Open Models

On Monday 27 January 2025 Nvidia's shares fell about 17%, and the company lost about $589 billion of market value, the largest one-day loss for any company to that date (market data). Broadcom, TSMC, power and data-centre stocks fell with it. The same day the chatbot app of DeepSeek, a Hangzhou lab of about 160–200 people (reported), passed ChatGPT to become the most downloaded free app on the US App Store, and DeepSeek limited new registrations to Chinese phone numbers after what it called large-scale malicious attacks. President Trump called DeepSeek "a wake-up call" for American industry (reported).

The trigger had been released a week earlier, on 20 January 2025. DeepSeek published R1, a reasoning model that roughly matched OpenAI's o1 on mathematics and coding benchmarks (company-reported, broadly confirmed by independent tests at the time, reported), together with six smaller distilled models, under the MIT licence, which lets anyone use, modify and distil them. Its API charged $0.55 per million input tokens and $2.19 per million output tokens, against $15 and $60 for o1 (list prices, January 2025). Behind the alarm was a number from DeepSeek's December 2024 technical report on its V3 base model: the final training run had used 2,048 Nvidia H800 GPUs for about 2.788 million GPU-hours, about $5.576 million at assumed rental prices (company claim). If frontier AI could be built for millions rather than billions, investors asked, perhaps the hyperscalers' planned spending, and Nvidia's earnings, were too high.

The fear passed within weeks. The lesson about open models did not. This chapter explains why companies give away models that cost a fortune to train, who makes money around them, and what they do to everyone else's prices.