Executive Summary
A company whose drugs have never been given to a person has taken the largest single investment in the history of biotech. Isomorphic Labs, spun out of Google DeepMind, announced a $5 billion Series C on November 17. The amount is more than double the $2.1 billion it raised six months ago, and Greenoaks and Sequoia led the round.
What the money is riding on shows up where two sets of statements part ways. Founder Demis Hassabis has said that other AI biotechs always ask for more data and that the request sounds to him like a crutch, while Sequoia said it invested after seeing scaling laws, meaning performance that keeps improving as more compute and more data go in. During the same stretch Isomorphic Labs put money, alongside two other companies, into a venture that manufactures new biological data.
Sections 1 through 5 report what the announcements, the coverage, and the papers say. Section 6 is the reading Pebblous draws from them.
Key Numbers
The first three numbers measure Isomorphic Labs, while the last is a tally for the AI drug discovery sector as a whole. The company's own figures say nothing about how far its scores carry into the clinic.
Source: Isomorphic Labs Series C announcement (2026-11-17) · BioSpace · IsoDDE technical report (2026-02-10) · Jayatunga et al. (2024).
$5 billion
Largest sum a biotech has raised in one round
The previous record was the $3 billion Altos Labs raised in 2022
Zero
Isomorphic-designed drugs given to a person
The target for a first trial has slipped once, from the end of 2025 to the end of 2026
50% vs 23.3%
IsoDDE and AlphaFold 3 on the hardest split
60 cases, with the 95% confidence interval on the gap running from 8.3 to 43.3 points
About 40%
Phase 2 pass rate for AI-discovered drugs
Phase 1 runs 80 to 90%, far above the industry average, while Phase 2 matches it
A Long Investor List, No Molecule Disclosed
The investor list in the announcement runs long. Greenoaks and Sequoia Capital co-led, Andreessen Horowitz came in as a new backer, and Alphabet, GV, CapitalG, Thrive Capital, MGX of Abu Dhabi, Temasek of Singapore, and the UK government's sovereign AI fund all returned. The company disclosed no valuation. Bloomberg reported that the round was discussed at around $50 billion, but that is not a figure the company has confirmed.
The size itself is a record. BioSpace wrote that it is the most a biotech has ever raised in a single round, passing the $3 billion Altos Labs took in 2022. Adding the first outside round of $600 million in March 2025 and the $2.1 billion Series B in May of this year brings the disclosed total to roughly $7.7 billion. That comes five years after DeepMind spun the company out in 2021.
The money is not the only thing that changed. TechCrunch described this round as the first the company has raised since its founder stepped back from running DeepMind. On August 5, 2026, Hassabis handed over the chief executive role at Google DeepMind and took the chairman seat there along with the title of chief scientist at Alphabet, and Koray Kavukcuoglu, until then chief technology officer, took over day-to-day operations as senior vice president. In making the move, Hassabis said he would remain chief executive of Isomorphic Labs and give it more of his time.
The company taking this money, though, has yet to disclose a single molecule. In BioSpace's words, Isomorphic Labs has never revealed an asset and has never reached the clinic. What the company has stated is that preclinical results have come out of internal oncology and immunology programs. In January, Hassabis said at Davos that a first trial would begin by the end of 2026, and that target had already moved once, from the end of 2025. At this announcement he offered no updated clinical timeline, saying only that regulatory filings are on schedule and that further news will come in 2027.
That is not to say nobody outside the company has priced this engine. Eli Lilly and Novartis each signed research collaborations in January 2024, and on January 20, 2026, Johnson & Johnson joined with a collaboration spanning small molecules, antibodies, and peptides. The shape of that money differs from an equity round. Lilly paid $45 million upfront with up to $1.7 billion in milestone payments tied to development stages, and Novartis paid $37.5 million upfront on three targets with up to $1.2 billion in milestones. Across the two deals, $82.5 million actually changed hands, and the remaining $2.9 billion arrives only once a drug clears those stages. Novartis added up to three more research programs on the same terms in February 2025, and the Johnson & Johnson deal came with no disclosed figure.
So the investors did not pay for a drug. They paid for the engine that designs drugs, and for the expectation that the engine keeps improving. The pharmaceutical contracts, where money moves only once a result arrives, pay on entirely different terms.
Hassabis Calls the Demand for More Data a Crutch
Hassabis himself has said more than once where he locates his company's advantage. In December 2024, in an Endpoints News piece reported from the London office, he said of other AI biotechs, naming none of them, that they always say they need more data, and he added that this sounds to him like something of a crutch. His instruction was to make the algorithms and the models better, and that instruction carried a condition. Rework the algorithm side far enough, he argued, and the data already in hand is enough. He placed the shortage not in the volume of data but in the skill applied to it.
This position is of a piece with how the company is built. Isomorphic Labs runs no laboratory of its own, and chief scientific officer Miles Congreve explained in the same piece that the choice was deliberate, since setting up a lab would pull attention away. Where others talked about bigger laboratories, this company talked about better models and better people.
