DeepSeek: The Chinese Lab Changing Everything
How a Chinese hedge fund's AI lab built models that match OpenAI at a fraction of the cost, and what DeepSeek's open-weight releases mean for the global AI race.
From Hedge Fund to AI Lab
DeepSeek is the AI research arm of High-Flyer, a Chinese quantitative hedge fund. The connection is not coincidental: quantitative trading requires massive computational infrastructure, strong mathematical optimization, and world-class engineering talent, all of which transferred directly to AI research. The lab started publishing AI papers in 2023 and rapidly progressed to releasing competitive models.
Unlike OpenAI or Anthropic, which are dedicated AI companies, DeepSeek operates as a research division of a profitable financial firm. This gives it unusual financial stability and patience, it doesn't need AI to be a commercial success to survive, enabling long-term research investments without venture capital pressure.
DeepSeek's Key Models
DeepSeek V3 (December 2024) is a 671B Mixture of Experts model (37B active parameters) that matches GPT-4o on coding benchmarks and outperforms it on mathematics, at roughly 10% of the cost. Input: $0.27/M tokens. Its technical report describing MLA (multi-head latent attention) and auxiliary-loss-free load balancing is among the most detailed publicly available. V3 is a Foundation Model; DeepSeek also releases instruction-tuned Instruct Model variants optimised for chat and tool use.
DeepSeek R1 (January 2025) is the reasoning model that sparked the January 2025 AI crisis: an open-weight model matching OpenAI o1 on most benchmarks. Released on Hugging Face for free, it demonstrated that reasoning model training was not an exclusive OpenAI capability. R1 was created using a pure RL approach on verifiable tasks with minimal SFT, showing a more efficient training path.
Technical Contributions
DeepSeek has contributed meaningfully to the technical literature. Multi-head Latent Attention (MLA) compresses the KV cache by using a low-rank projection for keys and values, dramatically reducing memory requirements. Their FP8 training approach, a form of Quantization applied during training rather than just inference, enables efficient training at reduced hardware cost with minimal quality loss. Chain of Thought reasoning is central to R1's design: the model generates extended thinking traces before committing to a final answer.
The claim that DeepSeek V3 was trained for approximately $6M in compute costs (vs. $100M+ for comparable US models) attracted enormous attention. While some skepticism about the exact figure is warranted, the general point stands: significant algorithmic innovation can compensate for compute deficits. This has major implications for AI policy and the effectiveness of export controls.
Geopolitical Significance
DeepSeek's success is geopolitically significant because it emerged despite US export restrictions on advanced Nvidia chips to China. The models were reportedly trained primarily on Nvidia H800 chips (an export-controlled chip inferior to H100), demonstrating that hardware constraints can be partially overcome through algorithmic innovation.
For the AI policy community, DeepSeek raises important questions: Can export controls effectively slow Chinese AI capability development? What is the optimal frontier between openness (which benefits the global scientific community) and control (which might preserve strategic advantage)? These questions will shape AI governance for years.
Read next
How DeepSeek Disrupted the AI Industry
The story of how a Chinese hedge fund's AI lab released models that matched OpenAI at a fraction of the cost, and what it means for the global AI race.
Open-Weight vs Open-Source Models: What's the Difference?
Why 'open-source AI' is often a misleading term, and what it actually means when a model is open-weight, what's included, what's not, and why it matters for developers.
The Rise of Reasoning Models: How AI Learned to Think
From GPT-4 to o3 and beyond. How reasoning models work, why they differ, and what they mean for the future of AI capabilities.