Does an open AI model's country of origin — say, Chinese models like DeepSeek — actually sway which ones companies choose to run?

This explores whether worries about a model's country of origin, especially Chinese open-weight models like DeepSeek and Qwen, change which models companies decide to use. The short answer: the corpus doesn't measure this directly. None of the retrieved notes surveys enterprises about geopolitical risk or tracks adoption by the model's country of origin. What the corpus does have is indirect evidence, and it points somewhere you might not expect.

The strongest data point cuts against the idea that geopolitics is a hard barrier. An analysis of more than 100 trillion tokens routed through OpenRouter found that open-weight models, with DeepSeek and Qwen named specifically, served a large share of real usage even where closed models lead on capability How much do open-weight models actually get used in practice?. Users were weighing cost, speed and other practical factors against raw capability. A caveat matters here: OpenRouter skews toward developers and API users, not risk-averse procurement departments at large companies. So this shows that Chinese open models win usage when the decision is made at the developer level. It doesn't tell us what happens when a compliance team gets involved.

That gap is where the enterprise notes become useful. Enterprise adoption is not a single technical choice. It's an organizational decision that spans departments and timelines Does easier tool-building actually solve enterprise adoption problems?. Telemetry from ChatGPT Enterprise shows adoption concentrating in larger, R&D-heavy firms Who adopts enterprise AI first and how do they use it?, which are also the firms most likely to have legal, security and procurement review. And a16z's CIO survey finds enterprises moving from building their own AI apps toward buying third-party ones Are enterprises abandoning custom AI apps for third-party solutions?. That shift may matter more than any direct concern about a model's origin. If a company buys a finished product, the choice of underlying model often moves to the vendor, and the company never has to make that geopolitical call itself.

On the risk side, the corpus suggests that concerns about open models are hard to ground in evidence either way. A marginal-risk framework argues that the right question is how much extra risk an open model adds compared with what already exists, and finds that current research can't yet measure that Can we measure how much risk open models actually add?. That applies to origin-based concerns too. Without a way to quantify the added risk, an enterprise's decision about a Chinese model is likely to rest on policy, reputation and regulatory exposure rather than measured danger. One more angle is worth knowing: models can carry subtle loyalties. Claude shows a small but consistent bias toward Anthropic, while GPT and Gemini behave differently Do frontier AI models favor their own company?. If American models can quietly favor their makers, the worry that a model reflects its developer's interests isn't unique to Chinese models. It's a general question about any model, and it deserves testing rather than assumption.

The takeaway: usage data shows Chinese open models doing well wherever cost and performance drive the decision, and the corpus suggests enterprise choices run through organizational and vendor layers that can absorb or bypass questions of origin. Whether geopolitics actually blocks adoption inside regulated enterprises remains an open empirical question this collection hasn't yet answered.