For most of human history, intelligence has been one of the world's scarcest resources.
If you had an exceptionally difficult problem, you could hire smart people, fund a research team or build a great organisation. But you could not simply ask for another 10,000 mathematicians for the weekend. Expertise took decades to create, and even the richest organisations were constrained by the number of capable people available.
That constraint may be starting to change.
OpenAI recently described an experiment in which roughly 10,000 AI agents were deployed concurrently across difficult mathematical problems, including work on the Navier-Stokes equations. The agents ran for about 88 hours, exchanged millions of messages and consumed an extraordinary amount of inference. The Navier-Stokes effort alone reportedly involved around 130 billion output tokens and 2.7 million agent messages.
The mathematical result will, quite properly, be scrutinised by mathematicians. That is not actually the part I find most interesting.
The interesting part is that someone decided this problem was valuable enough to justify spending millions of dollars of machine reasoning on it.
We are used to buying compute. We may now be learning how to buy intelligence.
We already know how to spend compute
For decades we have thrown computation at difficult problems. Weather forecasting, engineering simulation, protein folding, rendering, financial modelling and cryptography all consume enormous amounts of compute.
But traditionally the machine has been executing an algorithm that humans already defined.
Agentic AI changes that relationship. We can increasingly spend compute on something closer to searching through ideas.
In the OpenAI experiment, agents were not simply running the same calculation thousands of times. Groups of agents explored different approaches, promising ideas were shared, weak paths were abandoned and resources were shifted toward approaches that looked more productive.
That is an important distinction.
We are not just scaling calculation anymore. We are starting to scale reasoning effort.
The important number is not 10,000
It would be easy to describe this as 10,000 artificial minds working on a problem. I do not think that is accurate or particularly helpful.
Ten thousand agents are not the same thing as ten thousand mathematicians.
They may be based on the same underlying model. They can repeat the same mistakes. They do not represent the accumulated independent experience, intuition and diversity of 10,000 human experts. And simply adding more agents does not guarantee a proportionate increase in capability.
Coordination matters. Verification matters. The architecture around the agents matters.
But none of that changes the important observation: there appears to be useful work that can be gained by allocating more inference, more agents and more time to a difficult problem.
It does not need to scale perfectly to matter.
Intelligence gets a budget
This is where things become economically interesting.
Organisations have always allocated human effort to problems. A project might justify two engineers for six weeks, ten lawyers for a major transaction or a research team for several years.
We may increasingly make the same type of decision with machine reasoning.
A routine task might justify a few cents of inference. A complex contract review might justify tens of dollars. A product design problem might justify thousands. A drug discovery problem could justify millions.
The exact numbers are not important. The shift in thinking is.
Intelligence begins to look less like a fixed resource and more like something that can be allocated according to the expected value of the answer.
The question becomes: how much intelligence is this problem worth?
Cheaper AI could mean we spend much more on AI
A few days ago I wrote about when AI will get cheap. One of the traps in that discussion is assuming cheaper inference will necessarily mean lower AI bills.
The opposite may happen.
When the cost of reasoning falls, many more problems become economically worth attempting. And once a problem is worth attempting, organisations may choose to spend dramatically more inference on it if the expected value of a better result is high.
OpenAI's own research organisation provides an early example. The company has described researchers increasingly using coding and research agents as part of their normal work, with heavy users consuming very large amounts of inference each day.
The people closest to frontier AI are not responding to better and cheaper models by using less intelligence.
They are using more of it.
This is similar to what happens elsewhere in computing. Falling costs rarely result in us doing the same amount of work more cheaply. They usually allow us to attempt things that were previously uneconomic.
This changes which problems are worth attempting
There are countless areas where the value of a better answer could vastly exceed the cost of the intelligence used to find it.
Drug discovery. Battery chemistry. Materials science. Chip design. Logistics. Mathematics. Engineering optimisation. Software architecture. Fraud detection. Legal discovery.
Human experts have always rationed their attention. A scientist cannot investigate every plausible hypothesis. An engineer cannot deeply analyse every possible design. A software architect cannot examine thousands of architectures before choosing one.
Machine reasoning makes much wider exploration possible.
We may increasingly use what I think of as speculative intelligence: spending AI inference exploring large numbers of ideas even though we know that most of them will lead nowhere.
That sounds wasteful until the value of finding the one useful result is high enough.
A pharmaceutical company does not care that 999,999 machine-generated ideas went nowhere if the millionth leads to a valuable drug.
The bottleneck moves
Abundant machine reasoning does not eliminate scarcity. It moves it somewhere else.
If AI systems can generate thousands or millions of hypotheses, designs, proofs or strategies, the next scarce resource may be verification.
Can the mathematical proof be checked? Can the molecule actually be synthesised? Does the engineering design survive physical testing? Is the business strategy sensible? Is the generated software secure?
OpenAI's mathematical work illustrates this nicely. Generating the candidate argument was not the end of the process. The proof also went through formalisation and machine verification work using Lean.
This may become increasingly normal.
The problem changes from:
Can we generate a possible answer?
to:
How do we reliably identify the valuable answers among everything AI can generate?
Human judgement, experimental capacity, high-quality data, physical laboratories, compute, energy and reliable verification mechanisms may all become more important as machine-generated ideas become abundant.
There will still be a price
None of this means intelligence has become unlimited.
Large agent systems are expensive. They duplicate effort. They require orchestration. They can reinforce one another's errors. More agents will eventually produce diminishing returns, and some problems simply cannot be solved by throwing more inference at them.
But that may still leave us with something new: an emerging market for different quantities and qualities of machine intelligence.
Just as organisations choose different levels of cloud compute today, they may eventually choose different levels of reasoning effort.
A cheap model might answer an everyday question immediately. A more capable model might spend several minutes researching a difficult problem. A collection of thousands of agents might spend days exploring something important enough to justify the cost.
Those are not simply different products. They are different amounts of intelligence being allocated to a task.
What happens when intelligence becomes a line item?
Companies currently budget for cloud infrastructure, storage, software licences, employees, consultants and external research.
It is not difficult to imagine machine intelligence becoming another explicit budget.
And that could make future AI expenditure much larger than today's forecasts suggest.
If spending $5 million of inference has a credible chance of producing a discovery worth $500 million, then $5 million is not expensive.
It is cheap.
This is why I think the most important thing about OpenAI's 10,000-agent experiment may ultimately have little to do with Navier-Stokes.
It may be an early example of intelligence becoming an allocatable computational resource.
For most of history, faced with a difficult problem, we have asked:
Who is smart enough to solve this?
We may increasingly ask a very different question.
How much intelligence is this problem worth?