Build a top AI model for $5.6M. DeepSeek did it. GPT-4 cost about $78M.
DeepSeek-V3's final training run cost $5.6M in rented GPU time, against an estimated $78M for GPT-4.
We didn’t run this. DeepSeek-V3 Technical Report (arXiv 2412.19437) did. ModelSignal read the published work and checked every number below against it — the sources and the dates we read them are at the end.
DeepSeek-V3 · the bill
$5.6M
to train a top AI model.
Their own accounting.
2.788M GPU hours at $2 an hour. Dec 2024.
DeepSeek vs GPT-4
OpenAI's GPT-4
~$78M
Estimated compute
Gemini Ultra, est.~$191M
DeepSeek-V3
$5.6M
Final run, by them
Not countedresearch
Count the hours.
$5.6M covers one final run at a rental price they chose. Failed runs aren't in it.
DeepSeek report · Stanford AI Index 2024
The proof
DeepSeek-V3's report on arXiv, Dec 2024. Read 13 Oct 2026.
The full breakdown
Somebody else paid this bill; we read the report. Every number below is quoted from the paper and the Stanford AI Index, with the date we read them.
The report
DeepSeek-V3 Technical Report, arXiv:2412.19437, v1 dated 27 December 2024. Model: 671B total parameters, 37B activated per token, pre-trained on 14.8 trillion tokens.
Table 1 — training costs of the final run
| Stage | H800 GPU hours | At $2/hour |
|---|---|---|
| Pre-training | 2,664K | $5.328M |
| Context extension | 119K | $0.238M |
| Post-training | 5K | $0.01M |
| Total | 2,788K | $5.576M |
"Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M." The costs "exclude the costs associated with prior research and ablation experiments on architectures, algorithms, or data."
For scale
Stanford AI Index 2024: "Google's Gemini Ultra cost an estimated $191 million worth of compute to train, while OpenAI's GPT-4 cost an estimated $78 million."
Sources
The catches
- DeepSeek's own accounting of the final run, at a rental price it assumed; no outside audit.
- It excludes research, ablations, staff and the cluster; the GPT-4 and Gemini Ultra figures are outside estimates made by a different method, so the comparison is indicative, not like-for-like.
- Hours are H800 GPU hours; a different chip changes both the hours and the price.
Every number, and where it’s quoted from
- 2.788M H800 GPU hours in total; $2 per GPU hour assumed; $5.576M total“DeepSeek-V3 costs only 2.788M GPU hours for its full training. Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M.”arXiv 2412.19437v1, introduction
- Table 1: pre-training 2,664K (2664K) hours $5.328M; context extension 119K hours $0.238M; post-training 5K hours $0.01M; total 2,788K (2788K) hours $5.576M“Pre-Training 2664K $5.328M · Context Extension 119K $0.238M · Post-Training 5K $0.01M · Total 2788K $5.576M”arXiv 2412.19437v1, Table 1
- the figure excludes prior research and ablation experiments“exclude the costs associated with prior research and ablation experiments on architectures, algorithms, or data”arXiv 2412.19437v1, introduction
- 671B total parameters, 37B activated per token“671B total parameters with 37B activated for each token”arXiv 2412.19437 abstract
- 14.8T pre-training tokens“14.8 trillion diverse and high-quality tokens”arXiv 2412.19437 abstract
- $5.6M is $5.576M rounded to one decimal“$5.576M”arXiv 2412.19437v1
- GPT-4 ≈$78M and Gemini Ultra ≈$191M estimated training compute“Google's Gemini Ultra cost an estimated $191 million worth of compute to train, while OpenAI's GPT-4 cost an estimated $78 million.”Stanford HAI, AI Index 2024 (15 Apr 2024)
- arXiv identifier 2412.19437, version 1 dated 27 December 2024“arXiv:2412.19437v1 — 27 Dec 2024”arXiv
