; Anthropic issued no announcement, so the price is marked "unverified, reported." In September 2026 Kauderer-Abrams confirmed to Reuters that the company runs a robotic wet lab in the Bay Area and said it does preclinical work only, runs no human trials and does not sell finished drugs directly (as relayed). Protein design results disclosed at the same time: 354 of 1,320 AI designs bound their targets in the lab (26.8%), and 14 of 15 targets got at least one successful design (company figures, via media reports). In the Novo Nordisk partnership, NovoScribe cut clinical study report generation from months to minutes (partner's figure).
For comparison (showing that "almost one a week" is not an illusion): on April 16, 2026, OpenAI released GPT-Rosalind, a model for the life sciences (trusted access; early partners include Amgen, Moderna, Novo Nordisk, Thermo Fisher and the Allen Institute), and on October 20, 2026, it acquired Orchestra Bio, a platform for coordinating drug development programs (terms undisclosed), which it plans to fold into its Rosalind life sciences tools.
3 · Why they are coming: six drivers
Put the four moves side by side and the driver is not a single "bullish on biotech" view but six forces stacked together, each with concrete examples.
i · Compute and model capability spilling over
In the 2020s biology became a computing problem: AlphaFold showed that sequence-to-structure can be learned, and Nvidia's BioNeMo, Vera Rubin and DGX supercomputers are built for exactly this kind of workload. Examples: Lilly first built its own AI supercomputer with more than 1,000 GPUs (2025-10), then a joint lab with Nvidia (2026-01); Isomorphic's IsoDDE is built directly on AlphaFold 3. For Nvidia, this is also the most natural path from general capability to vertical-industry orders: sell the compute first, let the drugs take their time.
ii · Proprietary data as a moat
Models can be copied; failure data cannot. Public data covers only a few thousand approved drugs, while big drugmakers hold vast unpublished failed compounds and negative results, exactly the data missing for learning "what doesn't work." One official purpose of the Nvidia × Lilly lab is to generate new data at scale; Lilly's CEO spoke of combining "our volume of data and scientific knowledge" with Nvidia's compute and modeling. The pitch of Anthropic's Benchling connector is access to private lab records within permissions and audit logs. This explains why the model is "AI company plus incumbent drugmaker" rather than AI companies alone: half the moat (data, wet labs, clinical networks) sits with drugmakers and the other half (compute, modeling talent) with AI companies, and neither can assemble it alone.
iii · Market size and the shape of returns
The global pharmaceutical market was about $1.74 trillion in 2025; a single blockbuster can sell more than $30 billion a year (tirzepatide, $36.5 billion in 2025). Patent-period exclusivity plus insurance reimbursement make the returns look more like "regulated monopoly rent" than the rapid commoditization common in tech. The flip side matters just as much: a new drug takes 10 to 15 years on average at a capitalized cost of about $2.6 billion (the classic 2016 Tufts estimate by DiMasi, disputed; many think it overstated), and about 90% of candidates entering the clinic fail. High returns and high attrition are two sides of one coin. The AI companies' bet: improving success rates or speed by even a few percentage points changes expected value by tens of billions of dollars.
iv · Talent flow: people who understand both computing and biology are moving this way
Isomorphic's core team came straight from DeepMind (Hassabis himself and researchers from AlphaFold and AlphaStar), and its pitch is a cross-disciplinary computing-plus-biology team. Anthropic reportedly paid about $400 million for Coefficient Bio, fewer than 10 people, essentially a former Genentech computational biology team: about $40 million per head, paying for talent and industry credibility, not product revenue. OpenAI's purchase of Orchestra Bio (terms undisclosed) may reflect similar thinking. Colville gives the supply-side view of the same point: longevity's "biggest bottleneck is talent," and soaring demand for scarce people who understand both sides is one reason acquisition prices look outlandish.
