European enterprises favor open weights
- AI.cc said on May 9 its 2026 infrastructure report shows European enterprise workloads now lean heavily toward deployable open-weight model stacks over single APIs. - In Europe, Llama 4, Gemma 4, Mistral Small 4, and GLM-5.1 made up 61% of enterprise token volume, while blended token costs fell 67%. - That matters because Europe is pairing cheaper inference with sovereignty rules, pushing buyers toward multi-model, self-hostable AI procurement.
Enterprise AI buyers in Europe are starting to act less like SaaS customers and more like infrastructure teams. That is the real story here. The flashy part is model names and token prices, but the deeper shift is architectural — companies want models they can route, swap, fine-tune, and sometimes run themselves. A new AI.cc infrastructure report, released May 9 and based on 2.4 billion API calls from January through April 2026, says that shift is now visible in production usage, not just conference talk. ### What actually changed? The report’s broad headline is global — open-source models reached 38% of enterprise token volume for the first time on AI.cc’s platform, and multi-model routing has crossed from experiment to default setup. But Europe stands out even more. There, deployable open-weight families — specifically Llama 4, Gemma 4, Mistral Small 4, and GLM-5.1 — accounted for 61% of enterprise token volume. That is not “some teams are testing local models.” That is mainstream production traffic. (natlawreview.com) ### Why do open weights matter so much? Because “open” in enterprise procurement does not just mean ideology. It means control. If a company has model weights, or at least a deployable version it can run in a controlled environment, it gets more leverage over data handling, latency, vendor lock-in, and cost. Gemma 4 is a good example — Google released it under Apache 2.0 and pitched it explicitly as a model family developers can run on their own hardware, from edge devices up to accelerators. (natlawreview.com) Hugging Face framed the same launch around broad deployability across local tooling and inference stacks. ### Why is Europe leaning harder into this? Basically — sovereignty. Europe has spent the last two years turning “digital sovereignty” from a policy slogan into procurement logic. The European Commission has been openly arguing that open-source AI can strengthen competitiveness and sovereignty, while also building out AI Factories and a €20 billion InvestAI facility aimed at European AI capacity. In parallel, the EU’s guidance around general-purpose AI models has made compliance a live design constraint, not a future maybe. (blog.google) If you are a European enterprise, the appeal of a model you can inspect, host, or at least move between providers is obvious. ### Is this just about regulation? No — cost is doing a lot of the work. AI.cc says enterprise token costs on its platform fell 67% year over year, from a blended $18.40 per million tokens to much lower levels by April 30, 2026. Teams using multi-model routing saw median savings of 71% versus single-provider setups, with the top quartile above 80%. That lines up with the broader market direction — inference prices have been collapsing fast across tasks, sometimes by orders of magnitude per year at a fixed capability level. (digital-strategy.ec.europa.eu) ### Why does multi-model routing matter here? Because cheaper models alone do not explain the shift. Routing does. Instead of sending every task to one premium API, companies are breaking workloads apart — simple classification to one model, extraction to another, harder reasoning to a third. Deloitte has been making the same point from the finance side: enterprises are moving toward hybrid mixes of APIs, SaaS, and self-hosted infrastructure, and token spend now needs active governance. (natlawreview.com) Europe’s open-weight tilt fits that pattern almost perfectly. ### Are closed models losing? Not exactly. The best closed models still matter for frontier reasoning, coding, and agentic work. But the center of gravity is moving. More of the average enterprise workload is becoming “good enough plus controllable” rather than “absolute best model at any price.” Once that happens, open weights get a structural advantage — not because they win every benchmark, but because they fit the stack buyers are trying to build. (deloitte.com) ### What is the bottom line? Europe is not just buying AI anymore. It is choosing how AI should be owned, routed, and governed inside the company. The new report suggests open-weight models are becoming the default answer for a big chunk of that question — especially where cost discipline and sovereignty sit next to raw capability. (natlawreview.com)
