Nobody at a frontier lab is working on your disease

A Nobel-winning model, a decade of patience and billions of dollars still did not produce a drug. What was missing is the thing a seed-stage company can actually own.

After the last two pieces, one question kept coming back. The labs are walking into drug discovery. Do they just take the field?

A model does not produce a drug, not even the best one ever built. Frontier models are commoditising biological hypothesis generation, but what stays scarce is the loop that turns a hypothesis into a validated result, plus the endurance to carry one programme through the years where no such loop exists. Those are the two things the labs are now buying at enormous expense. They are also the two things a focused company can own.

The worry

Anthropic shipped Claude for Life Sciences in October 2025 and Claude for Healthcare in January. Vas Narasimhan, the Novartis chief executive, joined its board earlier in the year. In April it bought the drug discovery startup Coefficient Bio. On 30 June it said it would run its own preclinical drug programmes and launched Claude Science the same day. Last week it published wet lab results showing Claude designed working protein binders against 14 of 15 targets, at roughly twice the field’s hit rate. Meanwhile John Jumper, who co-led the AlphaFold team and shared the 2024 Nobel Prize in chemistry for it, left Google DeepMind for Anthropic in June after nearly nine years, and OpenAI put a biology-tuned model in front of Amgen, Moderna and Thermo Fisher.

If that is the competition, what is a seed-stage company supposed to defend?

Alphabet has been at this for twelve years across three vehicles, with more money committed to biology than Anthropic and OpenAI have put in between them. The three were built differently: one with unlimited time, one with full independence, one with a single job. None of those turns out to be what binds.

Patience was not the constraint

Google launched Calico in 2013 to work on ageing, the least tractable problem in medicine. Time put it on the cover and asked whether Google could solve death. AbbVie signed on in September 2014. By its own accounting in its 2024 annual report, it spent roughly $1.75 billion on the collaboration between 2013 and 2022.

In January 2025, Calico’s ALS drug fosigotifator did not slow the disease in a Phase 2/3 trial. Denali Therapeutics, working on the same biology, failed in the same trial. That November, STAT reported that AbbVie had ended the eleven-year partnership and cut around a hundred chemists.

Structure was not the constraint either

I implied in the first piece that the labs would prove too impatient for drug development. Alphabet was patient and still came up empty. It could fund twelve years of biology because Google Search prints money and Calico was a rounding error against group profits. That is a specific kind of balance sheet and not the one the current entrants have (yet).

The obvious lesson from Calico is that this work needs to sit outside the parent, on its own clock, with its own investors. Verily is the reason not to believe it.

Verily was the second vehicle and it had every feature separation is supposed to provide. Spun out of Google X in 2015 as its own company with its own chief executive, it took $800 million from Temasek in 2017 and $1 billion in a Silver Lake-led round in 2019, six years before Isomorphic raised a dollar externally.

In January 2023, months after a further $1 billion round led by Alphabet to expand in precision health, it cut about fifteen percent of staff and dropped most device work. “We are designing complexity out of Verily,” wrote chief executive Stephen Gillett. More cuts followed. In August 2025 it shut the devices programme and turned to AI and data infrastructure.

The org chart was never the problem. Verily came out of Google X doing everything at once: devices, clinical research, diabetes management, stop-loss insurance, surgical robotics. Every round moved money from the long uncertain programme toward the nearer one, because a nearer one always existed.

Focus worked, and it was still not enough

Isomorphic Labs is the third vehicle and the one Alphabet built properly. Out of DeepMind in 2021, with discovery deals with Novartis, Lilly and Johnson & Johnson, and $2.1 billion raised on 12 May 2026 in a round led by Thrive Capital.

What distinguishes it from Verily is not the corporate form. Isomorphic designs drugs. When a programme looks slow there is nowhere else for the money to go.

Isomorphic is raising billions partly to build wet labs. Demis Hassabis had aimed to have AI-designed drugs in trials by the end of 2025. It now expects its first trials by the end of 2026. Carrying one drug from discovery to approval runs to about $2.6 billion once you count the cost of capital, so its entire war chest is roughly the fully loaded cost of a single drug.

The best-constructed of the three is still queuing for the part that costs money and time.

Where the model stops

A predicted structure is not a drug. Between a predicted interaction and a molecule that survives a human trial sit years of bench work no model produces on its own.

The Anthropic experiment shows exactly where the line falls. Claude designed protein binders against 15 targets and produced working binders for 14, at hit rates between 22% and 35% depending on the setup, against a field norm Anthropic puts at 10% to 15%. It was not a protein model. It orchestrated RFdiffusion, ProteinMPNN and ESMFold2, all open source and in general use, working from a protocol prompt of roughly 16,000 words and up to 12,500 hours of H100 time.

Then it stopped. Adaptyv Bio and Twist Bioscience made every design and measured whether it bound, and Anthropic could not do that step. Adaptyv calls it an open-loop experiment. Claude designed, they tested, and that was the end. Design work that used to take a specialist months compressed into a 48-hour session. The measurement took weeks and had to be bought from two companies with robots.

Nothing in the result speaks to whether those targets matter, whether the molecules are stable, or whether any of them survive a human immune system.

Anthropic reached the same conclusion and acted on it. It bought a team, opened wet labs and hired biologists, which is not how you spend if you think a better model is enough.

What that leaves you to build

The labs can own drugs and they cannot own the field. Even total commitment from every frontier lab at once would leave almost all of medicine uncovered, because the unit of progress is a programme and programmes do not get cheaper by being run next to each other.

So the screen is one question, close to the test I set in the model piece. If a frontier lab entered your exact space tomorrow, would it be your competitor or your customer?

Competitor means your moat is the model, and the model is not yours. Customer, or acquirer, means you hold something they need and cannot make.

Two weeks ago I argued that data alone is the worst-paid position in this business, and I still think that. The acquirer route is the exception and it rests on a few buyers staying interested. So own the data and own a programme. Be the company making proprietary biology a lab would rather buy than rebuild, because Isomorphic raising billions for wet labs and Anthropic paying $400 million for ten computational biologists show what that costs even at their scale.

And hold each programme in a vehicle of its own, so the organisation bends around one drug at a time without the company riding on any of them. The legal wrapper is not what does the work. A team with one thing to work on cannot quietly move to the easier thing, which is what Verily did four times. The last piece got to the same shape from the other end, with Takeda buying a Nimbus subsidiary rather than Nimbus itself.

Why their success does not threaten yours

The market that buys those programmes runs on merit. On China’s own regulator’s numbers, Chinese biotechs licensed a record $135.7 billion of drugs to foreign buyers in 2025, mostly from companies with no name recognition in Boston whatsoever. Pharma bought unproven molecules in volume, from wherever they were. Safety failures and the politics can make buyers pay less for an asset because of where its data came from. They do not change what a working drug is worth.

Two things would change my mind. If the cost of carrying a drug collapsed, from $2.6 billion to say two hundred million. Or if proprietary data stopped being an edge because models got good enough to simulate the biology rather than needing it measured. Neither has happened.

The model generates hypotheses and there is no shortage of those. What is scarce is the loop that turns one into evidence, and the willingness to carry that evidence for the years it takes to become an asset. Own that and better models make your company more valuable rather than less.

Jan Buza
Partner @ Zaka