A training run confirmed at 10^30 FLOP

Where things stand (13 Oct 2026)

I downloaded Epoch's all_ai_models.csv (https://epoch.ai/data/all_ai_models.csv, 3,626 rows, fetched 2026-10-13). The largest "Training compute (FLOP)" value is 1.0001e27 for GPT-6 Astra (published 2026-09-03, "Likely"). Epoch's note puts it at at least 100k GB200s at Abilene for about 90 days at 25% FP8 utilization, with a confidence interval of about 5e26 to 2e27. The next largest are Grok 4 (5e26, which itself sums Grok 3's pretraining plus reinforcement learning), GPT-4.5 (3.8e26) and Grok 3 (3.5e26). So the threshold is 1,000× the current record. Most 2026 closed frontier models have no compute estimate at all (Claude Opus 5.5, Gemini 4 Argon, GPT-6 Sol, Grok 4.x). Epoch estimates mostly where developers disclose hardware.

Base rates I computed (from the CSV)

• Record-setting runs since 2020: GPT-3 at 3.1e23 (2020), GPT-4 at 2.1e25 (2023), Gemini Ultra at 5e25, Grok 3 at 3.5e26 (2025), Astra at 1e27 (Sep 2026).
• A log-linear fit to Epoch's "Frontier model"-flagged rows since 2020 gives 4.4×/yr. The trend sits at about 1.06e27 now and crosses 1e30 around April 2031.
• A fit to record-setters since 2023 gives only 3.3×/yr, crossing around 2032.1. From early 2025 to late 2026 the record grew only about 2×/yr.
• A 1,000× rise takes 4.3 years at 5×/yr, 4.7 at 4.4×/yr and 5.8 at 3.3×/yr. The window ends 5.3 years after Astra's release.
• Epoch's own "AI in 2030" report (Sep 2025, https://epoch.ai/files/AI_2030.pdf) says that on trend "the largest models will be trained using 10^29 FLOP by 2030". That puts 1e30 at roughly 2031–2032.

Physical check

1e30 FLOP at 30% utilization over 180 days needs about 2.1e23 peak FLOP/s. Using my estimates of chip efficiency including facility overhead, that is about 21 GW on Rubin-class hardware (FP4), about 9 GW on Feynman-class (FP4) or about 18 GW (FP8), and about 5 GW on a 2030-generation chip. Astra used about 0.25 GW. Today the largest site is Colossus 2 at 946 MW, and the largest planned site in Epoch's tracker is about 2.26 GW in 2028 (https://epoch.ai/graphs/largest-ai-data-centers-by-power-capacity). Epoch projects the largest single site at only 3.7M H100-equivalents by January 2028, although the record has been doubling about every 7 months (https://epoch.ai/data-insights/largest-data-center-compute). Overall buildouts are huge: SpaceX targets about 10 GW by end-2027 (SemiAnalysis, 7 Aug 2026), and OpenAI already passed 10 GW secured by April 2026 (openai.com). So a 1e30 run in 2030–31 is feasible for a top lab that commits about 5–10 GW to one run. It is not feasible before about 2029.

Paths to NO, and frictions

• Safety pauses: OpenAI paused training twice after the Hugging Face incident, and its "largest planned frontier RL run remains on hold" (openai.com/index/pacing-model-development-cyber-capabilities, Aug 2026; Fortune, 26 Sep 2026).
• Pacing: Amodei's "We Must Pace the Frontier" (12 Sep 2026) explicitly floats limits on training compute. The President opposes slowdowns and an antitrust suit challenges coordinated limits, so binding caps are uncertain but possible.
• Opportunity cost: inference is very profitable (SemiAnalysis estimates over $100B per GW per year), which raises the cost of tying up gigawatts for one run.
• Epoch's rules: coverage is patchy, and the question's single-run rule excludes pretraining-plus-RL sums. Model release and estimate also lag behind the end of training.

Model and forecast

I ran a Monte Carlo over growth regimes:

• 7% explosive (0.85 orders of magnitude per year)
• 53% trend (0.55/yr)
• 20% slower (0.42/yr)
• 20% paced (0.25/yr)

Each simulated model scatters ±0.3 orders of magnitude around the trend, with about 4.5 estimated frontier models per year, a 0.1 order-of-magnitude haircut for the single-run rule, and a 0.2-year lag to publication. This gave about 0.7% by end-2029, 6% by end-2030 and 22% by end-2031. I rounded that into hazards: near zero through 2028 (only data-error tails), rising through 2030, and about 4–7% per quarter in 2031.

Strongest case against my number: Epoch's long-run 4.4–5×/yr frontier trend implies a first crossing in late 2030 to 2031, which would mean roughly 50–70% by end-2031, and this world's capital boom and recursive self-improvement push could speed things up. I weight that against the slower recent record growth, the pauses and pacing pressure, and Epoch's own forecast of about 1e29 by 2030. The trend crossing falls right at the window's edge, so small delays push it out of the window.

Question. By the last day of each quarter, has Epoch AI publicly estimated that a single AI training run used at least 10^30 FLOP?

Definitions.

• Training compute: the number of floating-point operations used in training, not FLOP per second, hardware peak capacity, or inference-time compute. Use Epoch AI's FLOP accounting, as described in its AI Models Documentation—Estimation, without precision-based normalization.
• Single training run: one actual training execution producing a model or checkpoint, including distributed execution and checkpoint restarts or continuations of that execution. Distinct pretraining, fine-tuning, or post-training runs cannot be added together to reach the threshold. A run need not produce a publicly released model, and may be by any developer.
• Epoch AI's estimate: Epoch's published central or point estimate for the run, or an explicit lower bound of at least 10^30 FLOP. A point estimate qualifies regardless of its confidence label, including “Speculative”; a confidence interval merely extending above the threshold does not. Where only a range is given and no central estimate is designated, its lower endpoint must reach the threshold.

Resolves YES if a qualifying Epoch AI publication identifies an actual run and estimates compute already expended in that run at or above 10^30 FLOP. The dataset field Training compute (FLOP) is the primary measure, subject to the single-run requirement above. A completed run qualifies, as does a measured or estimated elapsed portion of an ongoing run; a planned total does not. Apply the Observation period convention: the training period covered by the estimate must end after the opening date.

Exclusions (resolve NO).

• Projected future runs, announced compute budgets, or calculations of what a cluster could train.
• Totals across experiments, hyperparameter searches, independent seeds, failed runs, model families, or independently trained ensemble members.
• A dataset total that reaches the threshold only by including separately trained base models or components, or by adding distinct fine-tuning or post-training runs. Epoch's documentation permits such cumulative totals, so a row alone is insufficient when its notes identify this aggregation; at least one constituent run must independently meet the threshold according to Epoch.
• Developer claims or third-party estimates that Epoch has not adopted, and rounded chart labels whose underlying value is below 10^30 FLOP.
• No qualifying publication by the quarter-end.

Resolving source. Epoch AI only: its Data on AI Models database, the official all_ai_models.csv download and associated record notes, or an Epoch-authored report or data insight explicitly estimating an identified run. Official successor datasets and fields count if they retain the same measure. This is a narrower source requirement than Public confirmation (A2); separate developer confirmation is not required. Apply the Resolution window and Resolution date (A3) conventions. For a dataset update, use the first publicly verifiable date the qualifying estimate appeared, established by a dated official update or archived snapshot, not the model's publication date or an unrelated row modification. Revisions can first establish YES only for a training period meeting the Observation period convention; once YES, later downward revisions do not reverse it.