The Nobel Was the Easy Part

There were two announcements at Google I/O this week, and they were told as one story. They are not one story. One was a pricing move that will compress a chunk of the API market by the end of the quarter. The other was a science narrative that will be cited in policy papers for the next decade and may, if it works, generate a kind of intellectual property that the AI industry has not yet learned to value. Google framed them under the same umbrella because the umbrella looks good on stage. I want to take them apart, because the strategic logic is different, and because conflating them obscures what each actually says about where DeepMind is going.

Let me start with the science narrative, because that is the louder claim, and the easier one to mistake for substance.

The shift Hassabis is selling. In 2024, Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for AlphaFold. The citation was specific: protein structure prediction, a fifty-year-old problem in computational biology, materially solved. The AlphaFold database now holds more than 200 million predicted structures, freely available. That is a real scientific achievement, externally validated by the most conservative prize committee in science, and it is the highest-water mark any AI lab has reached in a non-AI domain.

At I/O this week, Hassabis moved the frame. The pitch is no longer "we solved protein folding." It is now, in his words, that "AI can move from helping scientists work faster to actually doing science."¹ The product wrapper is something Google calls an "AI co-scientist," described as a multi-agent system that can generate, test, and refine hypotheses. AlphaFold 3 extensions were announced alongside. The umbrella phrase is "general scientific reasoning."

This matters because AlphaFold's credibility came from a specific source: CASP, the biennial protein-structure-prediction competition, where AlphaFold2 outperformed everything else in the field by a margin that left the assessors visibly surprised. That is what a real scientific benchmark looks like. It is adversarial, externally administered, and ruthless about whether the model actually did the thing. When Hassabis says the next chapter is "doing science" more broadly, the natural question is: doing what science, measured how, against whom?

The defensible version of Hassabis's argument requires naming the next AlphaFold-level benchmark. None was named.

The history here is not kind to scope-expansion claims. IBM marketed Watson as a cancer-curing AI in 2015. MD Anderson cancelled the project in 2017 after spending $62 million, with zero clinical outcomes to show.² The failure mode was not that Watson was bad at NLP, which it wasn't. The failure mode was that demonstrating one capability in one constrained setting was treated as a license to claim the general case. The Watson story is the case study every AI lab claims to have learned from. It is also the story that the I/O keynote was structurally closest to.

I am not saying DeepMind is IBM. The difference between AlphaFold and Watson-for-Oncology is the difference between a system that beat the field on a public benchmark and a system that won a marketing campaign. But the rhetorical move from a domain-specific win to a general claim is the same move, and it deserves the same scrutiny.

What would make the general claim serious. There is a version of the "AI co-scientist" architecture that would matter, and it is worth being precise about what it looks like. A closed-loop scientific system would generate a hypothesis, simulate it computationally, design a wet-lab experiment, ingest the results, and revise. The AlphaFold pipeline already does part of this for one narrow problem. Extending it to, say, materials discovery (where Google's GNoME project did publish 2.2 million new crystal candidates in 2023) or to enzyme design or to small-molecule drug discovery would constitute something qualitatively new.

The evidence from I/O on whether this is actually being built, as opposed to being talked about, is thin. The "co-scientist" framing is mostly product marketing language. There is no equivalent of the CASP result. There is no peer-reviewed paper attached to the announcement. There is a keynote and a demo. That is not nothing, but it is also not what AlphaFold was, and the difference matters.

Now the second story, which is the one I think will move money this quarter. Gemini 2.5 Flash is now the default model on the Gemini API. Input pricing is $0.15 per million tokens below 200k context, $0.40 above. Output is $0.60 to $3.50 per million depending on whether thinking mode is on. Gemini 2.5 Pro sits above it at $1.25 input and $10 output.³

For mid-tier API workloads where context stays under 200k tokens, this is a price compression of roughly 13x against Claude Sonnet 4 and 8x against GPT-4.1 on the input side. Output gaps are smaller but still material. Not every workload is input-heavy, but a great many agentic workloads are, because they involve stuffing large amounts of retrieved context into each turn. RAG pipelines, document review, code review across large repositories, customer support bots ingesting transcripts: these are all input-dominant, and they are the bread and butter of the paid mid-tier API market.

The price compression is not a science story. It is an inference-economics story, told on the same stage, to a different audience.

The science narrative is aimed at policymakers, regulators, journalists, and the talent market. (If you are a postdoc choosing between a DeepMind offer and an Anthropic offer, "we won a Nobel" is a recruiting line that "we have great unit economics on Flash" is not.) The pricing move is aimed at developers and procurement teams at the buying enterprise. These are different audiences, and the messages are doing different jobs.

I want to be careful about the pricing claim, because the counter-case is real. Two counter-cases, in fact.

