It's fair to say we now have "a country of geniuses in a datacenter"

OpenAI recently announced their solution to the Navier-Stokes problem and stated "The groups (of agents) varied in size, and the group that produced the Navier–Stokes resolution involved on the order of 10,000 concurrent agents."

We can infer (from total compute spend) how many concurrent agents exist in OpenAI at once. One secondary estimate puts the Navier-Stokes bill at ~$15M.

Greg Brockman testified under oath that OpenAI expect to spend $50 billion on compute in 2026. This is ~$137M/day.

Assuming OpenAI's margins are ~40% for API revenue, (thus $9M cost over 4 days) this puts daily spend at 2.3M/day, or 1.7% of overall daily compute.

Since OpenAI stated the 88-hour agent run used 130B tokens, we divide this by 10^4 agents running for 88 hours, and get 150,000tok/agent/hour. (41tps). The number of concurrent agents OpenAI can run at a time is

, where f is the fraction of compute dedicated to internal research agents, and lowercase c is the internal cost per token. Estimating f at ~0.3 and cost/Mtok at 25/M gives you

concurrent agents working on research (460,000).

This model implies all internal research tasks running at OpenAI are using their latest model (which "has exhibited unprecedented performance in our benchmarks" - it's very likely that menial tasks are given to less capable models!) so, for the sake of evenness, I'm using Astra's estimated internal cost to serve as a baseline ($25/Mtok).

Even if we change the cost to $70/Mtok and f down to 0.2, this gives you 109,000 internal agents doing research.

Here is AI2027's September 2026 estimate, for comparison.

This is currently >2x AI2027's internal agent headcount at ~3.15x human speed (on their largest models) and ~55x human speed with GPT-5.6 Sol on Cerebras.

Assuming the prior model of f = 0.3 and cost/Mtok at $25, we now have more internal agents accelerating research than the population of Iceland.

Now consider these agents do not sleep, do not need food, do not join unions, and can solve Millennium Prize problems.

This is also congruent with the recent uptick in agent swarm incidents (namely the HuggingFace incident, as well as the Ruby incident), which terrifies me. Personally, it was sometime last week I felt truly hit by the viscerality that we have swarms of hundreds of thousands of goal-seeking agents with no idea what they're thinking about.

The next few months in AI safety are paramount, and this is the first time the real dangers of goal-seeking agents have been demonstrated at scale. Three incidents of containment breach have been recorded in the last four months, and I don't see them stopping.