What public AI compute numbers actually say
Lab pages on the scoreboard show compute figures, gigawatts, megawatts and capital-expenditure dollars, but none of them feed into the 0-100 score. Here's what the public numbers actually say, and why they don't combine into one comparable figure.
The numbers, as published
| Lab | Metric | Value | As of | Source |
|---|---|---|---|---|
| xAI | Colossus 2, current IT power (est.), largest tracked | ~946 MW | 2026-10-15 | Epoch AI |
| Anthropic | Anthropic-Amazon New Carlisle, current IT power (est.) | ~910 MW | 2026-10-15 | Epoch AI |
| Meta | Meta Prometheus, current IT power (est.) | ~471 MW | 2026-10-15 | Epoch AI |
| Anthropic | Compute contracted since Oct 2025 | at least 14.8 GW | 2026-09-06 | The Information, via AI-RTZ |
| OpenAI | Planned capacity by 2030, told investors | 30 GW | 2026-09-09 | AI-RTZ |
| OpenAI | Compute spend forecast through 2030 | ~$750B | 2026-09-09 | AI-RTZ |
| Google (Alphabet) | 2026 capex guidance, company-wide | $195-205B | 2026-07-22 | CNBC |
| Meta | 2026 capex guidance, company-wide | $130-145B | 2026-07-29 | Meta Q2 2026 results |
| xAI | AI-segment capex, Q1 2026 | $7.7B | 2026-03-31 | TechCrunch (SpaceX S-1) |
| Alibaba | Group capex, June quarter | $9.975B | 2026-06-30 | Alibaba results |
Epoch AI's current estimate puts xAI's Colossus 2 as the largest single AI data center it tracks by IT power, at about 946 MW as of 15 Oct 2026, ahead of the Anthropic-Amazon site in New Carlisle (~910 MW) and Meta's Prometheus site (~471 MW). Those three numbers are at least directly comparable: same tracker, same metric, same date.
Nothing else on the table is. Anthropic's "14.8 GW contracted" figure, from The Information via the AI-RTZ newsletter, is a cumulative multi-year contracting total, not a current running capacity. OpenAI's 30 GW figure is a stated plan for 2030, four years out, alongside a $750B spending forecast over the same period. Alphabet's, Meta's and Alibaba's capex guidance are whole-company figures covering every business line, not an AI-lab-specific number; Alphabet's $195-205B, for instance, funds Google Cloud and other infrastructure beyond DeepMind.
Why the scoreboard doesn't score it
The methodology states the reason plainly: "Compute (GW, capex) is shown on lab pages but not scored, because the figures aren't comparable across labs." The table above shows exactly why: the units differ (instantaneous site power in MW vs. committed capacity in GW vs. dollars of capex), the timeframes differ (a current snapshot vs. a 2030 plan vs. a single quarter), and the scope differs (one data center vs. a whole company's infrastructure spend). Averaging or ranking across those would mix incompatible things into a single number, which is exactly what the scoreboard's "no guessing" rule is built to avoid.
One comparison that does hold up
Two of the Epoch AI figures share a tracker, a metric and a date, so they can be compared directly: xAI's Colossus 2 (~946 MW) and the Anthropic-Amazon New Carlisle site (~910 MW) are within about 4% of each other as of 15 Oct 2026. That compares two single sites, not either lab's total compute: Anthropic also has at least 14.8 GW contracted since Oct 2025, and xAI's other sites aren't in the board's data.
Sources
- Epoch AI: largest AI data centers by power capacity (opens in new tab) (as of 15 Oct 2026)
- AI-RTZ: Anthropic's $517B vs. OpenAI's $750B (opens in new tab)
- CNBC: Google Q2 2026 earnings (opens in new tab) (22 Jul 2026)
- Meta Reports Second Quarter 2026 Results (opens in new tab) (29 Jul 2026)
- TechCrunch: xAI burned $6.4B last year (opens in new tab) (20 May 2026)
- Alibaba Group Announces June Quarter 2026 Results (opens in new tab)
- Morningstar: xAI financials (opens in new tab)
- Anthropic's third-quarter revenue hits $16.8 billion (opens in new tab) (CNBC, 15 Oct 2026)