Is DeepSeek’s Low‑Cost AI Model a Real Disruptor or Just Hype?

According to a February 10 report by the Financial Times, Arm CEO Rene Haas said the rapid rise of Chinese AI startup DeepSeek was unexpected, but dismissed the claim that its training cost was only US$5.6 million as a rumor.

DeepSeek’s technical claims

Since the end of 2023, DeepSeek released two open‑source models, V3 and R1, that challenge the conventional belief that high‑performance AI requires massive compute budgets. The R1 inference model is said to match OpenAI’s o1 performance while costing roughly 1/20 of the training expense (V3 was trained on 2 048 H800 GPUs for two months at about US$5.6 million). Its API pricing is claimed to be about 1/28 of o1, reducing usage costs by roughly 97%.

Industry reaction

Venture capitalist Marc Andreessen called DeepSeek an “artificial‑satellite moment” for AI. Haas, however, remained skeptical, noting the speed of development and questioning the veracity of the low‑cost narrative.

Haas acknowledged that open‑source models are beginning to catch up with the best closed‑source inference tools and highlighted the significance of a Chinese‑origin model breaking the Silicon‑Valley‑led progress narrative.

Investment trends

Despite DeepSeek’s buzz, major cloud providers such as Microsoft and Meta reiterated in their January 29 earnings calls that AI‑related capital expenditures will continue to rise in 2025, with Meta planning $60‑65 billion and Microsoft projecting $80 billion for AI data‑center construction.

Strategic partnerships

On February 3, SoftBank (Arm’s parent) and OpenAI announced a joint venture in Tokyo, SB OpenAI Japan, with SoftBank committing $3 billion annually for OpenAI technology. Earlier, SoftBank, OpenAI, and Oracle had launched the “Stargate” partnership, pledging $500 billion over four years, with an immediate $100 billion investment in AI infrastructure.

Government restrictions

U.S. agencies have moved to block DeepSeek: NASA prohibited its use on government devices on January 31, citing foreign‑hosted servers and security concerns; the Department of Defense and Navy issued similar bans at the end of January. A bipartisan bill introduced by Senator Josh Hawley would criminalize the download or use of DeepSeek on U.S. government equipment, with penalties up to 20 years imprisonment and $1 billion fines for corporations.

Other governments—including South Korea, Italy, Australia, and India—have announced or are planning comparable restrictions.

Broader perspective

Haas compared the U.S. stance on DeepSeek to its approach to TikTok, suggesting a similar precautionary attitude. He emphasized that his comments were personal opinions without firm knowledge.

China’s foreign ministry spokesperson countered the politicization of technology, affirming legal protection of data privacy and defending Chinese enterprises’ rights.

Open‑source impact

Because DeepSeek’s models are open source, complete bans may have limited practical effect. Several U.S. tech giants—Amazon, NVIDIA, Intel, AMD—have begun supporting the models, and parts of the European tech community view DeepSeek as a means to break the monopoly of a few AI vendors and promote AI democratization.