DeepSeek's R1 model trained with a fraction of the compute of US frontier models yet matches or beats them on reasoning benchmarks. The AI world took notice.

In January 2025, Chinese lab DeepSeek released R1 — an open-source reasoning model that:

- Matches OpenAI o1 on many math and coding benchmarks

- Was reportedly trained for ~$6M in compute (vs estimated $100M+ for comparable US models)

- Uses a novel training approach: pure RL with process reward models, no SFT warm-up

**Why it matters:** Suggests compute efficiency gains are far from exhausted. The assumption that frontier AI requires massive datacenter build-outs may be wrong.

**The policy angle:** Export controls on Nvidia chips may be less effective than assumed if models can be trained efficiently on lower-tier hardware.

DeepSeek R1: Open-Source Reasoning at Frontier Level