Has frontier AI training-compute growth actually slowed?
The verdict
No, not at the ceiling: as of late May 2026 frontier training compute is still climbing at roughly 4 to 5x per year and the biggest run on record keeps getting bigger, but the curve is bending below it as per-flagship total training compute flattens (and the slowdown likely sits in pretraining as labs reroute spend into reinforcement learning).
My ruling, as of 29 May 2026: no, frontier training compute has not slowed at the ceiling. The curve is bending underneath it, and that distinction is the whole story.
Start with the top line. Epoch AI fits frontier training compute at 4 to 5x per year, with the running top-10 at 5.3x annually (90% CI 4.9x to 5.7x) over 2010 to May 2024. Worth saying plainly: the same analysis already clocks a post-2018 slowdown to 4.2x/year (90% CI 3.6x to 4.9x), so deceleration is not a fringe claim. Yet the ceiling keeps rising. Epoch's trends dashboard, updated February 5, names the largest known run as Grok 4 at roughly 5e26 FLOP. Epoch's separate Grok 4 analysis puts that run at about 246 million H100-hours. That is roughly 25x GPT-4's ~2e25 FLOP. The absolute frontier marched on.
Now the catch. GPT-5 landed at about 5e25 FLOP total. More than double GPT-4, but below GPT-4.5's >1e26 FLOP. Read that twice. A flagship used less total compute than the prior flagship. Note that this is total training compute including RL; pretraining-only is indeterminate from the public figures, so the read that the slowdown sits in pretraining is an inference, not a measured pretraining number. Epoch points to reinforcement learning as an increasingly important axis, though it hedges hard: RL's share could be anywhere from 10% to 200% of pretraining.
So the deceleration is real, just not where the wall crowd looks. Per-model total training compute is flattening while the record run climbs. This bend rests on a single recent flagship comparison (GPT-5 below GPT-4.5) plus Epoch's pre-2024 post-2018 deceleration, not a re-fit of 2025 to 2026 data, which is why I cap confidence at medium.
Caveats. Capex still rips at ~72% per year, and the AI chip stock compounds at a steady ~3.3x/year. FLOP figures for closed models are educated estimates, not invoices.
Bottom line: the frontier is not slowing, the per-model curve is, and spend is rerouting into RL. Anyone calling a flat scaling wall is reading one number and missing three.
The evidence
The data points behind this verdict. Each is cited so you can check the call against its source.
Frontier AI model training compute grows 4-5x per year; the running top-10 models grew 5.3x/year (90% CI 4.9x to 5.7x) over 2010 to May 2024, with a post-2018 deceleration to 4.2x/year (90% CI 3.6x to 4.9x). GPT-4 is estimated at ~2e25 FLOP.
5.3x/year full-window (90% CI 4.9x to 5.7x); 4.2x/year post-2018 (90% CI 3.6x to 4.9x); GPT-4 ~2e25 FLOP
Epoch AI, Training compute of frontier AI models grows by 4-5x per yearThe largest known training run is Grok 4 at around 5e26 FLOP; frontier LM compute has grown 5x/year since 2020, doubling roughly every 5.2 months (dashboard updated Feb 5 2026). At ~5e26 vs GPT-4's ~2e25, the ceiling is about 25x above GPT-4.
~5e26 FLOP (Grok 4), ~25x GPT-4; 5x/year since 2020; ~5.2-month doubling
Epoch AI, Trends in Artificial IntelligenceEpoch's Grok 4 analysis (dated Sept 12 2025) estimates ~246 million H100-hours; its subtitle frames Grok 4 as 'the largest AI training run to date,' while the article body quantifies resources without repeating the superlative.
~246M H100-hours; 'largest to date' is Epoch's subtitle framing, as of Sept 12 2025
Epoch AI, What did it take to train Grok 4?GPT-5 total training compute estimated at ~5e25 FLOP (pretraining plus RL): more than 2x GPT-4's ~2e25 FLOP but LESS than GPT-4.5's >1e26 FLOP. Epoch hedges on RL, calling it 'a key factor in near-term AI progress' with its share estimated at 10% to 200% of pretraining.
~5e25 FLOP total, below GPT-4.5's >1e26; RL share 10-200% of pretraining
Epoch AI, Notes on GPT-5 training computeHyperscaler capex has grown ~72%/year (90% CI 66% to 78%) since Q2 2023, projecting ~$770B in 2026.
~72%/year capex; ~$770B 2026 projection
Epoch AI, Hyperscaler capex has quadrupled since GPT-4's releaseThe total stock of AI chips (global AI computing capacity) has grown by approximately 3.3x per year since 2022, doubling roughly every 7 months.
~3.3x/year, ~7-month doubling, since 2022
Epoch AI, Global AI computing capacity is doubling every 7 monthsCaveats
FLOP figures for closed frontier models (GPT-5, Grok 4, Gemini, GPT-4.5) are Epoch AI estimates derived from public statements and modeling, not disclosed by the labs, so they carry wide uncertainty bands. The headline 4-5x/year fit (and the post-2018 4.2x/year figure) is anchored to data through May 2024; the 2025 to 2026 points (Grok 4, GPT-5) are individual data points consistent with, but not a re-fit of, that trend, which is why the per-model flattening claim is held at medium confidence. "Compute" here is training FLOP only and deliberately excludes inference and total R&D compute. The GPT-5-below-GPT-4.5 comparison is total training compute including RL; an apples-to-apples pretraining-only comparison is murkier, so the read that the slowdown sits in pretraining is an inference about where it likely lands, not a measured pretraining figure. The ~3.3x/year is the steady growth of the global AI chip stock since 2022, not a model-level rate and not a 2025-specific deceleration. The "largest run to date" label is Epoch's own subtitle framing as of its Sept 12 2025 Grok 4 analysis, not a body-text ranking claim, and the ~246M H100-hours comes from that separate analysis, not from the trends dashboard that carries the ~5e26 FLOP figure.
A TF Verdict is TensorFeed's own analysis over cited public data, not a republished dataset. We take a clear position, show the evidence and the sources, and date-stamp the call because the answer can change. Disagree with a data point? Follow the source link and check it yourself.