DeepSeek R1
DeepSeek positioned R1 as an open-weight reasoning model with performance "on par with OpenAI-o1" across math, code, and reasoning benchmarks, released under a permissive MIT License that explicitly allowed free commercial use and distillation. It argued, via the variant R1-Zero, that reasoning ability could emerge from pure reinforcement learning without human-labeled reasoning traces, and it shipped six smaller "distilled" models so the capability could run on modest hardware. The surrounding narrative emphasized extreme cost and training efficiency, anchored by the widely repeated ~$5.6M figure. API access was priced far below o1 — roughly $0.55 per million input tokens and $2.19 per million output tokens.
On launch day DeepSeek published the open weights, a technical report, the R1-Zero variant, and six distilled models (1.5B-70B) under MIT, plus a low-cost hosted API and a free consumer app. Within days third-party inference providers and major clouds were hosting it, and by January 26 the free app reached No. 1 on the U.S. Apple App Store, displacing ChatGPT.
The core technical achievement was real and largely survived scrutiny: independent analysts agreed R1 reached roughly o1-class reasoning, and a Chinese lab catching the frontier in months was genuinely significant. The headline economics, however, were oversold — the famous ~$5.6M number was the GPU pre-training cost of the underlying V3 base model, not R1, and it excluded R&D, infrastructure, data, and staff; analysts estimated DeepSeek's total compute investment in the billions with tens of thousands of GPUs. That misread number nonetheless drove a market panic: on January 27, 2025, Nvidia fell roughly 17% in a single day, the largest one-day market-cap loss in U.S. history. Two further caveats matter for decision-makers: R1 does not beat o1 on every benchmark, and as a Chinese-origin model it applies political censorship on topics sensitive to Beijing — a real data-governance consideration, partly mitigated by the option to self-host the open weights.
R1 was the moment a frontier-class reasoning model became open-weight, cheap to run, and freely modifiable — collapsing the assumed cost moat around proprietary AI and forcing every enterprise into a real build-vs-buy conversation. For executives it is also a clean case study in separating a genuine capability shift from a market over-reaction to a misread cost figure.