Langdock moves its parent company from the US to Germany
A Berlin AI platform spends millions to un-Americanise its corporate structure, and a Chinese bank takes first place on the LLM routing leaderboard.
Langdock, the Berlin-based enterprise AI platform used by roughly 13,000 organisations, has moved its legal parent company from the US back to Germany, converting into a Societas Europaea — a reversal of the transatlantic holding-company default European startups have followed for a decade. The company said the process began in early 2026, cost several million euros, and is now complete; its operations and customer data always stayed in Germany, and the US parent had no employees, infrastructure, or access to production systems. That last detail is the point. Customers' legal teams still had to assess whether an American parent created exposure under the Cloud Act, so a company with zero US operations was generating US legal risk anyway — and the fix was a structural one, not a contractual promise.
The numbers attached to the move are the ones that make it credible: Langdock says its annual subscription revenue run rate hit $50 million in August, up from $1 million in October 2024, and that about 80% of the company is owned by founders and employees living in the EU. It now plans three new services before the end of the year and will run open-source models on its own data centre in Germany, with the stated ambition of becoming a "sovereign, full-stack AI platform" that can compete with US hyperscalers over time. For scale, $50 million of run rate sits against the $128.7 billion AWS reported for 2025 — this is a bet on where demand goes, not a claim about where it is. Our read: European AI sovereignty stopped being a procurement subsidy story this year and became a corporate-structure story. The question buyers ask is no longer "where does your compute sit," it is "who can compel you." We covered the sovereign pitch when the money was still going into GPUs — Mistral raises €3B in Europe's largest tech round to push sovereign AI.
Ping An Bank's self-developed model router, Paix2, is holding first place on RouterArena — an open LLM-routing leaderboard built by a Rice University team — with a weighted Arena score of 77.6, ahead of KT's ModelRouter at 76.3 and Microsoft's Azure-Router at 70.4. The bank's number stayed on top for nearly a month, per Yicai, and the sub-scores explain why it is not a rounding artifact: 79.7 accuracy, $0.27 per 1,000 queries, an optimal-model selection score of 89.7, and a perfect optimal-accuracy score. Paix2 is routing between four candidates — MiniMax-M3, Agnes 2.0 Flash, DeepSeek-R1-0528-Qwen3-8B and GLM-4-9B-0414 — and has been deployed inside the bank's digital employee assistant, with further internal rollouts planned.
Read the leaderboard carefully and a second story appears. Berkeley's RouteLLM scores 99.7 on optimal selection — nearly perfect at naming the cheapest correct model — yet lands at 48.1 on the Arena score, because the metric is a weighted harmonic mean that trades accuracy against cost (91% accuracy weight, 9% cost weight). Routing is not one problem. There is the easy half, "is this query cheap or hard," and the hard half, "is the cheap answer actually correct." A commercial bank sitting on top of both is the tell: the routing layer gets optimized hardest by whoever pays the largest inference bill, and no one pays a larger one than a bank running the same questions a million times a day. We have tracked this layer as it went from startup pitch to shipped product — Ramp launches its own AI model router and Nvidia ships Nemotron 3.5 Lightning and an open model router.
What to watch: whether Paix2's routing code gets published the way its leaderboard entry implies, and whether other European AI companies follow Langdock into European holding structures.
Does sovereignty sell better as a corporate structure than as a compute subsidy? Tell us in the comments.