Then on October 7 of this year the same company's name appeared on an announcement of the opposite character. The Chan Zuckerberg Biohub's expansion of its Virtual Biology Initiative drew in the US Department of Energy and the National Institutes of Health for $1.8 billion in total, and within that, Google DeepMind, Meta, and Isomorphic Labs committed $300 million combined. The per-company amounts were not disclosed. The purpose of the venture is to manufacture new training data on how cells respond to stimuli. It fills the very shortage behind the request Hassabis called a crutch.
Speaking for Isomorphic Labs in that announcement was president Max Jaderberg, and his sentence pointed at the data side. "Generating the data to solve predictive systems biology requires scaling past the limits of what any single organization can produce today." Pebblous covered that announcement separately in an article on the Biohub.
The Investors Bought the Claim That More Data Keeps Helping
Konstantine Buhler, the Sequoia Capital partner who co-led this round, said in the official announcement that Isomorphic Labs is building the foundational technology behind the future of medicine. The more specific account of the investment case came in a Bloomberg interview. According to the coverage, he said he invested because the company had found scaling laws for its models, which is to say that its performance rises as the compute and the data behind it grow.
Setting the two statements side by side sharpens the difference. The founder locates the advantage in models and people and marks down requests for data as a crutch, while the investor names compute and data as the inputs to performance. Neither is easy to call wrong. But two sentences describing one company point at different things, and the $5 billion followed the second sentence.
Where the money goes matches that second sentence too. At the Series B in May, the company said the funds would keep scaling a drug design engine called IsoDDE, extend it to pharmaceutical partners, and push the oncology and immunology programs toward the clinic, and the November announcement added building clinical development infrastructure to that list. Scaling the engine comes down to feeding it more compute and more data.
Predicting a Structure and Becoming a Drug
The engine's performance is public as a number. Releasing the IsoDDE technical report on February 10, 2026, the company claimed the first step change in structure prediction accuracy since AlphaFold 3. The central sentence of the report is that on a demanding protein-ligand generalization benchmark, IsoDDE more than doubled the accuracy of AlphaFold 3.
Where that doubling comes from sits in Table 4 of the report. On the split least similar to the training data, which is to say the hardest one, IsoDDE got 30 of 60 cases right for a 50% success rate, against 23.3% for AlphaFold 3. Of the 30 it got right, 17 were cases AlphaFold 3 missed. The sample is 60 cases, though, and the 95% confidence interval on the gap between the two models runs from 8.3 to 43.3 percentage points. The direction is clear and the width is still wide open.
Checking that figure from outside is difficult. AlphaFold 2 and 3 came with papers and research code, while IsoDDE has released no code, no weights, and no API, and it has not been peer reviewed. Nature noted that the technical report says little about how anyone would reproduce the same results.
The larger problem is that accuracy at predicting structure does not carry over intact into the performance of a drug. A benchmark published in Molecular Systems Biology in 2022 showed that gap early. The researchers attached molecular docking to structures predicted by AlphaFold 2, predicted binding between 12 essential E. coli proteins and 218 antibacterial compounds, then checked those predictions against enzyme activity assays. Across all 12 proteins the AUROC ran from 0.18 to 0.71 and averaged 0.48. Random guessing sits at 0.5.
Drawing a structure and ranking which compound binds more tightly are different problems. That is precisely the point IsoDDE aims at, and the report states that its binding affinity prediction has passed physics-based methods. Even if that holds, one layer remains. Whether a compound works inside a human body is not a question a benchmark answers.
Phase 1 Went Up and Phase 2 Did Not Move
There is already a tally of what AI-discovered drugs do inside a human body. Jayatunga and colleagues at Boston Consulting Group tracked 75 AI-discovered molecules that entered trials from 2015 onward and published the results in Drug Discovery Today in 2024. The Phase 1 pass rate was 80 to 90%, far above the industry range of 40 to 65%. The Phase 2 pass rate, though, was about 40%, no different from the industry. The authors attached a caveat about the small sample.
The two phases ask different questions. Phase 1 looks mainly at safety and at how a compound moves through the body. That is work on the properties of a molecule, the very thing these models are trained to optimize. Phase 2 asks whether the drug works on a human disease. That answer is held by data coming out of patients rather than by molecular structure, and no amount of design can produce that data ahead of time.
The scorecard for the sector as a whole points the same way. As of November 2026 the US Food and Drug Administration has approved no AI-designed drug. Even the count of programs that have entered trials differs from tally to tally. The Boston Consulting Group analysis counts about 75, while other tallies record 173 or more than 200. The spread comes not from new data arriving but from the absence of any agreement on what counts as AI-discovered. Counting only programs where AI found the target, or counting any molecule refined computationally, moves the number by nearly threefold.
