The problem
By 2023 frontier AI looked like a rich-country club: training a top-tier model cost hundreds of millions of dollars, so only a handful of US labs with giant compute budgets could compete — developers everywhere else were stuck renting closed APIs at whatever price Silicon Valley set.
How it works
DeepSeek, founded by Liang Wenfeng in 2023 and funded through his quant hedge fund High-Flyer, trains large language and reasoning models and releases their weights openly. Developers can download, customise, and self-host models like R1 and V4 Pro, or call them through DeepSeek’s own low-priced API. The January 2025 release of DeepSeek-R1 — a reasoning model built for a fraction of typical US training costs — made global headlines as a serious ChatGPT rival.
Pain points
Prohibitive training and inference costs, lock-in to closed foreign APIs, and no way to inspect or fine-tune the models their products depend on — the exact frustrations DeepSeek’s open weights and cheap tokens address.
Business model
Volume-priced API and platform access (V4 Pro at about US$1.32 per million input tokens, far above fellow Chinese lab Z.ai’s US$0.15 flash pricing but well below US frontier rates), with free open-weight downloads acting as distribution: ubiquity first, paid inference and enterprise adoption second.
Challenges
US labs keep raising the spend bar; American chip export controls squeeze compute access; open weights are hard to monetise against closed-model competitors; and Washington now treats Chinese open-source momentum as a geopolitical threat, inviting regulatory pressure at home and abroad.
Funding
- Raised: no outside capital — owned and funded by Liang Wenfeng via High-Flyer (MISSING).
- Valuation: no public priced round (MISSING).
Latest — September 2026
DeepSeek released DeepSeek-V4-Pro, ranked 14th on Artificial Analysis intelligence benchmarks — trailing OpenAI, Anthropic, and domestic rival Moonshot’s newly released Kimi K3 — as Chinese open-source models close in on, but still trail, the top US offerings.