Overview
On January 20, 2025, the Chinese AI company DeepSeek released R1 — an open-source reasoning model that matched OpenAI’s o1 across standard benchmarks, including mathematics, coding, and scientific reasoning. It was released under the MIT license, meaning anyone could download, modify, and deploy it freely.
The model trained for approximately $6 million in compute costs — compared to estimates of $100 million or more for comparable US frontier models. It used roughly one-tenth the compute of Meta’s Llama 3.1 405B for comparable performance.
Seven days later, on January 27, Nvidia’s stock fell 18% in a single day, wiping approximately $593 billion in market capitalization — the largest single-day market cap destruction in US stock market history. The US AI industry had not expected this.
What R1 Achieved
DeepSeek R1’s benchmark performance:
- AIME 2024: 79.8% (OpenAI o1: 79.2%)
- MATH-500: 97.3% (OpenAI o1: 96.4%)
- Codeforces Elo: ~2,029 (OpenAI o1: ~1,819)
- GPQA Diamond (PhD-level science): 71.5%
These numbers were not just competitive — on several measures, R1 exceeded o1. And it was open-source.
Why the “Sputnik Moment” Framing
Venture capitalist Marc Andreessen coined the phrase “Sputnik moment” within days of the release, and it stuck. The analogy was precise:
- When the Soviet Union launched Sputnik in 1957, the shock was not that a satellite had orbited Earth — it was the gap in assumed capability. American experts believed Soviet technology was far behind. It wasn’t.
- With DeepSeek R1, the shock was not that China had built a good AI model. It was the assumption shattered: that frontier AI required massive, sustained compute investment protected by chip export controls, and that the US had a durable moat.
R1 demolished three assumptions simultaneously:
- Cost moat: Frontier reasoning models do not require $100M+ training runs
- Hardware dependency: China’s restricted access to H100 GPUs had not prevented model-matching performance
- Knowledge gap: Chinese AI research had absorbed enough from published Western papers to replicate and compete
The Technical Breakthrough: Pure RL Without SFT
A key reason R1 was so surprising was its training methodology. Most US reasoning models (including o1) used Supervised Fine-Tuning (SFT) on curated chain-of-thought data, followed by Reinforcement Learning from Human Feedback (RLHF).
DeepSeek demonstrated that pure reinforcement learning — training directly on outcome rewards (correct/incorrect answer) without any SFT data — could produce emergent chain-of-thought reasoning. This had been theorized but never demonstrated at scale. The model, incentivized only by whether its final answer was right, spontaneously developed behaviors including self-reflection, hypothesis testing, and backtracking.
This finding — later published in the technical report — reshaped understanding of how reasoning capabilities emerge and how cheaply they can be cultivated.
Geopolitical Fallout
The week of January 27, 2025:
- DeepSeek R1 became the most downloaded free app on the US iOS App Store, surpassing ChatGPT
- US lawmakers called for emergency hearings on AI competitiveness
- The Biden-era export controls on advanced chips to China were immediately questioned in their efficacy
- The Trump administration, inaugurated the same day R1 was released, signaled it would pursue AI dominance through deregulation rather than export restriction
- Taiwan semiconductor supply chain stocks fell; “compute scaling is the only moat” thesis received its most severe challenge to date
Why This Matters
DeepSeek R1 did not just release a competitive model. It redistributed the cost structure of frontier AI — which has implications for every company, government, and individual whose plans depended on AI capability remaining concentrated in a small number of well-capitalized US firms.
It also raised a question the field is still answering: If capable reasoning models can be trained for $6M, what happens to the AI economy when training costs fall by another 10x?