A training run confirmed at 10^29 FLOP

Where things stand (checked 12 Oct 2026)

I downloaded Epoch's all_ai_models.csv (https://epoch.ai/data/all_ai_models.csv; the page says it was updated 8 Oct 2026). The largest training-compute estimate in it is GPT-6 Astra: 1.0e27 FLOP ("Likely"), with a publication date of 3 Sep 2026. Epoch's notes base this on at least 100,000 GB200 chips, about 90 days of training and 25% hardware utilisation at 8-bit precision, with a range of roughly 5e26 to 2e27. Astra came out before the question opened, and its estimate is 100 times too small anyway. No qualifying estimate exists yet.

Epoch's coverage of frontier models is sparse. It has no compute estimate for the Claude 5.x models, Gemini 3 or 4, GPT-5.5, the Grok 4.x models or Meta's Muse models. It only estimates when a developer discloses hardware or token counts. In practice, the first estimate of 1e29 or more probably has to come from OpenAI, SpaceXAI or Meta, the labs that have disclosed such details.

Base rates I computed from the CSV

• Record progression: GPT-3 3.1e23 (2020) → GPT-4 2.1e25 (Mar 2023) → Gemini 1.0 Ultra 5e25 (Dec 2023) → Grok 3 3.5e26 / GPT-4.5 3.8e26 (Feb 2025) → Grok 4 5e26 (Jul 2025) → Astra 1e27 (Sep 2026).
• Growth of the record: about 3.6x a year since GPT-3 and about 3.05x a year since GPT-4. Over the last 14 months it was only about 2x.
• Regression on Epoch-flagged frontier models: about 4.4x a year since 2020. The fitted line crosses 1e29 around 2029.5–2029.9.
• Reference class: every five-year window in Epoch's record history grew at least about 460x. This question needs only 100x in 5.3 years (about 2.4x a year).

Rates of 3x, 4x and 5x a year from Sep 2026 reach 1e29 around late 2030, around 2030.0 and mid-2029 respectively.

Bottom-up check

Under Epoch's style of assumptions (25% utilisation, 8-bit precision, about 120 days), 1e29 needs roughly 5–6 GW of Rubin Ultra-class chips, or about 3 GW of next-generation (Feynman) chips, in a single run.

• Single sites: Epoch's data-centre tracker projects the largest single site at about 2.3 GW by May 2028 (https://epoch.ai/data-insights/frontier-data-center-power).
• Whole fleets: OpenAI expects "low double-digit GW" in 2027 (https://epoch.ai/gradient-updates/frontier-labs-dont-use-most-ai-compute).
• Epoch's own projections: clusters able to run about 1e29 by 2030, and 4–16 GW for the largest runs in 2030 (https://epoch.ai/publications/what-will-ai-look-like-in-2030; https://epoch.ai/publications/power-demands-of-frontier-ai-training).

Astra used only about 8% of OpenAI's fleet. If OpenAI's fleet grows about 2.5x a year, and labs devote somewhat more of it to one run and train longer or at lower precision, the first 1e29-class run lands around late 2029–2030. Release and Epoch's estimate would follow some months later.

What could slow it

• Pacing and pauses. OpenAI put its largest planned frontier RL run on hold (https://openai.com/index/pacing-model-development-cyber-capabilities/, 18 Aug 2026). Amodei's "We Must Pace the Frontier" essay (12 Sep 2026) was endorsed by Altman and Musk. The State of AI 2026 report says OpenAI's most capable models were still paused on 25 Sep.
• Spending. Analysts expect AI capital spending growth to slow in 2027–28 (Fortune, 7 Oct 2026).
• Shift away from pretraining. By SemiAnalysis's estimate, pretraining is a shrinking share of lab compute.
• Epoch's estimates. They tend to be conservative (Astra was assumed to train for 90 days, though it may have run about 6 months), and Epoch covers few labs.
• Counting rule. Pretraining plus post-training sums do not count. I assumed only a single-run estimate, such as one based on the cluster used, qualifies.

