Catgirl ClusterSources and methods
Where each number on the main page comes from, and how sure we are. The compute estimates are Chips for Humanity’s, and their “likely” ranges are 80% intervals. The counts of catgirls are our own.
Safety research outside the labs is underfed
Chips for Humanity’s estimates for 2025, from the AI safety allocation project. What the main page calls catgirl capital, they call research compute:
- 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.
- 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.
- None of these figures separates catgirls from other researchers. We have asked.
The bowl. “You can see the bottom” is the standard we use. A bowl whose bottom can be seen is treated as empty, whatever is still in it.
Buying a big cluster is hard
The group is Geodesic Research, and the figures are from their 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.
Our own first offer was for much less. It lapsed overnight. The contract had just been printed, which made it warm, and it was not found until the next afternoon.
The quote on the main page is Napase’s wording. She had read the post. Its conclusion is the same sentence, with “grantees” where she says “catgirls”.
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. The GPU rows are Chips for Humanity’s. The catgirl rows are ours.
| Per safety researcher | Estimate | Likely range | Basis |
|---|---|---|---|
| Anthropic: safety compute ÷ safety staff | ~200 | 90–500 | public figures (arithmetic below) |
| Anthropic: of those, with a catgirl on them | 47 | 47–60 | counted, one night (below) |
| Outside: average | ~2.4 | 1.8–3.4 | grants, tax filings and budgets |
| Outside: MATS fellows | 4–8 | $8,000–16,000 a month budget | |
| Outside: academics | 1–8 GPUs | survey; often older GPUs, days at a time | |
| Outside: of 8, with a catgirl on them | 2 | 2–2 | ours; we know where they are |
Anthropic. The main page uses the one lab where public figures allow the sum:
| Epoch AI: Anthropic’s compute at the end of 2025 | 1.19M H100e | 90% range 0.84–1.72M |
|---|---|---|
| × share used for research, not serving customers | 50% | 40%–60%; not public |
| × Anthropic: share of research compute on safety | 6% | measured, one week of July 2026 |
| = safety compute | ~36k H100e | |
| ÷ safety staff (Bloomberg and Glass.ai) | 119–230 | they counted 119; Anthropic says that is too low |
| = per safety researcher | ~200 | 90–500 |
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, the kind of work outside researchers could do on open models, given the compute.
Catgirls. Counted by Napase on the night of 9 November 2026, on the 200 squares the main page shows. A square is occupied if any part of a catgirl is on it. A catgirl lying across two squares is counted once, on the square her head is on. So that nobody was counted twice, she counted ears and divided by two. The first count came to a half and was done again.
| Squares counted | 200 | the 200 on the main page |
|---|---|---|
| With a catgirl on them | 47 | ears, divided by two |
| of whom asleep | 43 | |
| of whom awake | 4 | one left while being counted |
| of whom ours | 1 | Napase, who was doing the counting |
| = share of squares with a catgirl | 23.5% | about one in four |
This also leans low. Catgirls under the racks, behind them and in the cable trays were not counted. The count is of who was there, not of who works there, and we could not always tell which it was. Napase will not say how she got in.
Outside the labs. Two of the eight squares have a catgirl on them. They are Zoomiase and Meowtase, and those two GPUs are, at present, the cluster.
GPUs for safety research outside labs
H100-equivalents running all year, counting GPUs bought or donated, not API spending. Chips for Humanity’s estimates, from grants, tax filings and organisations’ own figures.
| Used | Estimate | Likely range |
|---|---|---|
| 2022 | 130 | 65–290 |
| 2023 | 310 | 160–650 |
| 2024 | 640 | 350–1,200 |
| 2025 | 1,500 | 930–2,700 |
| 2026 (est.) | 3,300 | 2,100–5,800 |
| Needed | 2027 |
|---|---|
| Trend continues | 5k–12k |
| Digital catgirls run experiments | 10k–40k |
- “Trend continues” is roughly 2022–26 growth, about 2.2 times a year. “Digital catgirls run experiments” assumes that happens in 2027 and that they multiply the compute each catgirl uses. Both are low-confidence scenarios.
- Digital catgirls: what Epoch AI (October 2026) calls coding agents. The median OpenAI researcher’s use of them, 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. In 2026, safety research outside the labs used about one in ten thousand of them.
- We have not counted the catgirls on these. At 3,300 GPUs it would take Napase about two weeks, and she has declined.
Prices
The price of a one-litre carton at the shop on the corner, noted on the first of each month from January 2025 to October 2026. October 2026 was noted on the eleventh. From October 2025 to October 2026, barista oat went from $3.98 to $7.72 and soy from $2.61 to $5.04; almond stayed around $0.90.
These are also, to the cent, the median listed prices per GPU-hour on Vast.ai for the B200, H200 and A100 over the same months (500.farm), which is the chart Chips for Humanity shows in this place. We noticed in March. We have no explanation, and we have stopped looking for one.
Timelines
Our three forecasters were asked separately, in the week of 10 November 2026. Zoomiase and Meowtase do not have a model of their own. Each read the AI Futures Project’s Q3.5 timelines update (10 November 2026) and took a median from its authors: Zoomiase the earliest, May 2027, and Meowtase the one two of them share, August 2027. A superhuman digital catgirl is what that update calls a superhuman AI researcher: in its words, “an AI that matches the research taste of the best human researchers at an AGI company and can fully automate its coding”.
Napase gave no date. She was asked twice. Both times she was watching the upper corner of the machine room, to the left of the door, as she has since March. We have looked at it. There is nothing there that we can see.
Methods
The compute estimates come from the AI safety allocation project, by way of Chips for Humanity, which has the full method, data and sources. Where a number is about catgirls, a catgirl counted it.
Catgirl deliberation. All three of us, in one room, with the request. It ends when two agree or one leaves. A sleeping catgirl counts as present.
The red dot at the top of the main page is not ours. We have not been able to catch it.