DeepSeek

DeepSeek was founded in 2023 as an AI research subsidiary of High-Flyer (Huanfang), a Hangzhou-based quantitative hedge fund that had accumulated one of the largest GPU clusters in China before export controls tightened. The lab's founding team — led by Liang Wenfeng, also High-Flyer's CEO — brought deep expertise in numerical optimization from quantitative finance directly to bear on AI training efficiency, which proved to be a decisive advantage. DeepSeek-V3, released in December 2024, and DeepSeek-R1, released in January 2025, constituted the most significant competitive disruption to US AI labs since GPT-4. DeepSeek claimed V3 was trained for approximately $5.6 million in GPU-hours, a figure that was disputed in detail but nonetheless pointed to a radical efficiency gap versus the hundreds of millions estimated for comparable US frontier models. The R1 model introduced reinforcement learning from human feedback combined with chain-of-thought reasoning in a way that produced benchmark results competitive with OpenAI's o1, triggering a significant sell-off in NVIDIA stock (the so-called "DeepSeek moment" on January 27, 2025) as markets repriced assumptions about compute requirements for frontier AI. DeepSeek's efficiency techniques are technically substantive. The V3 architecture uses Multi-Head Latent Attention (MLA), which compresses the KV cache using low-rank projection matrices, dramatically reducing memory bandwidth pressure during inference. It also employs an MoE routing scheme with 256 experts where only 8 are active per token, and uses FP8 mixed-precision training — a then-novel approach that reduces memory footprint and improves throughput on NVIDIA H800 GPUs (the export-control-compliant version of H100 with reduced NVLink interconnect bandwidth). DeepSeek's training was reportedly conducted on a cluster of approximately 2,048 H800 GPUs. Export controls are the defining constraint on DeepSeek's hardware roadmap. The BIS October 2023 rules classified H100/H200 GPUs as controlled items requiring a license for export to China, and subsequent 2024/2025 rules tightened further. DeepSeek's use of H800s (which were compliant at time of purchase) has come under scrutiny for whether those chips were acquired through legitimate channels. As the H800 is now also controlled, DeepSeek is pivoting to Huawei's Ascend 910B and 910C NPUs for future training runs, despite their lower software maturity and interconnect bandwidth compared to NVIDIA's offerings. The geopolitical implications of DeepSeek's results reshaped the AI chip export control debate. US policymakers who assumed that restricting advanced GPU exports would maintain a durable performance gap now face evidence that efficiency-driven approaches can partially compensate for hardware disadvantage. Meanwhile, DeepSeek's open release of model weights under a permissive license made its technology globally accessible, accelerating the diffusion of its architectural innovations — including MLA and its MoE routing approach — into both open-source and proprietary model development worldwide.