Is DeepSeek’s $5.6M Training Cost a Myth? Arm CEO’s Take on the AI Challenger

According to the Financial Times on February 10, Arm CEO Rene Haas said the claim that DeepSeek trained its models for only US$5.6 million is a rumor and warned the company might eventually be shut down.

Since the end of 2024, DeepSeek has released the open‑source V3 and R1 AI models, which challenge the industry’s assumptions about AI compute requirements. DeepSeek‑R1 is said to match the performance of OpenAI’s o1 while costing roughly one‑twentieth of its training budget (V3 used 2,048 H800 GPUs for two months, costing about US$5.6 million). Its API pricing is claimed to be about one‑twenty‑eighth of OpenAI’s, reducing usage costs by roughly 97%.

Venture capitalist Marc Andreessen called DeepSeek an “AI satellite moment,” but Haas expressed skepticism, noting the rapid pace of development and suggesting the story may change quickly.

Haas acknowledged the surprise of an open‑source model from China catching up to top‑tier closed‑source tools like ChatGPT, emphasizing that the Chinese origin is a significant development in a field traditionally led by Silicon Valley.

He found the low‑cost training claim impressive but labeled it a rumor, arguing that the market’s reaction to DeepSeek is overblown.

Despite DeepSeek’s impact on AI‑related stock prices, major tech firms such as Microsoft and Meta reiterated in their January 29 earnings calls that AI capital expenditures will continue unabated through 2025, with Meta planning $60‑65 billion and Microsoft $80 billion for AI infrastructure.

On February 3, SoftBank’s CEO Masayoshi Son announced a joint venture with OpenAI in Tokyo, SB OpenAI Japan, committing $3 billion annually to use OpenAI technology, alongside other large‑scale AI investments.

The United States has taken a hard line: NASA, the Department of Defense, and the Navy have banned DeepSeek on government equipment, citing security and privacy risks. A bipartisan bill introduced by Senator Josh Hawley would criminalize the use of DeepSeek on U.S. government systems, with penalties up to 20 years in prison and $1 billion for corporations.

Other countries—including South Korea, Italy, Australia, and India—are also considering or have implemented restrictions on DeepSeek.

Haas compared the U.S. government’s approach to DeepSeek with its treatment of TikTok, stating it is his personal view and not based on any insider knowledge.

China’s foreign ministry spokesperson Guo Jiakun responded, opposing the politicization of technology issues and affirming China’s commitment to data privacy, security, and the protection of legitimate corporate rights.

Because DeepSeek’s models are open source, the practical impact of outright bans may be limited. Companies such as Amazon, NVIDIA, Intel, and AMD have begun supporting DeepSeek, and the European tech community welcomes its potential to break the monopoly of a few AI vendors and promote AI democratization.