v · Founders' personal interest and mission narratives
Hassabis has said many times that AlphaFold and curing disease are "why I started DeepMind"; Altman not only invested in Retro but has long talked publicly about a longevity lifestyle, and his $180 million was a personal check, not a corporate budget; Anthropic wrote scientific progress into its public-benefit mission. Colville's distinctive contribution is naming both sides at once (see the "halo" argument in Part 1): founders' concern about biosecurity is real, but the labs also know that being seen as able to cure cancer makes them hard to restrict or shut down. The safest reading is not to choose between "real mission" and "empty narrative": the evidence supports both at once, and the mission narrative itself has quantifiable business value (hiring, regulatory goodwill, valuation story).
vi · Capital and the valuation story
Isomorphic raised about $2.7 billion with zero clinical data and is reportedly seeking a $40 billion valuation; Retro is valued at $1.8 billion pre-money with only a Phase I safety trial; Thrive Capital led both of Isomorphic's rounds. This pricing borrows the grammar of the AI industry (platforms, engines, economies of scale) rather than biotech's traditional grammar (pricing risk down stage by clinical stage). It holds only if capital believes faster discovery will eventually translate into higher clinical success rates, and the evidence in Part 5 shows that chain has not been proven. The valuation story then becomes a driver itself: the more leading AI companies step in, the more latecomers must follow so as not to be absent from the "next big market" narrative. That is one direct cause of the "almost one a week" rhythm.
4 · What AI genuinely helps with: hard evidence
Set the marketing aside and look only at what has third-party adoption, peer review or named customers behind it. The answer is layered: the earlier in the chain, the harder the evidence.
1 · Protein structure prediction: the only breakthrough the whole scientific community has ratified
200M+Predicted structures in the AlphaFold database (240 million per February 2026 reports, growing with each release)
3M+Researchers using the database, in 190+ countries, more than 1 million of them in low- and middle-income countries
35K+Papers citing the work; Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry
Limit: it predicts static structure, which is not the same as knowing a protein's function, dynamics or druggability. Structure is where drug design starts, not where it ends.
2 · Molecule generation and screening: the first positive Phase IIa signal
Case: Insilico Medicine's rentosertib. Its AI platform PandaOmics found a new target, TNIK, and its generative chemistry system Chemistry42 designed the molecule, making it the first candidate with both target and molecule AI-led. The Phase IIa GENESIS-IPF trial enrolled 71 patients with idiopathic pulmonary fibrosis and was published in Nature Medicine on June 3, 2025: it met its primary safety endpoint; at 12 weeks the 60 mg once-daily group gained an average of +98.4 mL in FVC versus −20.3 mL on placebo, a dose-dependent efficacy signal. The company says target to preclinical candidate took about 18 months (company figure). In July 2026 the company announced it would advance to Phase III.
A sobering note: this is still a 71-patient, 12-week Phase IIa proof of concept, not an approval; and it comes from an AI-native company, not from the four giants in Part 2. None of them has a clinical readout of that grade yet.
3 · The tally: how many reach the clinic, and stage success rates
Phase I · AI molecules80% to 90%
Phase I · historical40% to 65%
Phase II · AI molecules~40%
Phase II · historical29% to 40%
FDA approved · AI-discovered0
Source: analysis by Jayatunga et al. in Drug Discovery Today (2024), as reported by BioSpace and PitchBook; bar length uses the midpoint of each range. Historical Phase I baselines vary by study.
The sample is tiny: PitchBook stressed that AI companies had completed only about 10 clinical trials in total at the time. How to read it: Phase I mainly tests safety, and AI screening out obviously toxic or poorly behaved molecules early may genuinely raise Phase I pass rates; Phase II starts testing real efficacy, and AI's advantage has not shown up there. That squares with Colville's view that the bottleneck is proof. Watch the definitions too: "AI drugs in the clinic" of 67 (2024, narrow definition, AI-discovered molecules) and 173 to 175+ (2026, broad definition, AI-involved programs) are not comparable and cannot be joined into a trend line.
