On June 3, 2026, DeepSeek — the Hangzhou lab behind the MIT-licensed R1, V3 and V4 models — was reported to be closing its first-ever external funding round: roughly $7.4 billion (about 50 billion yuan) at a valuation of $52–59 billion. Tencent is weighing about 10 billion yuan, battery giant CATL around 5 billion, and China's state-backed AI investment funds are participating, while founder Liang Wenfeng contributes some 20 billion yuan — roughly 40% — himself (Sahm Capital). The headline is the number. The story is what the number signals: open-weight AI has become an instrument of Chinese industrial policy.
A subsidy stack, not a startup round
DeepSeek built its reputation by giving its weights away. That is unusual for a company now valued near $59 billion, and it only makes sense inside a policy environment engineered to reward it. China's national AI industry investment fund, launched in January 2025, was billed at 60.06 billion yuan (about $8.2 billion), co-financed through the third phase of the National Integrated Circuit Industry Investment Fund (Global Times). State capital of that scale does not chase open-weight labs by accident; it does so because openness has been designated a strategic asset.
The legal architecture points the same way. The scholarly Model AI Law, whose version 2.0 was unveiled on April 16, 2024, proposes that open-source AI "provided for free and with transparency can be exempt from legal liability," and pairs that carve-out with affirmative industrial support — an open-source AI foundation, tax incentives, development platforms and government procurement of compliant open-source products (IAPP). The exemption is conditional, not absolute: developers must meet transparency and risk-prevention duties such as documentation and basic technical guardrails (Just Security). But the direction is unmistakable. Liability relief, state equity and procurement preference are being stacked behind models that ship their weights.
The case for Beijing's approach
It is worth stating the strongest version of the strategy before contesting it. A liability safe harbor for genuinely open models is defensible policy. A developer who publishes weights for free, with documentation, cannot supervise every downstream deployment; holding them strictly liable for every misuse would simply end open releases, concentrating capability in a few closed incumbents. The European Union's own AI Act grants lighter obligations to free and open-source models precisely on this logic. China's draft Artificial Intelligence Law, circulated in scholarly translation in May 2024, similarly frames AI development as a state priority to be promoted, not merely policed (CSET). Even China's binding 2023 rulebook — the Cyberspace Administration's Interim Measures for the Management of Generative AI Services — was written, in the State Council's own words, to put forward "measures on boosting generative AI technology on the one hand" while setting norms on the other (gov.cn). A regime that rewards transparency and openness is, on its face, more pro-innovation than one that defaults to closed control.
Where the model breaks
The problem is not the safe harbor. It is the coupling of the safe harbor to state capital and procurement. Genuine open-source liability relief is content-neutral: it protects any developer who publishes transparently. What Beijing is assembling is something narrower — a subsidy channel that flows to nationally favored labs, with openness as the qualifying condition rather than the point. When the National AI Industry Investment Fund and Tencent co-anchor a raise alongside the founder's own capital, "open" weights become a distribution mechanism for state-aligned models, optimized to become the global default layer on which other developers build.
That default carries embedded costs the license does not advertise. China's binding regime requires generative content to align with state values and obliges providers to use "legally sourced" data and to limit illegal-information generation (Just Security). A foreign team fine-tuning a Chinese open-weight base inherits those alignment choices unless it actively strips them out — and most will not. The openness is real at the weights layer and constrained at the values layer.
There is also a security dimension that openness amplifies rather than resolves. The same downloadable ubiquity that makes open models attractive makes them a wide attack surface: 2026 has already seen a critical "BadHost" flaw in the Starlette package that underpins millions of AI-agent servers, and a supply-chain poisoning spree corrupting hundreds of open-source tools (Ars Technica). A safe harbor that reduces developer accountability while state policy maximizes distribution is a combination regulators elsewhere should study carefully.
The proportionate answer
The right response from Washington, Brussels and Delhi is not to mirror China by closing their own ecosystems or punishing open releases — that would forfeit the genuine advantages of open weights to the one government willing to subsidize them. It is to keep the part that works and reject the part that does not. Liability safe harbors for transparent open models: yes, and democracies should legislate them clearly. State equity steering openness toward a single national champion, and procurement mandates that lock public infrastructure onto value-laden foreign weights: no. The lesson of DeepSeek's raise is that open source is no longer just an engineering culture. It is contested industrial policy — and being pro-open-source now requires being precise about whose openness, on whose terms.