What could speed it

• Exploding revenue: Anthropic is on a pace above $100B a year and OpenAI near $70B.
• Hyperscaler capital spending of about $1.2T in 2027 (Goldman).
• SpaceXAI is heading for 1.44M GPUs at Colossus 2 by the end of 2026, and Musk discloses GPU counts.
• Rubin-generation chips, training at 4-bit precision, training across several sites, and AI automating AI research.

Forecast

I put the median date for Epoch's first estimate of 1e29 or more at about 2031.5–2032, with fat tails. That gives roughly 1% by end-2027, 4% by end-2028, 13% by end-2029, 28% by end-2030 and about 50% by end-2031. The quarterly chance rises from about 0.15% now to about 10% in late 2031.

Strongest case against this forecast: the historical rule that every five-year window grew by more than 100x points to more like 65–70%. I discounted it because the record's growth rate has fallen from about 4.5x to about 2–3x a year, and because Epoch's sparse, conservative estimates add delay.

Question. By the last day of each quarter, has Epoch AI publicly estimated that a single AI training run used at least 10^29 FLOP?

Definitions.

• Training compute: floating-point operations actually used for training, in Epoch AI's accounting, rather than FLOP/s, available hardware capacity, dollar cost, or algorithmically adjusted “effective compute.” Use the numerical central estimate in Epoch AI's Training compute (FLOP) field or an explicit Epoch AI estimate for the run. Exactly 10^29 FLOP qualifies. All Epoch confidence labels, including “Speculative,” are eligible; no lower-confidence-bound requirement applies. If Epoch gives only a range or bound and no central estimate, its lower endpoint must be at least 10^29 FLOP.
• Single training run: one coordinated training execution producing a model or checkpoint, including distributed execution, checkpoint resumptions, and stages within that execution. Do not sum independent runs, model-family members, hyperparameter searches, discarded restarts, or separately pretrained component models. A dataset total that includes such compute qualifies only if Epoch separately attributes at least 10^29 FLOP to one run. Separate pretraining and post-training runs cannot be added together, but either can qualify on its own if Epoch publishes a run-specific estimate. Inference, evaluation, and separate synthetic-data-generation jobs do not count.
• Scope: models from any developer, including unreleased models, are eligible. A completed run or a documented amount already consumed by an ongoing run can qualify; planned total compute cannot.

Resolves YES if Epoch AI publishes a qualifying run-specific estimate on or before the quarter-end, within the Resolution window. The training interval to which the estimate applies must end after the opening date, under the Observation period convention; for an ongoing run, use the date through which consumed compute is estimated. A revised estimate can qualify if that interval satisfies this condition. An estimate merely revising a run completed before the opening date does not qualify.

Resolves NO if no qualifying estimate has been published by the quarter-end. Exclude forecasts, announced training budgets, estimates only of installed compute capacity, an upper uncertainty bound above the threshold when the central estimate is below it, and aggregate training totals not attributable to one run. Developer announcements, press reports, and estimates by other analysts do not qualify unless Epoch AI adopts the numerical estimate in its own publication or dataset.

Resolving source. Epoch AI only: its Data on AI Models database, downloadable all_ai_models.csv, accompanying run notes, or an Epoch-authored research publication on epoch.ai. The relevant field is Training compute (FLOP), interpreted with its Training compute notes, Base model, Finetune compute (FLOP), and Post-training compute (FLOP) fields where applicable. Methodology is documented in AI Models Documentation – Estimation. Direct successor datasets and renamed fields count only if they still report actual training FLOP for an identifiable run. This is a narrower resolving-source requirement than Public confirmation (A2): an Epoch estimate itself suffices, without separate developer confirmation. Apply Resolution date (A3) to the first public appearance of the qualifying estimate, not the model's release date or training date. For dataset changes, use dated public versions or archived snapshots; a model's Publication date or an unrelated Last modified change does not establish when the estimate appeared. Apply the Resolution window and its once-YES rule; later downward revisions do not reverse a qualifying YES.