4 · Trial design, recruitment and operations: real gains, but at the operations layer
Anthropic's official January 2026 update ties the capabilities to concrete workflows: drafting clinical trial protocols (incorporating FDA and NIH requirements and institutional templates), tracking enrollment and site performance with Medidata data to surface schedule risks early, and finding and drafting gaps in regulatory submissions. NovoScribe, built by Novo Nordisk with Anthropic, cut clinical study report generation from "months" to "minutes" (partner's figure); Sanofi, named on Anthropic's official page, says Claude is embedded in its daily work and that it sees "efficiency gains across the value chain."
Limit: these gains compress document and coordination time, not biological time. Patient enrollment, accrual of endpoint events and long-term safety follow-up cannot be shortened by a single day.
5 · Literature and data integration: the least glamorous, the first to pay off
The Claude for Life Sciences connectors (PubMed literature, Benchling lab records, 10x single-cell data, Synapse collaborative data) plus Agent Skills essentially automate the manual shuttling scientists do among dozens of systems; on the official page, James Zou's team at Stanford describes Paper2Agent, built with Claude Code to turn papers into interactive AI agents. This layer's value does not depend on any biological breakthrough, so it is currently the most certain to pay off. It also explains why Anthropic and OpenAI both sell the workbench first and talk about their own drugs later.
5 · Where it gets stuck: bottlenecks and the skeptics' evidence
The skeptics do not deny AI's value at the front end. They point to one thing: the scoreboard is cherry-picked. Phase I pass rates are quoted again and again, while Phase II looks no different from history, and Phase III and approval data are zero.
1 · Wet-lab validation speed: generation is free, validation is still billed by the day
Anthropic's self-reported protein design experiment is the clearest yardstick: models can produce thousands of designs at once, but of the 1,320 tested in the lab only 354 (26.8%) actually bound; about 3 of every 4 computational predictions fell flat in the physical world, and lab capacity (synthesis, expression, assays) grows far more slowly than generation. Colville's point about Retro's "50x" is the same: an efficiency number in a dish is a completely different question from getting into human tissue without causing cancer, and delivery remains unsolved.
2 · Named skeptics
Only two things decide success: picking the right target, and knowing early whether a compound is toxic in humans. That is exactly where machine learning has made the least progress, because the data that can be produced in bulk (binding constants, cell activity, animal models) are all proxies, and none replaces the one measurement that counts: whether it works in people.
Derek Lowe, medicinal chemist and author of Science's "In the Pipeline" column, with more than 35 years in drug discovery. In August 2026 he cited and endorsed the review by Bender and colleagues: evidence of real clinical impact from AI is "very thin" so far. In 2024 he also criticized packaging "AI rediscovering known targets" as discovery: if the target-disease link was already known, the model's "success" systematically overstates what AI can really do.
"Despite the development, application and benchmarking of various AI methods, evidence for their clinically relevant impact has so far been disappointingly limited ... we are still at the stage of absence of evidence rather than evidence of absence."
Andreas Bender, Jack Scannell, David Shaywitz and colleagues, review in Nature Reviews Drug Discovery, August 7, 2026. Their prescription is a new scoreboard: measure not only a model's accuracy on clean datasets but whether it improves real decisions in drug discovery.
3 · Biological data: scarce, noisy, and missing its negative results
Unlike internet text, biological data has four problems: it is small (the compounds humans have seriously synthesized and tested number in the tens of millions, not the internet's trillions of tokens); measurements are noisy and not comparable across labs; failures and negative results are rarely published, so training sets carry publication bias by design; and the most crucial data, human efficacy and toxicity, are the scarcest. That is why big drugmakers' private failure data are a moat (Part 3), and why everyone is building their own wet labs to "make data" (the Nvidia × Lilly continuous-learning loop, Anthropic's robotic wet lab, Retro's own manufacturing facility): their actions already concede that the bottleneck is data and validation, not models.
