Chinese AI Now Leads Global API Calls. Europe Must Choose.

Chinese AI models now account for 61% of token consumption on OpenRouter, with MiniMax's M2 alone processing 1.7 trillion weekly tokens. For European founders and CTOs, this is no longer a theoretical debate. The binary trap — adopt and risk dependency, or abstain and fall behind — is now a live operational decision.

What happened in China

As of late February 2026, Chinese-built models crossed a threshold that would have seemed implausible two years ago: for the first time, API calls to Chinese AI models surpassed those to US models on global platforms. OpenRouter data published on February 24, 2026 shows Chinese models at 61% of total token consumption on the platform. MiniMax's M2 model sat at the top, processing approximately 1.7 trillion weekly tokens. Chinese models have taken disproportionate market share particularly in agentic flows — the automated, multi-step AI pipelines that are increasingly central to how software is being built. Even US firms are running their agentic workloads on Chinese models.

With DeepSeek-R2 expected to launch in 2026 and anticipated to intensify competition further, this is not a peak. It is a trend line.

How it works / the detail

The driver is not novelty. It is cost and accessibility. Chinese open-weight models offer significant cost advantages over closed US alternatives, and their open-weight nature means they can be self-hosted, fine-tuned, and integrated without vendor lock-in to a US API provider. For agentic workloads — where token volumes scale fast and inference costs compound — these advantages are material.

But the compliance picture complicates the arithmetic. The EU AI Act's rules for providers of general-purpose AI models came into effect on 2 August 2025. All GPAI models placed on the EU market are now subject to those requirements. DeepSeek's terms of service indicate that user data may be stored in China and used for training purposes — a direct conflict with GDPR obligations that govern how European companies handle user data. Italy's data protection authority, the Garante, has already launched a compliance probe into the companies behind DeepSeek. That probe is a signal, not an outlier.

Running Chinese hosted models through an external API compounds the risk. Self-hosting open-weight variants on European infrastructure is a different calculation — one that may satisfy data residency requirements — but it requires engineering capacity and ongoing maintenance that not every startup has.

European comparison / the gap

The EU leads on AI policy frameworks. It does not lead on AI model capability. The US and China are racing ahead in both AI innovation and investment while Europe builds regulatory architecture. No European AI model currently matches Chinese or US capabilities at scale.

The implementation of Chinese technologies across EU member states is uneven, according to the Council on Foreign Relations. There is no unified European position. Some member states are more exposed, some more cautious. The Economist has noted that in manufacturing, European firms hold advantages, and that using open models — including those from China — could extend that lead. The implication is pointed: blanket abstention from Chinese AI tools may cost European industrial firms a competitive edge they currently hold.

What European founders/investors should do

Three practical positions are worth distinguishing:

If you process EU user data through your product, Chinese-hosted APIs are a compliance liability until the legal picture clarifies. The Garante probe into DeepSeek illustrates the regulatory exposure. Don't wait for enforcement to find out where you stand.

If you are running internal tooling or agentic infrastructure, self-hosting open-weight Chinese models on EU-based infrastructure is a materially different risk profile. Evaluate it on its merits. The cost and performance case is real.

If you are building or investing in AI-enabled industrial or manufacturing applications, The Economist's framing is worth taking seriously. European firms have existing advantages in this sector. Tools that extend capability — regardless of origin — deserve evaluation on performance and compliance grounds, not reflexive national preference.

For investors: the uneven member-state implementation of Chinese technology creates arbitrage. Companies that build compliant, EU-infrastructure-based wrappers around open-weight Chinese models — capturing performance and cost advantages while managing data residency — have a structural opportunity. Watch for that wedge.

Closing line

The question European founders need to answer is not whether to engage with Chinese AI. It is whether they have the infrastructure, legal clarity, and engineering discipline to do it without handing away control — because the cost of abstaining is becoming as real as the risk of adopting.

Sources

Sources

Sources

Sources