A Reddit user pointed out something with strategic weight: on Openrouter, the platform that aggregates inference providers for LLMs, the Chinese model GLM 5.2 is served by sixteen suppliers. The list includes z.ai, Wafer, NovitaAI, Ambient, Together, Cloudflare, Fireworks, Friendli, Parasail, AtlasCloud, StreamLake, io.net, DeepInfra, Morph, Phala and SiliconFlow. All operate from the United States or, in two cases, Singapore and China. None from Europe.

The user’s question is blunt: are there any European inference providers for open-weight models, especially Chinese ones like GLM 5.2 and DeepSeek V4 Flash? Looking at Openrouter, the answer is no. And the issue goes far beyond hobbyist curiosity.

Europe’s absence in the inference market

Openrouter works as a marketplace: it aggregates dozens of inference services, letting developers pick providers based on price, latency and availability. For Western models, European presence is not a problem — think Mistral or AWS/GCP instances in EU regions. But for Chinese open-weight models, the gap is almost total. The providers running them — Cloudflare, Together, Fireworks and the rest — rely mainly on US or Asian infrastructure.

This is not a technical footnote but a market signal. Chinese models are gaining ground in capability and cost: GLM 5.2 and DeepSeek V4 Flash deliver competitive performance in code generation and reasoning. Yet their adoption in Europe hits an infrastructural wall: there is no local inference offering.

Digital sovereignty and open questions

For a European company wanting to integrate these models into applications handling personal or sensitive data, the lack of EU-based providers forces a stark choice: send data to servers outside the EU, with all the GDPR uncertainties that entails. Even when the vendor claims compliance, the physical residency of data carries growing weight in risk assessment, particularly in regulated sectors like finance, healthcare and public administration.

Latency is another factor. Inference on American or Asian servers introduces delays that can make real-time applications impractical. Moreover, dependence on third-party providers exposes companies to price swings, unilateral changes in terms of service and lock-in risks.

Self-hosting: rethinking infrastructure

Faced with this landscape, the alternative is on-premise deployment or private cloud in European data centers. Running GLM 5.2 or DeepSeek V4 Flash on one’s own hardware — or on dedicated instances in EU regions — restores control over data and performance. It is not a simple path: it requires investment in GPUs, quantization management and serving orchestration. But it is a route many organizations are already taking, driven precisely by the need to combine innovation with sovereignty.

AI-RADAR regularly explores the trade-offs of such choices: from TCO to fine-tuning pipelines to evaluating inference frameworks. The disappearance of European providers for Chinese models is a warning bell: demand for open-weight models is growing, but local service supply is not keeping pace. Anyone wanting to use them without sacrificing control will likely need to get their hands dirty with hardware.

A gap to fill

The void highlighted on Openrouter is also an opportunity. European cloud providers could differentiate themselves by offering inference for Chinese open-weight models, perhaps in partnership with the labs that release them. Until that happens, self-hosting remains the main road for those unwilling to compromise on data residency and latency.