Sources and methods

Safety research outside the labs is underfunded

Our estimates for 2025, from the AI safety allocation project:

• The frontier labs’ research compute was about $33 billion (likely $27–41 billion). That is training and research at Anthropic, Google DeepMind, Meta, OpenAI and xAI, not compute that serves customers.
• About $742 million of it went to safety research (likely $0.4–1.5 billion), about 2%. This is the least certain part. Only Anthropic has measured its share: 6% of its research compute in a week of July 2026. The other labs’ shares are our estimates.
• Safety research outside the labs used about $59 million of compute (likely $44–90 million), about $1 for every $558 of the labs’ research compute. The main page rounds this to $500.
• Safety research means work whose main purpose is safety. Dual-use work, such as post-training, counts as capabilities.

Buying a big cluster is hard

From Geodesic Research’s post (16 September 2026) and its appendix on their procurement. Contracts of one, two or three years were the only options at the size they needed. Prices rose about 10% a month from May to August 2026. Offers disappeared within 24–48 hours, because nodes are not reserved until a contract is signed. The quote on the main page is from the post.

GPUs per safety researcher

H100-equivalents running all year, counting GPUs bought or donated, not spending on AI model APIs. No lab publishes compute per safety researcher, so these are rough, right to within a factor of two or three.

Anthropic. The main page uses the one lab where public figures allow the sum:

It leans low: the compute figure is from December 2025, and the 6% is from July 2026, when Anthropic had more. Much of it is large runs on big models (interpretability, evaluations and model organisms), the kind of work outside researchers could do on open models, given the compute. Inside the labs it also needs access to the closed models.

Published papers that report their compute, such as Anthropic’s circuit tracing and Google DeepMind’s Gemma Scope, come to roughly 1–30 per author. Large runs inside production training are not published.

Outside the labs. Our model averages about 2. Redwood Research’s 2024 tax filing lists $473,000 of GPU rental for about 7.5 researchers, about 2.4 each; it lists only vendors paid over $100,000, so that is a floor. MATS budgets $8,000–16,000 a month of compute per fellow, about 4–8. Academics typically have 1–8 GPUs for days at a time, often older ones.

GPUs for safety research outside labs

H100-equivalents running all year, counting GPUs bought or donated, not API spending. Our estimates, from grants, tax filings and organisations’ own figures.

• The two 2027 paths come from our compute forecast (September 2026). “Trend continues” is roughly 2022–26 growth, about 2.2 times a year, with AI automating coding around 2028–29. “Agents run experiments” assumes that happens in 2027 and that research agents multiply the compute each researcher uses. Both are low-confidence scenarios. GPUs aren’t the only limit: agents also need well-posed problems and ways to check their results.
• Agent use: Epoch AI (October 2026). The median OpenAI researcher’s coding-agent use, valued at API prices, went from under $1 a day in January to $601 a day by August.
• All AI chips (the “All chips” button): Epoch AI chip sales (August 2026), every AI chip sold since 2022, in H100-equivalents, averaged over each year. 2027 carries the recent trend forward. Chips sold before 2022 are missing, so 2022–23 are too low. In 2026, safety research outside the labs used about one in ten thousand of them.

Prices

Median listed price per GPU-hour on Vast.ai, an open marketplace of mostly small hosts, from the public 500.farm exporter. It is the price to rent single machines; big clusters are sold elsewhere, on long contracts. October 2026 covers the first eleven days. Their dashboard shows prices with Vast’s fee taken out, about a quarter lower, split into rented and free GPUs, so its numbers look lower than ours. From October 2025 to October 2026, B200 went from $3.98 to $7.72 and H200 from $2.61 to $5.04; A100 stayed around $0.90.

Timelines

AI Futures Project, Q2.5 2026 update (16 August 2026). An automated coder is an AI that can do the work of an AI company’s coders. Their forecasts assume progress goes as fast as is technically feasible. Medians: Kokotajlo November 2027, Halstead January 2029, Lifland January 2030. They publish 10th, 50th and 90th percentiles; our curves pass through those three points, and the shape in between is our smoothing.

Methods

The estimates come from our AI safety allocation project, which has the full method, data and sources.