What is DeepSeek, and why did everyone panic?
In January 2025, a Chinese AI lab released a reasoning model as good as America's best — built for a fraction of the cost. The panic wasn't about one model. It was the question: what if frontier AI is cheap?
At a glance
Jan 2025: DeepSeek's R1 matches top US models, open-sourced · Nvidia's record one-day market drop · Oct 2026: new open model + open-sourced inference engine
Catch: benchmark and cost figures are vendor-reported, not independently verified — and open weights aren't the full recipe.
The basics
DeepSeek is a Chinese AI lab. Its models are genuinely top-tier — and it gives them away: open weights anyone can download and run. That's the part that broke the old story, where frontier AI was supposed to be expensive and locked up.
The panic
January 2025: the market flinched. A reasoning model as good as America's best, for a fraction of the cost — and Nvidia lost more value in a single day than any company in history. The panic wasn't about one model. It was the question: what if frontier AI is cheap?
Still happening
October 2026: DeepSeek shipped another open model coding near the top of the charts at a fraction of the going API price — and open-sourced the engine that serves it. The moat keeps shrinking. Cheap, open models keep arriving, and the winners are whoever builds on them fastest.
The catch
Those benchmark scores come from DeepSeek itself, not yet independently verified. And open weights aren't the full recipe — the training data stays secret. Treat the headline numbers as claims, not facts, until independent testing confirms them.
Our takeaway
Don't bet on the moat — bet on the builders. Cheap, open models keep arriving; the advantage goes to whoever ships on them fastest.
Sources & further reading
DeepSeek's models and announcements are at deepseek.com. Independent benchmark verification lags vendor announcements — check current leaderboards before treating any score as settled.