OpenRouter has moved US in-region routing into general availability, guaranteeing that requests sent to its US endpoint are decrypted, processed and served entirely inside the country — and rejected outright if no compliant provider can take them. The marketplace published the change in an announcement post, extending a guarantee it has offered in the European Union since October 2025.
The timing is not incidental. OpenRouter says the share of tokens going to open-weight models has risen steadily for requests originating in both the US and the EU, and that models from Chinese labs still make up most of that volume. Enterprises want the price and speed of those weights. Their procurement teams want to know where the prompts go.
Key takeaways
- US In-Region Routing is generally available on Business and Enterprise plans through the
us.openrouter.aiendpoint, which fails closed rather than falling back to an out-of-region provider. - The open-weight share of OpenRouter's US token volume has risen steadily, with Chinese labs including DeepSeek, Moonshot and Zhipu supplying most of it.
- Eligible models span OpenAI, Anthropic, Google, NVIDIA and Thinking Machines alongside DeepSeek, Moonshot and Zhipu, because US providers such as Baseten, Fireworks and Azure host them domestically.
How the routing guarantee works
The change happens at the routing layer rather than in the models. On OpenRouter's standard global endpoint, any eligible provider in any region can serve a request, which means even calling a US-developed model does not guarantee the request itself is handled on US soil. Requests sent to the regional endpoint are decrypted on OpenRouter infrastructure inside the country, and the pool of candidate providers is filtered down to endpoints the company has approved as operating there.
If no compliant US provider can serve the requested model, the call fails with a 404 rather than silently routing abroad. Teams can pin the restriction through OpenRouter's guardrails at the workspace, team or API key level, and tools that would push prompt data outside the region are disabled on the regional endpoint. The feature documentation lists the models currently eligible.
Why the open-weight share matters
The pitch for open-weight models has always included control: download the weights, run them where you choose, keep the data close. That argument weakens for the large share of developers who reach those models through a hosted marketplace instead of self-hosting them.
According to The New Stack, Hugging Face data from February showed Chinese developers accounted for 41% of model downloads over the preceding 12 months, ahead of US developers at 36.5%. The same report cites Deloitte's 2026 State of AI in the Enterprise, which found 77% of companies now factor country of origin into vendor selection and nearly 60% build their AI stacks primarily with local vendors. Nvidia's $12.9 billion move for Hugging Face underlines how much commercial weight the open ecosystem now carries.
OpenRouter product team member Cailee Moberg framed the problem in procurement terms, noting that models from Chinese labs still make up most open-weight volume and that approval for those models can be difficult to obtain. The company's argument is that in-region routing lets teams with residency requirements capture the price and performance gains from those models without involving the originating lab at all: when a US or EU provider hosts the weights, the request goes to that provider.
What it does not change
Residency routing governs where inference happens, not where a model was built. DeepSeek V4 Pro, Kimi K3 and GLM 5.2 remain Chinese-developed models; what changes is which hosted copies a US customer can be routed to, and where the prompt is handled along the way. Companies could already achieve the same outcome by self-hosting or contracting a US provider directly. OpenRouter, which Stripe has reportedly agreed to acquire in a deal valued at about $7.5 billion, is betting that most teams would rather buy the guarantee than operate it.
FAQ
What happens if a model is not available in the US region?
The request fails closed. OpenRouter returns a 404 error instead of falling back to a provider outside the region, so a misconfigured call cannot quietly leak prompt data abroad. Administrators can additionally enforce the restriction through Guardrails at the workspace, team or API-key level.
Which models support US In-Region Routing?
Eligible models include proprietary frontier systems from OpenAI, Anthropic and Google, US open-weight models from NVIDIA and Thinking Machines, and Chinese open-weight models such as DeepSeek's V4 Pro, Moonshot's Kimi K3 and Zhipu's GLM 5.2. They qualify because US providers including Baseten, Fireworks and Azure serve them from domestic data centers. OpenRouter maintains a live eligibility list.
Is EU in-region routing also available?
Yes. OpenRouter says European in-region routing has been available since October 2025, and the US launch extends the same guarantee to a second region. Both are offered to Business and Enterprise customers.
