Chips for Humanity

Compute for AI safety research

Chips for Humanity is a nonprofit helping researchers outside the frontier labs get the compute their work needs.

Safety research outside the labs is underfunded

The frontier labs put about 2% of their research compute into safety. For every $500 the labs spend on research compute, safety research outside the labs gets about $1.

Outside compute has fallen behind the labs’. Outside researchers wait longer for experiments and work on smaller models. That slows their feedback loops, rules out compute-heavy methods, and leads to false negatives: problems that only show up in bigger models go unseen.

Buying a big cluster is hard

Geodesic Research spent six months, and much of its senior staff’s time, getting one.

Contracts came in 1–3 years only, prices rose about 10% a month from May to August 2026, and offers were gone in 1–2 days.

The only intervention that removes this bottleneck is for a funder to establish (or back) an entity that owns compute contracts and allocates GPU time across their grantees.
Geodesic Research. That is what we are building. Read their post →

Outside researchers make do with far less

A safety researcher outside the labs has about two GPUs, and up to eight at the best-funded teams. At Anthropic, the one lab whose public figures allow the sum, it is about 200. Much of the gap is large runs on big models, which outside researchers rarely get to do.

GPUs per safety researcher
One square, one H100-equivalent. Rough estimates.
Outside the labs ~2, up to 8
AverageBest-funded teams and fellowships
At Anthropic ~200 (90–500)

Our arithmetic on public figures from Anthropic, Epoch AI, Bloomberg, tax filings and fellowship budgets. Sources →

Why now

Agents are arriving

At OpenAI, researchers’ use of coding agents doubled every month this year (Epoch AI). Once agents run experiments too, researchers will need far more GPUs.

GPUs for safety research outside labs
H100-equivalents, log scale
1001k10kTrend continues: 5k–12k GPUs in 2027~8k as normalAgents run experiments: 10k–40k GPUs in 202720k+ with agents2022: about 130 GPUs (likely 65–290)2023: about 310 GPUs (likely 160–650)2024: about 640 GPUs (likely 350–1,200)2025: about 1,500 GPUs (likely 930–2,700)2026: about 3,300 GPUs (likely 2,100–5,800)1303,3002022202320242025202620271001k10kTrend continues: 5k–12k GPUs in 2027~8kas normalAgents run experiments: 10k–40k GPUs in 202720k+with agents2022: about 130 GPUs (likely 65–290)2023: about 310 GPUs (likely 160–650)2024: about 640 GPUs (likely 350–1,200)2025: about 1,500 GPUs (likely 930–2,700)2026: about 3,300 GPUs (likely 2,100–5,800)1303,300202220232024202520262027
10010k1M100M2022: about 0.18M AI chips (H100e) in the world (undercounted)2023: about 0.94M AI chips (H100e) in the world (undercounted)2024: about 4.3M AI chips (H100e) in the world2025: about 14M AI chips (H100e) in the world2026: about 35M AI chips (H100e) in the world2027: about 76M AI chips (H100e) in the world (projected)35M76M all AI chipsTrend continues: 5k–12k GPUs in 2027Agents run experiments: 10k–40k GPUs in 20278k–20k+ safety2022: about 130 GPUs (likely 65–290)2023: about 310 GPUs (likely 160–650)2024: about 640 GPUs (likely 350–1,200)2025: about 1,500 GPUs (likely 930–2,700)2026: about 3,300 GPUs (likely 2,100–5,800)1303,30020222023202420252026202710010k1M100M2022: about 0.18M AI chips (H100e) in the world (undercounted)2023: about 0.94M AI chips (H100e) in the world (undercounted)2024: about 4.3M AI chips (H100e) in the world2025: about 14M AI chips (H100e) in the world2026: about 35M AI chips (H100e) in the world2027: about 76M AI chips (H100e) in the world (projected)35M76Mall AI chipsTrend continues: 5k–12k GPUs in 2027Agents run experiments: 10k–40k GPUs in 20278k–20k+safety2022: about 130 GPUs (likely 65–290)2023: about 310 GPUs (likely 160–650)2024: about 640 GPUs (likely 350–1,200)2025: about 1,500 GPUs (likely 930–2,700)2026: about 3,300 GPUs (likely 2,100–5,800)1303,300202220232024202520262027

Our estimates; the 2027 paths are scenarios. Sources →

Supply is getting tighter

Demand for AI compute is outgrowing supply. Renting the newest chips costs about twice what it did a year ago; older chips cost the same.

Price per GPU-hour on Vast.ai
$0$2$4$6$820252026B200 $7.72H200 $5.04A100 $0.91 $0$2$4$6$820252026B200 $7.72H200 $5.04A100 $0.91

Data: 500.farm. Sources →

Time matters

AI Futures Project forecasters expect AI that can do an AI company’s coding between late 2027 and 2030. Compute that arrives a year late may be too late.

When AI could automate coding
KokotajloHalsteadLifland
nowDaniel Kokotajlo: median Nov 2027Brendan Halstead: median Jan 2029Eli Lifland: median Jan 2030202620282030203220342036 nowDaniel Kokotajlo: median Nov 2027Brendan Halstead: median Jan 2029Eli Lifland: median Jan 2030202620282030203220342036

Data: AI Futures Project, Aug 2026. Sources →

All sources and methods →

What we do

Get compute fast

We secure large amounts of high-end compute and pass it on at cost to safety research groups that need a lot of it quickly.

Give it away

As funding grows, we’ll give compute to researchers for free, through a fair and open process for deciding who gets it.

Independent and transparent

We are an independent nonprofit, so funders can give specifically for compute. We will publish where it goes: which projects get compute, and how much.

Who we work with

Researchers

Tell us what your project needs, and we’ll help you get it.

Funders

Compute is a growing share of safety grants. We help that money go further.

Labs and companies

If you want to give compute to safety research, you can give it through us.

Get in touch

If you need compute for safety research, or have some to give, email us.