4 · Regulation and payment: rules still in draft, and the indication does not exist
In January 2025 the FDA issued its first draft guidance on using AI to support regulatory decisions for drugs (drawing on more than 500 submissions with AI components from 2016 to 2023), proposing a risk-based credibility framework by context of use; as of September 2026 it remained a draft. Longevity faces an extra hurdle: "aging" itself is not an FDA-recognized indication, so companies must prove their case indirectly through specific diseases (Alzheimer's, sarcopenia and others) or the veterinary route (Loyal's dog drug). That is why Colville says the first proof point will look like "delaying a specific age-related disease," not a product called a longevity drug. Payment is the same: there is no established framework for who pays, or at what price, to "prevent aging" in healthy people. (Note: we found no formal FDA position on a "longevity indication"; this paragraph's judgment is drawn from common industry practice and the cases above.)
5 · A history of failures: what happened to the last wave of AI drug stars
| Company | What happened |
|---|
| BenevolentAI | Listed via SPAC in 2022. In April and May 2023 its lead drug BEN-2293 (atopic dermatitis) met safety but failed on efficacy in Phase IIa, followed by layoffs of up to 180 people and pipeline cuts; more layoffs in 2024; delisted from Euronext Amsterdam in March 2025. |
| Exscientia | An AI drug-discovery IPO star, acquired by Recursion in November 2024 for about $688 million in stock and delisted; commentators noted this was a fraction of its roughly $2.9 billion market value at IPO (a comparison from commentary sources, disputed); its early showcase programs such as DSP-1181 did not make it to a drug. |
| Recursion | After the Exscientia merger, it shelved 3 programs in May 2025 and cut about 20% of staff (about 160 people) the same year to extend its cash runway; its AI-selected repurposed molecule REC-994 met its Phase II safety endpoint but was ended in May 2025 after long-term extension data showed no sustained improvement. The combined company has no approved drug. |
| Industry | Fortune reported in 2025 that after more than 200 AI drug companies and 75 drugs and vaccines reached the clinic, investor confidence and funding fell sharply, with PitchBook deal counts and amounts well below their 2021 peak (as relayed; use specific figures with caution). |
What this history means for today's story: the last wave almost never died because "the model wasn't smart enough," but because of wrong targets, efficacy that did not translate to people, and cash running out first, all problems at the proof stage that Colville and Lowe point to. The giants entering now change the depth of funding and the data conditions, but not the biological rules of clinical trials.
6 · Other structural bottlenecks
- Manufacturing and cost: in the primary transcript Colville says replacement therapies (cells, tissues) have been clearly de-risked on safety and efficacy over the past year and a half, but production costs are extremely high; Retro and Anthropic have both built their own manufacturing or wet-lab facilities, indirect evidence of how expensive outsourced capacity and closed data loops are.
- Safety and dual use: the stronger the capability, the greater the dual-use risk. Anthropic reportedly launched a "life sciences verification program" in September 2026 that relaxes biology-related safeguards only for vetted institutions, showing that the cost of safeguards is now part of product architecture, a real-world footnote to the "caring about biosecurity" in Colville's halo argument.
Our reading: how to read and use this piece
That is what he says. What follows is this desk's (Desk 3, Medical AI) framework for reading it, and the long-term use of this file: not a one-off roundup of hot news, but each move pinned to a cell of evidence and checked cell by cell later. Desk 3 has one spine: an announcement is not a launch.
The yardstick: idea problem or proof problem
Idea problems: getting cheaper with AI
Thinking up molecules, predicting structures, screening out obvious losers, writing regulatory documents, integrating literature and lab data. Each "announcement" in this cell is probably real: the evidence has third-party adoption (AlphaFold), peer-reviewed readouts (rentosertib Phase IIa) and named customers (Sanofi, Novo Nordisk).
