SIREN LABS · THE COMPUTE CLOCK · NOT IN THE SCORE
How fast is frontier training compute growing?
A live clock of the largest disclosed training run, the record ladder, and the doubling time from Epoch AI’s notable-models data.
SIREN does not invent FLOP estimates. Epoch AI maintains the public catalogue; this page recomputes the record holders and an ordinary-least-squares doubling time on the frontier subset, and shows Epoch’s AI data-centre stock beside it. The numbers are watched, not scored into the main index.
RECORD LADDER · LOG₁₀ TRAINING FLOP
Each point is a model that set a new all-time high for disclosed training compute in Epoch’s catalogue. Gaps mean no larger run was recorded, not that nothing trained.
RECENT RECORD HOLDERS
| DATE | MODEL | ORG | FLOP |
|---|---|---|---|
| 2026-09-03 | GPT-6 Astra | OpenAI | 9.40×10^26 |
| 2025-07-09 | Grok 4 | xAI | 5.00×10^26 |
| 2025-02-27 | GPT-4.5 | OpenAI | 3.80×10^26 |
| 2025-02-17 | Grok 3 | xAI | 3.50×10^26 |
| 2023-12-06 | Gemini 1.0 Ultra | Google DeepMind | 5.00×10^25 |
| 2023-03-15 | GPT-4 (Mar 2023) | OpenAI | 2.10×10^25 |
| 2022-06-29 | Minerva (540B) | 2.74×10^24 | |
| 2022-03-15 | GPT-3.5 (davinci-002) | OpenAI | 2.58×10^24 |
| 2021-09-03 | FLAN 137B | Google Research | 2.05×10^24 |
| 2021-08-11 | Jurassic-1-Jumbo | AI21 Labs | 3.70×10^23 |
NEWEST WITH A FLOP ESTIMATE
| DATE | MODEL | ORG | FLOP |
|---|---|---|---|
| 2026-09-09 | DeepSeek V4.1 Flash | DeepSeek | 4.32×10^24 |
| 2026-09-03 | GPT-6 Astra | OpenAI | 9.40×10^26 |
| 2026-09-03 | WeatherNext 3 | Google DeepMind,Google Research | 2.07×10^23 |
| 2026-08-20 | GLM-5.3-Flash | Z.ai (Zhipu AI) | 3.24×10^24 |
| 2026-08-07 | Motif-3 | Motif Technologies | 1.10×10^24 |
| 2026-07-31 | K-EXAONE 2.0 | LG AI Research | 3.55×10^24 |
LARGEST AI DATA CENTRES (EPOCH)
| # | SITE | OWNER | H100-EQ | MW | |
|---|---|---|---|---|---|
| 1 | Colossus 2 | SpaceXAI | 1,111,673 | 946 | |
| 2 | Anthropic-Amazon New Carlisle | Amazon | 818,428 | 1,037 | |
| 3 | Microsoft Fairwater Atlanta | Microsoft | 768,769 | 636 | |
| 4 | Google Pryor (North) | 636,685 | 368 | ||
| 5 | DayOne Nusajaya | — | 583,837 | 473 | |
| 6 | Meta Prometheus | Meta | 536,343 | 471 | |
| 7 | OpenAI Stargate Abilene | Oracle | 509,348 | 421 | |
| 8 | DayOne Kempas | — | 480,639 | 422 | |
| 9 | Microsoft Fairwater Wisconsin | Microsoft | 445,680 | 369 | |
| 10 | Google Columbus | 331,986 | 303 | ||
| 11 | Google New Albany | 323,901 | 333 | ||
| 12 | Google Mesa | 289,540 | 183 |
IT power and H100-equivalents are Epoch’s independent estimates from satellite imagery, permits and disclosed chip counts. Coverage is partial by design.
Related: Takeover Watch · The power race · SIREN Labs · US data-centre map
Source: Epoch AI notable AI models and AI data centers, CC BY 4.0. Data from Epoch AI (https://epoch.ai/), used under CC BY 4.0. SIREN recomputes the doubling-time fit; Epoch publishes related trend estimates on epoch.ai. Refreshed up to twice a day. JSON. Updated 2026-10-21 16:16 UTC.