The story
On January 20, 2025, DeepSeek, a start-up in Hangzhou, China, released DeepSeek-R1, a reasoning model that performed close to OpenAI's o1 on math, coding and science tests. DeepSeek published the model's weights under an open MIT license, so anyone could download and run it. The company, backed by the hedge fund High-Flyer and led by Liang Wenfeng, said an earlier base model had cost about $5.6 million in computing for its final training run, a small fraction of what U.S. labs were thought to spend. Its chatbot app climbed to the top of Apple's U.S. App Store.
R1 was built on the ChatGPT-era recipe of large language models and on reasoning models that spend extra computation thinking before answering, which OpenAI introduced in late 2024. DeepSeek's contribution was a way to train this skill largely with reinforcement learning, rewarding the model for correct answers and letting it learn step-by-step reasoning on its own. The lab also worked around U.S. chip export limits with less powerful chips and efficient engineering. On January 27, Nvidia's stock fell about 17 percent, erasing roughly $590 billion in market value in one day, the largest single-day loss for any company on record. The $5.6 million figure excluded research, earlier experiments and hardware, so the true cost was higher.
Why it mattered
- Open-weight models from China became a serious rival to closed U.S. systems, and U.S. labs responded with faster releases.
- Markets reconsidered how much computing power AI progress requires, as questions about chip demand lingered.
- Policymakers revisited chip export controls and debated the security of using Chinese AI services.
Sources
- DeepSeek: What you need to know MIT CSAIL Alliances
- DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning Nature
Drafted by AI (Claude) from the sources above and checked against them by AI; no historian has reviewed this page. Spot a mistake? Suggest a correction. How pages are made
Community notes
Loading notes…