DeepSeek-R1 — the open source shock

The shot heard around the AI world

In January 2025, the Chinese AI lab DeepSeek released DeepSeek-R1 — an open source reasoning model that matched or exceeded GPT-4o and Claude 3.5 Sonnet on multiple benchmarks, including math, coding, and scientific reasoning. The shock was not just the performance, but the cost: DeepSeek claimed to have trained R1 for approximately $5.6 million in compute costs — a fraction of the hundreds of millions reportedly spent on comparable American models. Nvidia's stock dropped 17% in a single day.

How DeepSeek achieved this

DeepSeek-R1 used reinforcement learning from scratch — starting from a base model and using RL to develop reasoning capabilities without supervised chain-of-thought data. This was a methodologically different approach from OpenAI's o1 and was reportedly significantly more compute-efficient. DeepSeek also used mixture-of-experts (MoE) architecture and aggressive quantization. Researchers who studied the paper noted it was technically rigorous and the results reproducible.

Geopolitical and strategic implications

DeepSeek-R1 shattered the assumption that frontier AI required the enormous compute resources and capital available only to American tech giants. It demonstrated that Chinese labs — despite export restrictions on advanced NVIDIA chips — could reach or approach the frontier through algorithmic innovation. It reignited the debate about AI export controls, compute governance, and whether hardware restrictions are an effective strategy for maintaining AI advantage. It also validated open-source AI as a viable path to frontier capability.

Sources

• DeepSeek-AI (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948.
• Wikipedia — DeepSeek