What AI Labs Actually Disclose About Training Costs

Every argument about what an AI model costs to train ends the same way: someone quotes a number, and nobody can say where it came from. So we collected only the figures labs have published themselves, in their own papers, model cards and release notes, and recorded for each what it covers and what the lab says it leaves out.

The result is 21 rows from 12 organisations between May 2022 and September 2026, plus two academic fine-tunes kept apart. Six rows give a dollar amount. The rest are GPU-hours, carbon, water or days-times-chips, and they cannot be added together. No dollar figure has been checked by anyone outside the lab. The four labs most people mean when they say “frontier” publish nothing for their frontier models.

The newest row, Xiaomi’s September 22 disclosure of its MiMo-V2.6 reinforcement-learning run, is also the only 2026 figure we found that prices post-training. Our post on that release covers it in full; here it takes its place in the table.

One row per disclosure. The figure column keeps the lab’s own unit and, where a phrase is quoted, its own words. The last column says what the figure covers, because that is the part every secondary citation drops.

Seven more model reports give hardware or relative compute but no absolute cost, hours or energy figure, and are not counted above: NVIDIA’s Nemotron-4 (768 DGX H100 nodes), Apple’s AFM-server (8,192 TPUv4 chips), Gemma 3 (chip counts per model), Kimi K2 (an H800 cluster), Kimi K3 (relative RL-compute plots only), Qwen3 (a relative one-tenth claim for distillation) and GLM-4.5 (nothing found).

The table has four families of unit, and a reader who wants one ranking has to convert between them with assumptions the labs did not make. We have not converted anything except chips times days, and we have not ranked the rows.

The labs are more candid about this than their quoters. Read the exclusions in their own words first; the pattern follows.

1. The research before the run. Ablations, baselines and abandoned experiments. Meta’s OPT footnote doubles the total; BigScience’s share puts the final run at just over a third; Ai2 measured development at half of training again.
2. The base model under a post-training figure. DeepSeek-R1’s $294K sits on DeepSeek-V3’s $5.576M. MiniMax-M1’s $534,700 follows 7.5 trillion tokens of continued pre-training. Xiaomi’s RL cost sits on 30 to 48 trillion tokens of pre-training it did not price.
3. People, capital and data. Five of the six dollar rows are rental-priced GPU time, and Xiaomi does not say how it priced its run. Epoch AI’s estimate for GPT-4 puts the hardware at about $800 million to acquire against $40 million amortised for the run; staff at tens of millions. No lab row says it includes any of the three.
4. Post-training, evaluation and fine-tuning. The Llama cards cover pre-training. Ai2 notes that repeated checkpoint evaluations and post-training consume a non-trivial share that pre-training hours alone do not capture.
5. Inference and hardware life cycle. Only Mistral’s row includes use after training and the emissions of making the chips, and it is a different kind of number: 18 months of a deployed model, not a training run.

Pre-training compute is the figure labs used to disclose, when they disclosed anything. The rows that matter for a 2026 buyer are the ones that price the reinforcement-learning phase, because that is where this year’s agent capability was added. There are four.

• DeepSeek-R1: $294K, or 147K H800 GPU-hours, published with the Nature paper in September 2025. That is about 5% of the $5.576M DeepSeek stated for the V3 base, our arithmetic on two figures priced at the same $2 rate. Epoch AI had estimated the RL cost at around $1 million eight months earlier.
• MiniMax-M1: $534,700 for three weeks on 512 H800s, June 2025. RL only.
• NVIDIA Nemotron Ultra 253B: about 140K H100 hours for the reasoning-RL stage, May 2025. No dollar figure.
• Xiaomi MiMo-V2.6: about $850K for Flash and $2.62M for Pro, six days, 30 steps, September 2026. Xiaomi ties the spend to a result: on its own harness, DeepSWE v1.1 moved from 48.8 to 65.7 for Flash and from 58.4 to 72.6 for Pro over those 30 steps.

Xiaomi’s report plots its benchmark score against cumulative dollars spent, which is why it is a more useful disclosure than the others even though it is the least specified on hardware. It is also the largest, and Xiaomi’s report says the run used thousands of GPUs. Kimi K3’s August 2026 report shows RL compute only as a relative plot, so the comparison a reader wants most, Xiaomi against Moonshot, is not possible from primaries. Our post on RL as the new moat covers why labs now guard this number.

Where labs are silent, two estimators are quoted most. They disagree with each other by about 2× on GPT-4 and about 6× on Gemini Ultra, because they use different methods, and a citation should say which.

• Epoch AI (Cottier et al., May 2024, revised February 2025) amortises hardware plus energy for the final run: GPT-4 about $40 million, Gemini Ultra about $30 million. The same paper puts GPT-4’s hardware at about $800 million to acquire and says frontier run costs have grown at 2.4× a year since 2016.
• Stanford AI Index (2024 and 2025, using Epoch data at cloud-rental prices) gives GPT-4 about $78 to $79 million, Gemini Ultra $191 million and Llama 3.1-405B about $170 million. Meta’s own card for that model gives 39.3M GPU-hours and no dollar figure. The 2025 report adds that labs are disclosing less about their training processes, which makes the estimates harder.

Two numbers in wide circulation are not lab figures. The “$6 million DeepSeek model” rounds $5.576M and drops DeepSeek’s own exclusion of prior research; our post on DeepSeek’s first funding round covers the round it was reported to be raising. The “$50 reasoning model” for s1 does not appear in the s1 paper, which states 26 minutes on 16 H100s and nothing in dollars.