On-prem and open-weight models in regulated, high-security enterprises

Narrative

The clearest regulator-published baseline for financial services remains the Bank of England and FCA's third Artificial Intelligence and Machine Learning survey, covering 118 banks, insurers and financial market infrastructure providers and published in November 2024. It found 75% of firms already using AI, with foundation models making up 17% of use cases and 55% of use cases involving some autonomous decision-making, though only 2% were described as fully autonomous. That contrasts with Menlo Ventures' December 2025 cross-industry enterprise survey, which found open-source model share falling to 11% from 19% even as total enterprise generative AI spend roughly tripled to $37bn, a reminder that regulated-sector behaviour and general enterprise behaviour are moving in different directions and that VC-authored surveys are analyst-estimated rather than independently audited.

OpenRouter's own published usage data shows US proprietary models' share of routed tokens falling from roughly 70% in June 2025 to about 30% by June 2026, with Chinese open-weight families such as DeepSeek, Tencent, Xiaomi and Minimax reported by Dataconomy to have reached around 61% share by February 2026. An arXiv analysis of over 100 trillion OpenRouter tokens corroborates the direction of travel. This data is real and independently measured at the platform level, but coding workloads grew from 11% to over half of all OpenRouter tokens through 2025, so the figures describe a developer-tooling market, not a proxy for how banks, insurers or defence ministries actually route regulated workloads.

Sovereign-provider activity accelerated sharply through 2026. Microsoft's multibillion-dollar July 2026 agreement with Mistral, reported by SiliconANGLE and France 24, expands French GPU capacity and spans Azure public cloud, customer-controlled Azure Local and fully disconnected deployments, though Microsoft confirmed no new equity stake and declined to disclose the deal's dollar value. Cohere's April 2026 acquisition of Germany's Aleph Alpha, reported at a combined $20bn valuation with Schwarz Group committing $600m, created a Toronto-Berlin dual-headquartered group explicitly pitched at finance, defence and healthcare customers. In the UK, Cohere's memorandum of understanding was signed with the Department for Science, Innovation and Technology, a department dissolved in July 2026 when Prime Minister Andy Burnham created a cabinet-level AI minister, per Bloomberg. The UK's separate £500m Sovereign AI Fund, launched in April 2026 per The Register and GOV.UK, made its first disbursements to UK-based firms rather than to Cohere or Mistral, while a parliamentary committee warned the government still lacks a coherent sovereign AI strategy.

In high-security environments, the clearest measured discontinuity was the US Commerce Department's June 2026 export-control order forcing Anthropic to disable its Fable 5 and Mythos 5 models for non-US users over a jailbreak vulnerability, an action Fortune reported as the strongest single catalyst for European government interest in sovereign alternatives. The Pentagon followed in May 2026 with classified-network AI agreements covering eight companies including OpenAI, Google, Microsoft and NVIDIA while explicitly excluding Anthropic, after Anthropic declined to guarantee unconditional military use. On the build-versus-buy side, JPMorgan's in-house LLM Suite, developed rather than bought partly for data-control reasons, reached roughly 200,000 users within eight months of its 2024 launch, still the largest publicly documented proprietary financial-sector deployment.