AI for Scientific Discovery Looks Like a Great Investment
The evidence so far points to substantial returns, and I expect those returns to improve as the systems get more capable and cheaper to use.
I think scientific discovery is already one of the strongest cases for spending money on AI. We have systems producing useful research tools, finding better algorithms and reporting solutions to long-standing mathematical problems. The costs are imperfectly disclosed, but the results are substantial enough to justify a much larger effort.
The recent Navier–Stokes announcement makes the economics especially interesting. In its September 8 account, OpenAI reports that a coordinated system of agents produced a proof of finite-time singularity for the forced, three-dimensional Navier–Stokes equations, addressing the breakdown alternatives in the Millennium Prize formulation. It reports a discovery time of about 88 hours, followed by another 17 hours for Lean formalization and verification.
That is OpenAI’s account of its result. I am not independently adjudicating the proof, and Clay’s problem page still labels it active as of September 29. The economic argument is conditional on the result holding up. A formal proof also needs to formalize the intended mathematical statement correctly.
With that qualification, this is an extraordinary research output to get from a run measured in days. What would it have cost?
OpenAI reports approximately 130 billion output tokens for the Navier–Stokes effort. At the current GPT-6 Astra Standard short-context API price of $50 per million output tokens, that is $6.5 million of output charges. The arithmetic is straightforward: 130,000 million tokens, multiplied by $50.
This is a public-price comparison. OpenAI used an internal model it describes as significantly more capable than Astra. The company has not disclosed its actual bill, total input tokens or hardware allocation. Input charges would add to that hypothetical bill; caching, service tiers and context lengths would change it. Pricing Astra’s output gives us a useful reference point without telling us what the actual system cost to operate.
There is also a larger denominator we should include. OpenAI reports 300 billion output tokens across all attempted problems, equivalent to $15 million at the same output rate. A research program has to pay for the attempts that fail as well as the one that makes the announcement. This broader total gives a better sense of the disclosed campaign, although it still excludes inputs and does not provide a complete project budget.
Even that larger figure looks attractive to me for a program producing a valid result of this importance. It is an inference about scientific value, not a calculation of cash profit. We do not have a dollar valuation of the theorem’s downstream benefits. We do have a reason to take seriously the possibility that a modest research budget can now buy an enormous amount of productive intellectual work.
A hardware-based cost model could put the direct expense well below the retail comparison. The assumptions are too uncertain to turn that into a claim that this particular run cost $300,000, or that AI was thousands of times more productive than human mathematicians. I have included the calculation below so the assumptions are inspectable. The case for doing more of this does not require that dramatic multiplier.
What interests me is the price of making further progress from here. A lab deciding whether to fund another attempt needs to compare the expected scientific contribution with the additional resources required. The possibility of obtaining major results at these costs is compelling even when the precise historical multiplier is unknowable.
Navier–Stokes also sits alongside evidence from systems whose usefulness is already visible in practice. In its 2022 AlphaFold database expansion, DeepMind and EMBL-EBI made more than 200 million predicted protein structures available. The announcement reported that more than half a million researchers had already accessed the database. Predicted structures still need appropriate validation, and they do not solve all of biology. They give many researchers a useful starting point without each lab having to recreate the capability.
AlphaEvolve provides a more direct economic example. Google DeepMind reports that a scheduling heuristic it discovered recovers an average of 0.7% of Google’s worldwide compute resources, and that a separate kernel improvement reduced Gemini training time by 1%. These are company-reported results, not an independently audited return calculation. But a useful algorithm deployed across a large computing fleet keeps producing savings after the discovery run ends.
That is a particularly attractive feature of this kind of research. The expense of finding something can be incurred once while its benefits accumulate across many uses. A proof becomes part of mathematics. A database supports thousands of projects. A better algorithm can improve every subsequent execution. Public scientific value can also spread far beyond the organization that paid for the work.
The cost of building the AI systems belongs in the accounting. Training, failed experiments, engineering and infrastructure are real investments. Evaluating the entire enterprise requires allocating those costs across what the systems produce. For a lab using an existing model, however, the decision to run another experiment has a different cost boundary. Both questions matter. Charging one discovery for an entire reusable platform, or pretending that platform was free, would give us a misleading answer.