The first counter-case. Gemini Flash pricing may not reflect a sustainable margin. Google's TPU stack gives it a structural cost advantage that nobody else can match without building comparable silicon, but "structural advantage" and "willing to lose money on Flash for two years to build the ecosystem" are different claims, and from the outside it is hard to tell which one is operating. If it is the second, then Anthropic and OpenAI do not need to match the price; they need to wait it out. The historical analogue is AWS pricing in 2008-2012, where the question of whether the prices were profitable or strategic was answered only in retrospect. We are in the unanswered part of the equivalent question for inference.

The second counter-case. Open-weight models are eating the mid-tier from a different direction. Llama 4 and Mistral's recent releases give enterprises the option to self-host at marginal compute cost, which for high-volume workloads is lower than any API price including Flash's. If you are running a billion-token-a-day RAG pipeline, the calculation that matters is whether your DevOps team can run Llama 4 on your existing GPU fleet, not whether Flash is cheaper than Sonnet. The Gemini Flash price cut may be squeezing Claude and GPT, but the floor underneath all of them is moving down regardless.

Both counter-cases are real. Neither of them, in my view, changes the conclusion that Flash is the actually-consequential announcement from this I/O. The science narrative is a ten-year story whose payoff is uncertain. The pricing change is a this-quarter story whose effects will be visible in Anthropic's and OpenAI's response within weeks.

Back to the science, because there is a serious version of the moat argument and I want to give it its strongest form. If DeepMind's "AI co-scientist" tooling produces novel compounds, novel materials, novel enzymes, novel anything that is patentable, the intellectual property generated is non-fungible. It does not commoditise the way model weights commoditise. A patent on a specific molecule with a specific therapeutic effect is owned, defensibly, for twenty years. If Google is using its AI infrastructure to generate a portfolio of such patents at scale, that is a moat of a kind the industry has not previously seen from a model lab.

The interesting question is not whether Gemini is better than Claude. The interesting question is whether DeepMind is filing patents on AI-discovered science, at scale, and on what.

This is the thread worth pulling, and it is also the thread on which the public evidence is thinnest. DeepMind published over 500 papers in 2023, which is the most recent figure I have seen cited.⁴ How many patents that has produced, and in what domains, and with what commercial structure, is much harder to find. Isomorphic Labs, the DeepMind drug-discovery spinout, signed deals with Novartis and Eli Lilly in 2024 worth up to $3 billion in milestones; that is the closest public signal that the science-to-IP pipeline is operating commercially. But the deals are structured as services with milestone payments, not as DeepMind-owned drug portfolios. The IP question is not yet answered by the public record.

If I had to guess where this lands, I would guess that Isomorphic and its analogues across materials and biology are the actual moat strategy, and that the I/O "AI co-scientist" framing is the consumer-facing wrapper over a B2B IP business that is harder to talk about on stage. The keynote claim is "AI is doing science." The actual business is "we are running a portfolio of science-IP-generating partnerships, and the IP we generate is owned in ways that survive model commoditisation." Those are very different stories, and only one of them is in the public materials.

The counter to my counter, which I should name. There is a reading of the AlphaFold database release where the moat argument fails entirely. DeepMind made 200 million protein structures free. That is the opposite of moat behaviour; it is ecosystem-building. If the AI-for-science strategy continues in that direction, the IP value accretes to the academic labs, pharma companies, and biotech startups that build on the public data, not to Google. The freely-available database is a deliberate gift, and there is no obvious mechanism by which Google captures the derivative value.

I find this counter half-persuasive. The half I find persuasive is that AlphaFold-the-database is genuinely a public good, and the moat argument cannot rest on it. The half I do not find persuasive is the implication that everything DeepMind does will follow the same pattern. The AlphaFold database was a credibility play, made at a moment when DeepMind needed to establish that AI-for-science was real. Mission accomplished. The next-generation tooling (AlphaFold 3, Isomorphic's pipelines, GNoME's materials candidates) is being released on much more controlled terms, and the commercial wrapping is visible if you look for it.

The AlphaFold database was a gift. The next chapter is not.

Where this leaves the reader. Two things to track, in my view, neither of which was the headline at I/O.

The first is whether Anthropic and OpenAI respond to Flash pricing within the quarter. They have three options: match, segment (introduce a cheap-tier model that undercuts Flash on specific workloads), or hold and bet that quality differentiation survives the price gap. Each option tells you something about how they read their own unit economics. Watch for it.

The second is whether DeepMind, Isomorphic, or any of the adjacent Google science efforts produce a peer-reviewed result, a regulatory submission, or a patent filing that is concretely attributable to the "AI co-scientist" tooling rather than to AlphaFold-the-existing-system. That is the test of whether the I/O framing was real or aspirational. The answer will arrive in months and years, not days, and it will not arrive on an I/O stage. It will arrive in journals and patent gazettes, which are exactly the unglamorous places where AlphaFold's reputation was built in the first place.

The Nobel was the easy part. The hard part is doing it again, in a different domain, without the fifty-year-old benchmark waiting to validate you when you arrive.

Footnotes

¶