In a rare disclosure, Chinese AI company DeepSeek has revealed that it trained its flagship R1 reasoning model for just $294,000, significantly undercutting the reported training budgets of U.S. giants like OpenAI and Anthropic — where costs have run well into the tens or hundreds of millions. The announcement, made in a peer-reviewed article in Nature, has reignited global debate over the economics, transparency, and competitiveness of AI model development.
A fraction of the cost, but not without questions
According to DeepSeek’s paper, the R1 model was trained using 512 Nvidia H800 chips over just 80 hours. These chips are designed specifically for the Chinese market and are less powerful than the H100s restricted by U.S. export bans. The company says it used A100 chips (previous-gen Nvidia GPUs) during the prep phase but switched to the more export-compliant H800s for full training.
DeepSeek’s claims challenge the notion that only companies with vast compute resources can train advanced large language models (LLMs). For comparison, OpenAI’s Sam Altman said in 2023 that training GPT-4 had cost “well over $100 million,” although exact figures remain undisclosed.
However, U.S. officials previously alleged that DeepSeek may have access to prohibited H100 chips, and some industry observers question whether the low training cost figures reflect model performance compromises or opacity in methodology.
A model built on ‘distilled’ foundations?
The Nature paper also quietly acknowledges that training data for DeepSeek’s V3 model includes a “significant number” of OpenAI-generated answers. While the company insists this was incidental (as part of web-crawled data), critics argue this borders on model distillation — a controversial method where one model learns indirectly from another, raising concerns over IP ethics and originality.
In earlier product updates, DeepSeek had openly used Meta’s open-source LLaMA models for fine-tuning. However, the latest admission blurs the lines between open training and shadow-copying proprietary models, an issue that has sparked tensions between Chinese and Western AI companies.
Competitive disruption or hype?
Despite limited public appearances by DeepSeek founder Liang Wenfeng in recent months, the company has continued to generate global attention. Its January 2025 launch of low-cost models triggered a tech stock sell-off as investors feared their potential to undercut Western incumbents.
Now, with the Nature paper offering official validation (albeit with caveats), DeepSeek positions itself as a formidable low-cost contender — especially in markets looking to adopt advanced AI without U.S. dependencies.
Still, some AI experts caution that training costs alone don’t paint a full picture. Factors like inference efficiency, multi-turn consistency, bias control, security, and scaling benchmarks remain crucial — and often, more resource-intensive.
