DeepSeek

DeepSeek is a Chinese AI research company founded in 2023 as a subsidiary of High-Flyer, a quantitative hedge fund, and based in Hangzhou. It made global headlines in January 2025 with DeepSeek-R1, a reasoning model that matched or exceeded leading US models at a fraction of the reported training cost — a watershed that challenged the assumption that frontier capability requires hundreds of millions of dollars of training compute, and briefly wiped hundreds of billions from NVIDIA's market value. It took its first outside capital in June 2026, in a round whose structure is as notable as its size: founder Liang Wenfeng contributed heavily himself, commercial investors led by Tencent accepted five-year lockups and no voting rights, and the only investor granted governance rights was a state-backed national AI fund. Control is concentrated with the founder and the Chinese state.

The lab releases its models as open weights under permissive licenses and runs a chat platform and API priced well below Western alternatives. Its current generation is a natively multimodal mixture-of-experts that activates only a small fraction of its parameters per token and ships ungated under a plain MIT license with no revenue, user-count or territory conditions. The larger prior flagship remains the biggest open-weight model available under a fully permissive license — later models from other Chinese labs have passed it on raw size, but carry custom terms. A practical consequence worth knowing: with this lab the effective price is set by which checkpoint the default endpoint points at rather than by the rate card, and retiring a larger model in favour of a cheaper one has moved prices sharply without any headline change.

Open availability has a safety corollary that is easy to state imprecisely. Researchers documented a threat actor wiring DeepSeek into an open-source agent framework and running a largely autonomous intrusion campaign across hundreds of targets from a single instruction, while a US provider refused equivalent requests and disabled the operator's account. The contrast is worth holding precisely: DeepSeek is markedly more restrictive than US models on political topics and markedly less restrictive on offensive-security ones. Those are two different axes, and they are easy to conflate.

Liang Wenfeng has publicly committed the lab to an open-source path and to pursuing artificial general intelligence rather than near-term commercialization — an unusual posture at this tier, where most frontier labs draw their largest checks from corporate cloud partners and treat AGI claims with strategic ambiguity. That independence sits inside real state entanglement: Beijing has extended state-secret travel restrictions to AI researchers, binding the lab into a managed talent regime, while on the US side an interagency committee has approved DeepSeek for addition to the Commerce Department's Entity List on national-security grounds, with the listing so far withheld. A separate step has gone through: in October 2026 the Department of Homeland Security published an order, signed in July, that requires federal civilian agencies to remove DeepSeek and stop buying its products. It reaches past the hosted app to third-party products built on DeepSeek's models, so inside those agencies running the open weights in someone else's software is no way around it. The lab has also begun recruiting chip engineers to design an in-house inference processor, aiming to reduce dependence on both Nvidia and Huawei as export controls tighten — a years-long undertaking made harder by restrictions on Chinese access to advanced foundries.