Back
MLQ.ai | AI for investors
webReports on DeepSeek's disclosure of R1 model training costs ($294K) published in Nature, relevant to AI safety governance debates about compute access, export controls, and the economics of frontier AI development.
Metadata
Importance: 42/100news articlenews
Summary
DeepSeek disclosed in a Nature peer-reviewed paper that training its R1 reasoning model cost $294,000 using 512 Nvidia H800 GPUs, far below comparable U.S. model costs. The figure excludes $6 million spent on the foundational base model. R1 is open-weight and has become the most downloaded model on Hugging Face.
Key Points
- •DeepSeek's R1 training cost $294,000 per Nature disclosure, far below U.S. model training estimates of tens of millions.
- •Training used 512 Nvidia H800 GPUs, chips subject to U.S. export controls since 2023.
- •The $294K excludes $6M spent developing the foundational base model underlying R1.
- •R1 is open-weight with 10.9 million Hugging Face downloads, widely used for math and coding tasks.
- •Disclosure challenges assumptions about necessary investment scale for state-of-the-art AI performance.
1 FactBase fact citing this source
| Entity | Property | Value | As Of |
|---|---|---|---|
| DeepSeek | Description | DeepSeek-R1 final training cost reported as $294K for RL stage; full pipeline including base model ~$5.6M | Jan 2025 |
Cached Content Preview
HTTP 200Fetched Apr 7, 20268 KB
MLQ.ai | AI for investors
-->
Key Points
DeepSeek revealed in Nature that its R1 AI model was trained for $294,000, markedly below typical U.S. model training costs.
The R1 training relied on 512 Nvidia H800 GPUs, chips subject to U.S. export controls.
The $294,000 figure excludes $6 million spent developing the foundational base model for R1.
R1 is available as an open-weight model and has been widely downloaded by the AI community.
The disclosure intensifies scrutiny of AI infrastructure costs and cross-border technology competition.
Advertisement
-->
Chinese AI developer DeepSeek has publicly disclosed that training its R1 large language model cost just $294,000, according to a peer-reviewed article published in Nature. This figure is far below widely cited estimates for comparable AI models produced by U.S. firms.
DeepSeek Reveals Detailed Cost Breakdown in Nature
In materials accompanying a peer-reviewed Nature paper, DeepSeek reported that the total training cost for its R1 AI model was $294,000. This disclosure marks one of the first times a major AI developer has released a line-item cost breakdown for model training in a peer-reviewed venue. The company stated that R1 was trained primarily on 512 Nvidia H800 graphics chips, a model placed under U.S. export controls in 2023. The $294,000 sum covers the reasoning-focused 'R1' model's training, not all prior foundational work by the company.
Foundational Model Costs Excluded
The $294,000 price tag does not include the $6 million DeepSeek reported spending to build the general-purpose large language model (LLM) that R1 is based on. Industry analysts noted that while the R1-specific training was inexpensive by global standards, the total investment in DeepSeek's broader AI development remains significant. Experts have said competing models from U.S. firms can cost tens of millions of dollars just for training.
Open-Weight Model with High Adoption
R1 is offered as an open-weight model, making it available for anyone to download. It has become the most popular open-weight model on the AI community platform Hugging Face, with 10.9 million downloads to date[1]. This availability has allowed broader research and adoption, with DeepSeek positioning R1 for advanced mathematical and coding reasoning tasks.
Implications for AI Development Economics
DeepSeek's cost disclosure is likely to impact ongoing debate about the real economics of training competitive AI models. Many in the industry have cited escalating infrastructure expenses as a barrier to global AI development. By releasing detailed cost figures, DeepSeek challenges common assumptions about the necessary
... (truncated, 8 KB total)Resource ID:
kb-2f0bec9dabc3d06e