Capability
Frontier Model Training Compute
Record-high estimated training compute among Epoch AI frontier models since 2023, shown in zettaFLOP (10²¹ FLOP) on a log scale. It tracks leading runs' resource scale—relevant to accelerator, data-center and capital needs—but not capability or cost.
Frontier Model Training Compute
Epoch AI's frontier-model file records a training-compute high of 940,000 zettaFLOP as of 2026-09-03, up from 21,000 zettaFLOP for GPT-4 in March 2023.
Methodology
Source cohort: use Epoch AI's official Frontier AI Models CSV, re-read on every scheduled check. Epoch defines this subset as models that were in the top 10 by training compute at the time of release.
Metric: Epoch defines Training compute (FLOP) as total compute used to train a model, including pretraining and fine-tuning. Values may be reported directly or estimated from operation counts, hardware and training time, cost, benchmarks, or comparisons with other models. For Sanity-compatible storage and readable labels, chart values divide source FLOP by 10²¹ and are displayed as zettaFLOP. Model names, Epoch confidence labels, estimation methods, and the most decision-relevant caveats are retained in point notes.
Frontier construction: parse valid positive Training compute (FLOP) values, sort the full source history by Publication date, and retain a model only when its compute strictly exceeds every earlier source model. Display only record events dated from 2023-01-01 through the check date. Computing records over the full history before clipping prevents a merely local post-2023 maximum from being mislabeled as an all-time record.
Line-only representation: use one unstacked step-after line for the running record rather than a separate connected line for every model. Each record event is a different named model, but connecting successive record states has a defined meaning: the observed record changed at that release. One-second hold vertices immediately before each new record encode discrete steps in a line-only renderer; the CSV does not invent monthly observations.
Scale and interpretation: a logarithmic y-axis is required because the observations span more than an order of magnitude. One zettaFLOP equals 10²¹ FLOP. Compute is a count of operations, not training cost, elapsed time, hardware quantity, energy use, or model capability.
Coverage limit: the line ends at the latest dated record event in Epoch's file and is never carried forward to the check date. It does not claim that no undisclosed or unrecorded model used more compute. When Epoch leaves a model's estimation method blank, the point note says the method is not stated.
Frequently asked questions
What makes a model a frontier model here?
Epoch AI defines a frontier model as one that was in the top 10 by training compute at the time of release. This chart applies a stricter display rule within that source cohort: it plots only models that set a new all-time training-compute record.
Why does frontier training compute matter to the AI industry?
Training compute indicates the computational scale committed to developing a model. Higher record estimates provide context for potential accelerator demand, data-center requirements and barriers to reproducing leading runs. The chart converts source FLOP to zettaFLOP, but compute still does not reveal spending, energy use, training efficiency or capability.
Why not draw one line through every frontier model?
Successive models are unrelated observations, so connecting all of them would imply a continuous model-level trajectory. A single stepwise record frontier has a defensible time-series meaning: the highest observed training-compute estimate changes only when a new record model appears.
Are these training-compute values reported facts?
Not generally. The displayed values are Epoch AI estimates with confidence labels such as Likely or Speculative. The point notes preserve the model name, confidence label, estimation method and key caveats; the values can change when Epoch revises evidence or assumptions.
How should an increase in the record be interpreted?
It means a newly released model has a higher Epoch training-compute estimate than every earlier model in the source history. It does not establish that the model is more capable, less efficient or more expensive, because architecture, data, algorithms, utilization and hardware economics also differ.
Why use a logarithmic y-axis?
The record estimates span well over an order of magnitude, from about 21,000 zettaFLOP for GPT-4 in March 2023 to the latest record. A log scale makes proportional changes readable without compressing the earlier records near zero.
Does the final point show the training-compute frontier today?
No. It is the latest dated record event in Epoch's file, not a value carried forward to today. The chart makes no completeness claim for releases Epoch has not yet recorded or estimated.