When DeepSeek released its R1 model in early 2025, the AI world was caught off guard. Here was a Chinese startup producing a reasoning model that matched — and in some benchmarks exceeded — OpenAI's o1, at a fraction of the training cost. The release sent Nvidia's stock tumbling and triggered a reassessment of just how much compute is actually needed to build frontier AI. Now, with R2 on the horizon, the pressure on Western labs is intensifying.
Architecture: Mixture of Experts Done Right
DeepSeek's secret weapon is its Mixture of Experts (MoE) architecture. Unlike dense models that activate all parameters for every token, MoE models selectively activate only a subset of expert neural networks for each input. DeepSeek-R1 used 671 billion total parameters but only activated 37 billion per forward pass — giving it the quality of a massive model at the inference cost of a much smaller one. R2 is expected to push this architecture further, with more granular routing and improved expert specialization.
This architectural choice is economical. Training costs for R1 were reported at around $5.6 million, a figure that drew gasps from an industry accustomed to spending hundreds of millions. GPT-4's training reportedly cost over $100 million. The efficiency gap has profound implications for who can afford to build frontier AI.
Benchmark Performance: Matching the Best
On AIME 2024 (a rigorous math olympiad benchmark), DeepSeek-R1 scored 79.8%, compared to OpenAI o1's 79.2%. On MATH-500, R1 achieved 97.3%. On Codeforces, R1 placed in the top 96.3% of human competitive programmers. These numbers represent genuine frontier performance, not marginal improvements over previous systems.
R2 is anticipated to improve on these numbers, particularly in multi-step reasoning tasks requiring maintenance of context across long chains of thought. Early leaks suggest improved performance on scientific reasoning and complex code generation tasks that stumped R1.
The Cost Disruption
Perhaps the most disruptive aspect of DeepSeek's models is their pricing. DeepSeek-R1 API access costs approximately $0.55 per million input tokens — compared to OpenAI o1's $15 per million. That is a 27x price difference for comparable performance. For companies building AI-powered products, this changes the economics of everything.
OpenAI responded with price cuts and accelerated the release of its o3 and o3-mini models. Google similarly began offering more competitive pricing on Gemini. The competitive dynamics that DeepSeek introduced to the market may ultimately benefit developers and end users, even if it creates pressure on AI labs' margins.
Geopolitical Dimensions
DeepSeek's success has become a flashpoint in the broader US-China technology competition. The US has imposed export restrictions on advanced Nvidia chips — the H100 and A100 — to prevent China from acquiring hardware needed to train frontier models. Yet DeepSeek achieved this with older H800 chips, specifically designed to comply with export rules.
This raises uncomfortable questions for US policymakers: if export controls can be circumvented through algorithmic efficiency gains, what is their actual effectiveness? Some analysts argue the controls slow Chinese progress. Others contend that DeepSeek proves talent and ingenuity can overcome hardware limitations. The debate will intensify as R2 approaches.
Open Source Strategy
DeepSeek has released model weights openly, allowing researchers and companies to run the models locally. This contrasts with OpenAI and Google's closed-source approaches. The open release has sparked an ecosystem of fine-tuned versions, distillations, and derivative models. Companies like Ollama have made it trivially easy to run DeepSeek-R1 on consumer hardware.
This open-source strategy may be calculated. By releasing weights, DeepSeek gains goodwill in the developer community, attracts research talent, and makes its architectural innovations the foundation for an ecosystem — while OpenAI and Anthropic remain perceived as closed corporate gatekeepers. R2 will test whether this strategy scales to even more capable models.
Conclusion
DeepSeek has shattered the assumption that frontier AI requires frontier spending. With R2, the company is poised to deepen its challenge to Western AI dominance. For developers, this means more capable models at lower prices. For policymakers, it is a wake-up call about the limits of hardware-focused containment strategies. And for the AI industry as a whole, it is proof that the race is far more global — and far more competitive — than anyone expected.