DeepSeek-R1 and the efficiency shock (2025)
A Chinese startup released an open-weights reasoning model that rivaled the best at a fraction of the reported cost, and on January 27, 2025 it wiped a record sum off Nvidia's value.
On January 20, 2025, the Chinese startup DeepSeek released R1, a reasoning model that matched OpenAI o1 on many public tests. Two things made it land like a thunderclap: it was released with open weights, so anyone could download and run it, and the company claimed it had trained its base model for a few million dollars, a tiny fraction of what Western labs were spending.
Why the market panicked
For two years the AI story on Wall Street had a simple shape: progress needs enormous compute, so buy the chips. DeepSeek questioned that premise. If a strong model could be trained cheaply and given away, the case for hundreds of billions in spending looked shakier.
On Monday, January 27, 2025, the doubt hit the market. Nvidia fell about 17 percent in a single day and shed close to 600 billion dollars in value, the largest one-day loss for any company in Wall Street history to that point, dragging chipmakers and power suppliers down with it.
What was actually new
R1 leaned on reinforcement learning to teach the model to produce long chains of thought before answering, much like o1, but the recipe was described openly. It also built on a mixture-of-experts design, activating only part of the network per query to save compute. DeepSeek even released smaller “distilled” versions, extending the open-weights tradition of Llama 2.
Analysts later noted the training-cost figure covered only one run and left out much of the real bill, and that the company still owned a large fleet of Nvidia chips. The efficiency was real; the framing was generous.
Why it matters
DeepSeek-R1 reset expectations about who can build a frontier model and how much it must cost. It strengthened the open-weights camp, complicated the geopolitics of chip export controls, and forced a harder question about the Stargate-scale spending that the same month was being announced as the future of AI.