This is also why the $5 billion is no freak outlier. When Xaira Therapeutics launched in April 2024 with more than $1 billion in committed capital, it had no clinical assets either. Pricing a platform and a team where clinical evidence does not yet exist is the ordinary trade in this sector, and this round is the largest instance of it.
Why Pebblous Is Watching This Boundary
To a team that has spent its working life on data quality, the heart of this story is not the amount of money but the gap between Phase 1 and Phase 2. That gap makes visible the line where the model's achievement ends and the data's limit begins. The models clearly won at the layer where the properties of a molecule are tuned, and that shows up in the Phase 1 pass rate. On the question of whether a drug works on a human disease they have not won yet, and that question is answered by data from patients, not by design work.
The doubling claimed for IsoDDE reads on the same scale. The value was measured on one split of a benchmark the company chose, with 60 cases. The human body lies outside that split. A high benchmark score and a guarantee that the score travels are separate things, and what separates them is how well the evaluation data represents actual use. In a field like the life sciences, where failure is expensive, answering that question is worth more.
The same question holds well outside drug discovery. Three things are worth confirming whenever an AI result is being assessed.
- Does the data behind the reported score represent the distribution of the real operating environment, or was it assembled from the range where the model performs well?
- Are the sample size and the confidence interval published alongside that number? The doubling in IsoDDE carries an interval running from 8.3 to 43.3 percentage points.
- Is the layer the model handles well kept separate from the layer the data has to answer? Without that separation, success in Phase 1 reads as success in Phase 2.
Calling data a crutch while funding a venture that manufactures data looks less like a contradiction than a matter of sequence. Push the model as far as it will go and what is left is data, and that point has now arrived. What the $5 billion is buying is time to build the data for the layer that comes next.
References
The Endpoints News piece is behind an access block that could not be opened directly, so the remarks by Hassabis were confirmed through secondary sources citing that article. The remarks Buhler of Sequoia made in the Bloomberg interview were likewise confirmed through coverage citing that interview.
Primary Sources
- 1.Isomorphic Labs. (2026-11-17). "Isomorphic Labs announces Series C investment round." Company announcement.
- 2.Isomorphic Labs Team. (2026-02-10). "Accurate Predictions of Novel Biomolecular Interactions with IsoDDE." Technical report, Zenodo DOI 10.5281/zenodo.18606681. The 50% against 23.3% and the confidence interval in Section 4 are cited from Table 4 of this report.
- 3.Isomorphic Labs. (2025-02-18). "Isomorphic Labs announces Novartis collaboration expansion." Company announcement. Up to three additional research programs on the same financial terms as the original deal.
- 4.Isomorphic Labs. (2026-01-20). "Isomorphic Labs enters into a research collaboration with Johnson & Johnson." Company announcement. A multi-target collaboration spanning small molecules, antibodies, and peptides, with no disclosed figure.
Industry and Press
- 5.TechCrunch. (2026-11-17). "Isomorphic Labs raises $5B as it looks to take drug candidates to clinical trials." Source for the framing of this round as the first raise since Hassabis stepped back from running DeepMind.
- 6.BioSpace. (2026-11-17). "Isomorphic Labs Breaks Biotech VC Fundraising Record With $5B Series C." Source for the previous record of $3 billion at Altos Labs and for the fact that the company has not reached the clinic.
- 7.BioSpace. (2024-01-08). "Lilly, Novartis Sign AI Partnership with Alphabet's Isomorphic." Source for the upfront and milestone figures in Section 1 ($45 million and up to $1.7 billion for Lilly, $37.5 million and up to $1.2 billion for Novartis).
- 8.TNW. (2026-11-17). "Isomorphic Labs raises $5bn from Sequoia and Greenoaks for AI drug design." Source for Buhler's remarks on scaling laws and for the mention of further news in 2027.
- 9.Dunn, A. (2024-12). "Inside Isomorphic Labs: Demis Hassabis' lab-free vision for biotech's AI future." Endpoints News. Source for the crutch remark, the condition attached to it, and the lab-free strategy.
- 10.Engadget. (2026-08-05). "Google DeepMind CEO Demis Hassabis is stepping down, but will remain with the company." Source for the move to DeepMind chairman and Alphabet chief scientist and for the succession of Koray Kavukcuoglu.
Academic
- 11.Jayatunga, M. K. P., Ayers, M., Bruens, L., Jayanth, D., & Meier, C. (2024). "How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons." Drug Discovery Today, 29(6), 104009.
- 12.Wong, F., Krishnan, A., Zheng, E. J., Stärk, H., Manson, A. L., Earl, A. M., Jaakkola, T., & Collins, J. J. (2022). "Benchmarking AlphaFold-enabled molecular docking predictions for antibiotic discovery." Molecular Systems Biology, 18(9). Source for the average AUROC of 0.48.
Pebblous Blog
- 13.Pebblous. (2026-10-12). "Biohub Is Building the Cell Data Library AI Can Read." Source for the $1.8 billion Virtual Biology Initiative and the Jaderberg quote.