DeepSeek

What they built

DeepSeek ships foundation models optimized for reasoning, code, and efficient inference — with a research culture that prioritizes architectural innovation over raw scale. The model family spans DeepSeek-V3 (general-purpose), DeepSeek-R1 (reasoning), DeepSeek-Coder, and the upcoming V4 (trillion parameters, 1M context). The open-weight release pattern has built an enormous developer community that deploys DeepSeek models in environments where OpenAI and Anthropic APIs are inaccessible or undesirable.

How they got here

Liang Wenfeng founded DeepSeek in 2023 as a spin-out from High-Flyer Capital, the Chinese quant hedge fund he also founded. The unusual founding structure — parent-funded, with no external VCs for the first two years — gave DeepSeek the freedom to prioritize research quality over GTM velocity. The team stayed small (~200), the research output stayed disproportionate, and the model releases built a compounding reputation.

The inflection was January 2025. DeepSeek-R1 matched OpenAI’s o1 on reasoning benchmarks at a tiny fraction of the training cost, and the open-weight release let anyone verify the result. US equity markets moved on the news — a single Chinese open-source release caused a notable correction in AI infrastructure stocks. Through 2025, DeepSeek shipped V3, R1.5, Prover, and a steady cadence of research papers that tracked the frontier closely. V4 launches in late April 2026 with a trillion-parameter MoE architecture and 1M-token context.

As of April 2026, DeepSeek is running its first external fundraise — $300M+ at a valuation of at least $10B, with reported offers up to $20B from Tencent, Alibaba, and others. The shift from fully internal funding to an open cap table signals the company is preparing for a longer, more capital-intensive phase.

What’s ahead

Three questions dominate. First, V4 and beyond: can DeepSeek maintain capability parity with US labs as training runs cross $1B in compute cost? The architectural-efficiency thesis is credible but unproven at the frontier. Second, geopolitics: US export controls have tightened through 2025 and 2026; DeepSeek’s continued access to compute depends on either domestic Chinese alternatives (Huawei Ascend) reaching parity or smuggled inventories holding out. Third, commercial model: DeepSeek has never optimized for enterprise revenue the way US labs have; the external fundraise signals that’s changing.

Why it matters

DeepSeek is the single most consequential AI company outside the US — and the one whose trajectory will most shape whether AI converges toward a US-dominated oligopoly or a multipolar landscape. For AI founders globally, DeepSeek’s story is a proof that a small team with strong research discipline can genuinely compete at frontier. For investors, understanding DeepSeek is how you avoid being caught on the wrong side of the compute-moat narrative.