DeepSeek R1: The $6 Million Model That Shook a $3 Trillion Industry

On January 20, 2025, DeepSeek, a Chinese AI lab backed by quantitative hedge fund High-Flyer, released DeepSeek R1 with open weights publicly available. On most benchmarks, R1 matched or exceeded OpenAI's o1, which had launched four months earlier after significant compute investment and with no public weight release. DeepSeek's paper stated the pre-training compute cost was approximately $5.6 million. NVIDIA's market cap fell by roughly $600 billion the following trading day: the largest single-day market cap loss in US stock history at that point.

The technical approach was distinctive. R1 was trained primarily using reinforcement learning from sparse reward signals: the model was rewarded for arriving at correct answers on math and coding problems, without supervised fine-tuning on human-written chain-of-thought data. Extended reasoning traces emerged from the RL process rather than being explicitly taught. DeepSeek also published DeepSeek-V3, a Mixture of Experts base model trained with unusual efficiency through FP8 mixed precision and optimised GPU communication patterns, squeezing more out of hardware that, due to US export controls, was less powerful than what US labs used.

The cost figure was the most destabilising element. The prevailing assumption in the AI industry was that frontier AI required hundreds of millions or billions of dollars in compute. If a Chinese lab could produce a competitive reasoning model for $6 million, the economic moat of frontier AI training appeared more fragile than assumed. Whether the figure was complete (it excluded data collection costs, failed experiments, and years of prior research infrastructure) was debated widely. But even at 10× the stated cost, it was a fraction of what US labs had spent on comparable capability.

The NVIDIA stock reaction exposed a thesis embedded in the AI bull market: AI performance scales with compute, compute requires NVIDIA GPUs, therefore NVIDIA supply is rate-limiting and premium pricing is justified indefinitely. DeepSeek's efficiency demonstrated that the scaling-compute story had assumptions embedded in it that might not hold. Efficiency research could substitute for raw scale. The gap between expensive and cheap training could close faster than assumed. Whether this would play out at the frontier remained an open question, but the market had believed the thesis was certain, and R1 introduced uncertainty.

DeepSeek R1 weights were downloaded millions of times in the week following release. Chinese labs had been assumed to be 12–18 months behind the US frontier. R1 showed that gap was either narrower than assumed, or that "frontier" in reasoning tasks was reachable by different routes. US export controls on advanced chips, designed to slow Chinese AI development, had not prevented DeepSeek from building competitive reasoning capability. The geopolitics of AI compute had acquired a new data point that fit nobody's prior neatly, and the conversation about what "AI leadership" means had become more complicated.