I expect the returns to improve for several reasons. More capable models can raise the probability that an attempt produces something useful. Better inference software and hardware can reduce the cost of an attempt. Better tools, evaluators and coordination can reduce wasted work. Reusable discoveries can then make later research more productive. AlphaEvolve’s reported improvements to AI training are already an example of one part of that feedback loop.
None of those mechanisms guarantees that every future project gets cheaper. Researchers will spend more compute on harder questions. Some results will take years of physical experiments to validate. Cheap hypothesis generation is less valuable when the experiment needed to test each hypothesis is expensive. Mathematics and software have an advantage where checking can be automated, although choosing the right specification and evaluating significance still require care.
The measure I care about is useful, validated scientific progress per dollar, including failed attempts, checking and the human work needed to make the system effective. Token prices alone do not measure that. A more expensive run can be a better investment if it makes a previously inaccessible result attainable.
The visible successes are also a selected sample. They do not tell us the average return across every AI research project, and they do not establish that all scientific fields are equally ready. They do show that the opportunity has moved beyond a hypothetical capability. We can now fund research around systems that have produced useful outputs, measure their failure rates and expand the approaches that work.
My conclusion is that we should be putting substantially more effort into applying AI to scientific discovery. The evidence so far makes it look like a great investment. As capability, efficiency and research methods improve together, I expect that investment to become more attractive. I want more laboratories testing that expectation with real problems and publishing enough of the costs and outcomes for the rest of us to learn from them.
The cost model and the human-effort comparison
The underlying infrastructure expense could be considerably below the retail comparison. A simple hardware model illustrates the scale. Assume a rack costs $5 million, lasts three years and is productively occupied 80% of the time. Those are modeling choices, not disclosed OpenAI procurement terms. They imply roughly $238 per productive rack-hour in hardware depreciation: $5 million divided by three years, 8,760 hours per year and 80% utilization.
Here is what an 88-hour run would allocate under that model:
| Average occupied rack-equivalents | Hardware depreciation for 88 hours |
|---|---|
| 5 | About $105,000 |
| 20 | About $419,000 |
| 100 | About $2.1 million |
Electricity, networking, facilities, tools, staff and the additional formalization phase would add costs. Model training and development would add another category altogether. This is allocated hardware cost, not a complete operating budget or a pure marginal cash expense. Most importantly, we do not know which row, if any, resembles OpenAI’s deployment.
But “10,000 agents” does not settle that question. An agent is a software process; many inference streams can share accelerators. The disclosed output volume averages about 410,000 tokens per second over 88 hours. For a sense of scale, NVIDIA reports peak throughput above 4,000 tokens per second per GPU for Qwen3.8-2.4T-A95B on GB300 NVL72. With 72 GPUs per rack, the output rate is equivalent to roughly 1.4 racks at that benchmark throughput.
That comparison cannot establish OpenAI’s rack count or put an upper bound on it. These are different models, and long contexts, input processing, memory requirements and utilization can change the answer enormously. It does explain why I would not infer a gigantic hardware bill simply from the number of agents. A few-hundred-thousand-dollar run is a scenario worth considering; the public data do not establish it as the actual cost.
The human comparison is harder still. Nobody has kept a timesheet for the decades spent on Navier–Stokes regularity. As a thought experiment, 1,000 to 10,000 full-time-equivalent researcher-years, valued at $180,000 to $250,000 per year including overhead, produces a range of $180 million to $2.5 billion. A middle scenario of 4,000 years at $215,000 gives $860 million.
Those researcher-years are assumptions, not an estimate derived from a comprehensive census of the literature. The dollars value that assumed effort at present-day research costs; they are not a reconstruction of historical spending. And the scope must stay narrow: work on the theorem and directly related regularity questions. Including the entire history of computational fluid dynamics would inflate the comparison with work that solves different problems. Aircraft designers still need fluid simulations after a theorem is proved.
Dividing the $860 million human scenario by a hypothetical $300,000 AI run gives roughly 2,900. That is where a claim of “thousands of times cheaper” can come from. I would not publish it as a measured productivity gain. It divides two uncertain quantities with different accounting boundaries, and the earlier research helped create the knowledge the AI system could use. Nor does historical research spending tell us the value of the final proof.