Nairobi, Kenya – Chinese artificial intelligence startup DeepSeek has revealed that training its R1 model cost just Sh37.9 million ($294,000), a fraction of the hundreds of millions reportedly spent by U.S. rivals. The disclosure, published Wednesday in Nature, has reignited debate about China’s role in the global AI race and raised fresh questions about transparency in model development.
A Low-Cost Challenger in the AI Race
DeepSeek’s announcement marks the first time the Hangzhou-based company has publicly shared cost estimates for its technology. According to the paper, the reasoning-focused R1 model was trained using 512 Nvidia H800 chips over 80 hours—hardware designed specifically for the Chinese market after Washington restricted exports of the more advanced H100 and A100 chips in 2022.
The cost comparison is stark. OpenAI’s Sam Altman noted last year that training foundational AI models had cost his firm “much more” than $100 million, although detailed figures have never been published. DeepSeek’s relatively low expenses, if accurate, underscore how Chinese companies could potentially scale AI innovation more cheaply—shaking global markets when the firm first unveiled its systems in January.
Yet questions remain. U.S. officials previously told Reuters that DeepSeek had access to “large volumes” of restricted H100 chips, though Nvidia insists the firm only used lawfully acquired H800s. In supplementary materials released alongside the Nature article, DeepSeek acknowledged it does own some A100 chips, but said they were only used in preliminary experiments before the R1 was trained.
Distillation Controversy and Use of Foreign Models
Beyond hardware, DeepSeek has also faced scrutiny for its methods. Critics, including a top White House adviser, accused the firm of “distilling” OpenAI’s proprietary models into its own—a practice where one AI system learns from another to cut down on costs and computing power.
DeepSeek defended the technique, arguing that distillation boosts efficiency and lowers barriers to access in an industry dominated by resource-heavy Western firms. In its Nature paper, the company admitted that training data for its V3 model included web pages containing OpenAI-generated answers. However, it maintained this was incidental, not deliberate copying.
The company also acknowledged leveraging Meta’s open-source Llama model for some distilled versions of its technology, a common practice in AI research but one that has fueled concerns about intellectual property boundaries.
A Flashpoint in U.S.–China Tech Rivalry
The disclosure comes at a tense moment in U.S.–China tech relations. Washington has steadily tightened export controls on advanced chips, while Beijing has sought to showcase domestic players like DeepSeek as evidence that it can still compete on the world stage.
For DeepSeek’s founder, Liang Wenfeng, the revelation could serve as both validation and provocation: proof that Chinese labs can build competitive AI models at a fraction of the cost, but also an invitation for heightened scrutiny from global regulators and rivals alike.
As AI accelerates into a critical front of geopolitics and economics, DeepSeek’s $294,000 figure may be remembered less as a technical detail and more as a symbolic marker—showing how innovation, controversy, and rivalry now define the race to shape the future of artificial intelligence.
