In short: China’s DeepSeek threw the multi-billion-dollar AI business into chaos lately with the discharge of its R1 mannequin, which is alleged to compete with OpenAI’s o1 regardless of being skilled on 2,048 Nvidia H800s and at a price of $5.576 million. Nonetheless, a brand new report claims that the true prices incurred by the agency have been $1.6 billion, and that DeepSeek has entry to round 50,000 Hopper GPUs.
The declare that DeepSeek was capable of practice R1 utilizing a fraction of the assets required by huge tech corporations invested in AI wiped a report $600 billion off Nvidia’s share value in someday. If the Chinese language startup to might make a mannequin this highly effective with out spending billions on Staff Inexperienced’s strongest AI GPUs, what would cease everybody else doing it?
However did DeepSeek actually create its Combination-of-Specialists mannequin, which nonetheless tops the Apple App Retailer charts, at such a low price? SemiAnalysis claims that it did not.
The market intelligence agency writes that DeepSeek has entry to round 50,000 Hopper GPUs, together with 10,000 H800s and 10,000 H100. It additionally has orders for a lot of extra China-specific H20s. The GPUs are shared between Excessive-Flyer, the quantitative hedge fund behind DeepSeek, and the startup. They’re distributed throughout a number of geographical places and are used for buying and selling, inference, coaching, and analysis.
SemiAnalysis writes that DeepSeek has invested far more than the claimed $5.5 million determine that despatched the inventory market right into a tailspin – the report states that this pre-training price is a really slender portion of the overall. The corporate’s total funding in servers is round $1.6 billion, with round $944 million spent on working prices. The GPU investments, in the meantime, account for greater than $500 million.
As a reference instance, Anthropic’s Claude 3.5 Sonnet price tens of tens of millions of {dollars} to coach, however the firm nonetheless wanted to boost billions of {dollars} of funding from Google and Amazon.
It is famous that DeepSeek has sourced all its expertise completely from China. That could be a distinction to experiences of different Chinese language tech corporations, corresponding to Huawei, attempting to poach staff from abroad, with Taiwanese workers of TSMC being extremely sought-after targets. DeepSeek allegedly affords salaries of over $1.3 million for promising candidates, far more than competing Chinese language AI companies pay.
DeepSeek additionally has the benefit of principally working its personal datacenters, fairly than having to depend on exterior cloud suppliers. This enables for extra experimentation and innovation throughout its AI product stack. SemiAnalysis writes that it’s the single finest “open weights” lab at this time, beating out Meta’s Llama effort, Mistral, and others.
Masthead: Solen Feyissa