Proof problems: not cheaper because models improve
Proving safety and efficacy in humans: patients, time, regulators, Phase III trials costing hundreds of millions. Every "announcement" in this cell is still a promise: Isomorphic's roughly $2.7 billion corresponds to 0 clinical candidates, Nvidia × Lilly's $1 billion to 0 disclosed products. For any new headline, first ask which cell it falls in.
The ledger: four moves pinned to their cells
| Move | Posture | Promise at announcement | Status as of 2026-10 | Next proof point (to watch) |
|---|
| Nvidia × Lilly | Supplier joint lab | Up to $1B over five years, closed-loop wet and dry lab | Lab ramping up, no outputs disclosed | First public model or data output, or first jointly found molecule entering the pipeline |
| Isomorphic Labs | Spinout drug company | AI drug design engine plus own pipeline | About $2.7B raised; 0 clinical candidates; target slipped to end-2026 | First candidate enters the clinic (IND or trial start), the first test of the $40B valuation rumor |
| Altman × Retro | Founder's personal investment | Add ten years to healthy lifespan | RTR242 in Phase I (safety only); "50x reprogramming" questioned by experts as markers, not efficiency | Phase I safety readout and a Phase II timeline; verifiable data from the reprogramming track |
| Anthropic | Product-line expansion plus own preclinical work | Have a substantial share of life sciences work run on Claude | Workbench and named customers live; own drug programs preclinical; acquisition and wet lab per media reports | Whether own preclinical programs reach candidate stage; whether the wet-lab binding rate (now 26.8%) improves |
Why keep a record like this: most people who write about AI do not understand the clinical and regulatory stretch, and most who write about pharma do not know where models' limits lie; "where it gets stuck" is exactly the part both sides miss. The value of this piece is not excitement but discipline: the four moves are four different species, usually lumped together. Separate them, check them cell by cell, and in a year or two, when you look back at who delivered and who bluffed, the person with the record and the person with only clippings will write very different things.
A writing rule: when citing Colville, relayed is relayed. Write "according to the author who spoke with Colville," without quotation marks posing as his words; only passages from the Free Radicals transcript get quotation marks.
Unverified and disputed items
Read before citing.
- Colville's "Stanford PhD in genetics": the primary account is a PhD in aging biology (research focused on genetic screening); see Part 1.
- The PANews article is the main single source for Colville's views on this topic, and it is secondhand: "everyday work 30% to 95% faster," "longevity companies worth about $15 billion combined, one-fiftieth of oncology," "NewLimit has more than $600 million in cash," "Anthropic market map, second-largest market" and similar claims have no independent primary source; each is flagged in the text.
- Semaglutide's 2025 sales: $34.5 billion or $36.1 billion depending on the tally; tirzepatide's $36.5 billion is more consistent.
- "AI drugs in the clinic": 67 (2024, narrow definition) and 173 to 175+ (2026, broad definition) cannot be joined.
- Isomorphic's $40 billion valuation: a rumor of talks, not a completed round; its $2.1 billion Series B and roughly $2.7 billion total are corroborated by several outlets and fairly reliable.
- Anthropic's roughly $400 million acquisition of Coefficient Bio, its own wet lab and internal drug programs: all per media reports, with no full official Anthropic announcement (Claude for Life Sciences and the January 2026 expansion do have official announcements).
- Retro's $1 billion Series A (Financial Times, January 2025): no confirmation it closed; what is officially confirmed is the first close of a 2026 round at $1.8 billion pre-money, amount undisclosed.
- Exscientia "$688 million versus $2.9 billion IPO value": the comparison comes from commentary sources; the deal price itself (all-stock merger) is widely reported, but use the market-value comparison with caution.
- The "$2.6 billion" average cost of a new drug: the classic 2016 Tufts estimate, disputed in academia (including how capital and failure costs are counted); cite it with this caveat.
- Anthropic's "life sciences verification program" and the 26.8% binding rate: company figures via media reports, with no full official technical report.
