Why DeepSeek's Rise Could Be Temporary
Chinese startup DeepSeek claims to have produced AI models that perform as well as ones developed in the U.S. at substantially lower cost. The news has rocked markets worldwide, based on the view that more efficient designs could displace the need for costly chips from NVIDIA and obviate the need for high-end datacenters to develop AI models by OpenAI and the hyperscalers.
AI tech stocks affected by the news tumbled in trading today. NVIDIA shares were down over 15% by midday. The news dragged down shares of Microsoft, Meta, and Alphabet.
Still, the rout could be temporary, though it may affect AI pricing worldwide.
China’s Buzzing Triumphantly
The buzz about DeepSeek seems to have been timed perfectly to coincide with the Chinese New Year and last week’s announcements in the U.S., specifically the introduction of the $500-billion Stargate venture announced by President Donald Trump and word that Meta is set to spend up to $65 billion on AI infrastructure this year. (This follows Microsoft’s claim to be spending up to $80 billion in 2025 on AI.)
An article in the China Daily on Monday, January 27th, trumpeted the news:
“In benchmark tests on Friday [January 25], DeepSeek-R1 ranked third across all categories on the international large model leaderboard, Arena. In the style control model (StyleCtrl) category, it tied with OpenAI's o1 for first place, with an Arena score of 1,357, slightly surpassing o1's score of 1,352.
“In addition to matching o1 in performance, the release of DeepSeek-R1 sent shockwaves through the US tech industry because it is fully open-source and achieved this breakthrough at an exceptionally low cost.”
Let’s take a closer look.
Technical Breakthroughs Lead to Lower Costs
The technical specifics of DeepSeek’s methods are esoteric, but the gist appears to be that by using a clever combination of well-known approaches, including Mixture of Experts (MoE) and reinforcement learning (RL), along with innovations such as Multi-Head Latent Attention (MLA), the startup has been able to craft models that consume less resources than are typically required by the likes of OpenAI for model development. By most accounts, DeepSeek's latest model cost under $6 million to develop, compared with billions claimed by leading AI firms in the West.
There is talk that costs in China may be cheaper than in Western countries, contributing to the figures. According to Reddit posters, cheaper electricity and more efficient supply chains in China may have contributed to the startup’s claims.
At any rate, cheaper production costs are reflected in substantially lower pricing by DeepSeek for its wares. How much lower depends on the model, but according to some sources, the difference can be 20 to 40 times cheaper for DeepSeek’s models than for OpenAI’s. Indeed, in May 2024, DeepSeek reportedly launched a price war in China based on the low cost of its DeepSeek-V2 model.
Questionable Infrastructure
A selloff of NVIDIA shares followed the DeepSeek news because apparently the startup used less GPU power than rivals such as OpenAI and Anthropic claim to need. But reports on DeepSeek’s platform are confusing. According to a DeepSeek technical report posted on research hub arXiv and cited by research analyst Janakiram MSV on Forbes, DeepSeek used NVIDIA chips to train its DeepSeek-V3 model:
“DeepSeek-V3 is trained on a cluster equipped with 2048 NVIDIA H800 GPUs. Each node in the H800 cluster contains 8 GPUs connected by NVLink and NVSwitch within nodes. Across different nodes, InfiniBand (IB) interconnects are utilized to facilitate communications.”
Another detailed article cited by Janakiram MSV appears on AMD’s site touting: “AMD Instinct GPUs Power DeepSeek-V3: Revolutionizing AI Development with SGLang.” A note on GitHub seems to say that DeepSeek-V3 runs on both NVIDIA and AMD infrastructure.
There’s more: According to at least one report, DeepSeek founder Liang Wenfeng also acquired an estimated 50,000 NVIDIA A100 chips before they were banned by the U.S. government for sale in China. Using these and other, smaller chips still allowed in China may have given DeepSeek the horsepower needed to create its models.
What’s Next?
While DeepSeek roils markets, it also raises questions. Its innovations appear to be impressive and real, but its development costs seem confusing. Still, experts, including Marc Andreessen, have praised the startup warmly: “Deepseek R1 is AI’s Sputnik moment,” Andreessen wrote on X. That Russian rocket launch impelled the U.S. to advance its space mission in the late 1950s and 1960s, ultimately resulting in the 1969 Apollo 11 moon launch.
Another caveat has to do with Western communications with China. In the era of Salt Typhoon, claims and strategies are questionable. Liang himself, who started DeepSeek in 2023 with backing from a hedge fund he founded, recently appeared at a meeting hosted by PRC Premier Li Qiang in Beijing. That event signals his close ties with the Chinese government.
At the same time, the competency of NVIDIA, AMD, and other U.S. companies is established, and by nearly all accounts leading U.S. and European tech firms are well positioned to regroup and innovate as necessary to meet market demand. If this is a Sputnik moment, expect DeepSeek competition to follow quickly outside of China.
Pricing, however, could change as a result of the DeepSeek models. NVIDIA may be forced, for instance, to lower prices in view of competing models from U.S. startups mimicking open-source DeepSeek.
Futuriom Take: DeepSeek appears to have broken new ground in AI, but its ties to the PRC government encourage caution. At the same time, tech firms in the West are prepared to quickly address the innovations DeepSeek has introduced. Pricing of chips and other infrastructure could be pressured as a result.
