Digital Sovereignty

A Chinese Model May Be Answering Your Prompts

On OpenRouter, Chinese-built models have taken over 30 percent of US company tokens since February, peaking near 46 percent. If you route through an aggregator, you cannot say which model read your data.

Digital SovereigntyBy Servola Tech Desk2026-07-093 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

A Chinese Model May Be Answering Your Prompts

Key takeaways

On OpenRouter, Chinese-built models have carried more than 30 percent of US companies' tokens every week since February 8 and peaked near 46 percent, up from an 11 percent yearly average, and because European firms route through the same aggregators, most cannot name which model served a given request.
The pull is price: open Chinese models run 60 to 90 percent cheaper than the leading Anthropic and OpenAI models, and Z.ai's GLM 5.2 saw daily token volume grow about 27 times in its first full week on Vercel, so the cheapest-model router is quietly changing what runs your workload.
A US House investigation by the Homeland Security and China committees is probing censorship, distillation, and security risk, but Europe's exposure is different: from August 2 the EU AI Act's transparency duties and the GDPR both assume you can state model provenance, which aggregator routing hides.
The operator move is not a ban but a provenance policy: pin the providers your router may use, log the exact model per request, and treat cheapest-available-model as a compliance decision rather than a billing default.

The number that changed under the hood

Chinese-built models now carry a large share of Western AI traffic, and the clearest measure comes from OpenRouter, the marketplace that lets a developer send a prompt to whichever model is cheapest or fastest. CNBC reported that the share of tokens from US companies going to Chinese models has stayed above 30 percent every week since February 8 and reached as high as 46 percent, against an 11 percent average over the prior twelve months and just 4.5 percent in the first half of 2025.

That is not a fringe experiment. It is a third of measured demand on a mainstream routing platform, moving in a single quarter. The models are open-weight releases from labs such as DeepSeek and Z.ai, reached through the same Western rails that European teams use every day.

Why the router keeps choosing them

Price is doing the work. Open Chinese models run roughly 60 to 90 percent cheaper than the leading models from Anthropic and OpenAI, and they have closed enough of the quality gap that the discount is no longer paid for in capability. Z.ai's GLM 5.2, released in June, became the fastest-adopted model Vercel tracked in 2026, with daily token volume growing about 27 times and its customer count about 80 times in its first full week.

The decisive detail is that most of this is automatic. A cost-optimizing router does not ask permission before switching the model behind an endpoint, so a workload that ran on a US model in January can be running on a Chinese one in July without a single line of your code changing. Coinbase has said openly that it cut its AI spend by half by moving to Chinese models. Yes, but the saving arrives with a question most buyers have not answered: which model, exactly, is now in the loop.

Europe's exposure is not America's

Washington is treating this as a security file. A joint investigation by the House Committee on Homeland Security and the House Select Committee on China is examining censorship built into model outputs, the risk of distillation, and data exposure. That frame is American, and it points at American firms.

Europe inherits the same models through the same aggregators but sits under a different rulebook. From August 2, 2026, the EU AI Act's transparency obligations for general-purpose models take effect, and the GDPR already expects a controller to know where personal data is processed and by whom. Both assume you can state provenance: which model handled a request, trained on what, hosted where. A router that silently swaps a US model for a Chinese one does not break your budget, but it can break that answer, and the answer is the part a regulator asks for.

Make provenance a setting, not a surprise

The fix is governance, not a boycott. Start by listing the models your router is permitted to use and removing the ones you cannot account for, so routing is a decision you made rather than one the price feed made for you. Log the exact model and provider on every request, because an audit trail you can produce after the fact is worth more than a policy you only assert.

Then decide provenance per workload. A public-facing translation may not care which open model runs it; a contract review touching personal data should be pinned to a named, documented provider. The teams that get caught out this year will not be the ones that chose a Chinese model on purpose. They will be the ones that never chose at all and could not say what they were running.

Frequently asked questions

Are Chinese AI models banned in Europe?

No. Open-weight models from labs such as DeepSeek and Z.ai are freely available and are reached through the same platforms European developers already use. The issue is not legality but provenance: knowing and being able to prove which model processed a given request, which the EU AI Act and GDPR both assume.

How would a Chinese model end up in our stack without us choosing it?

Through a cost-optimizing router or aggregator such as OpenRouter, which sends each request to whichever model is cheapest or fastest. If you have not pinned the allowed providers, the model behind your endpoint can change automatically, so a workload that ran on a US model can move to a Chinese one with no change on your side.

What should we actually do about it?

Pin the providers your router may use, log the exact model and provider on every request, and set provenance rules by workload, so sensitive tasks run only on named, documented models. That turns model choice into a governed decision rather than a billing side effect.

The surprise is not that Chinese models got good and cheap. It is that they can enter your stack without anyone deciding to let them in. Name the models you allow, record the one that runs each request, and the shift from a US model to a Chinese one becomes a choice you can defend rather than one you discover during an audit.

Chinese AI modelsOpenRouterDeepSeekGLM 5.2EU AI ActModel provenanceDigital Sovereignty

A German State Moves 50,000 Staff Off Microsoft

Mecklenburg-Vorpommern is moving more than 50,000 public workers off Microsoft to Nextcloud, and running its AI on EU models. What the shift signals for owners.

4 min read

Digital Sovereignty

IQM Took Europe's Quantum Bet Public in New York

IQM, Europe's flagship quantum firm, went public on 2 July 2026 - on Nasdaq, not a European exchange, and its own prospectus warns quantum may never pay off. The sovereignty gap is capital, not science.

3 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.