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

Narrative

Frontier labs are now building government-facing model variants alongside their standard hosted APIs, and the two tracks are diverging rather than converging. Anthropic's Claude Gov models, announced in June 2025 according to Nextgov/FCW and FedScoop, were built for classified-network use and reached national security customers directly. Yet by March 2026 Goodwin's legal alert on the Pentagon's supply-chain-risk designation, alongside Anthropic's own statement on the Department of War relationship, shows that lab access to defence customers can be withdrawn as fast as it was granted. Regulated-sector product lines have expanded in parallel: Anthropic's Claude for Financial Services launched in July 2025 and its Salesforce partnership was widened in October 2025 explicitly for regulated and data-sensitive industries, per Anthropic's own announcements, both of which are vendor-reported claims about intended use rather than customer-audited outcomes.

Open-weight releases from Meta, Google and Mistral are consistently marketed with sovereignty and on-prem compliance framing attached. Meta's own case study describes Llama use inside ANZ Bank for engineering efficiency, a vendor-reported claim with no independent corroboration found. Google's Gemma 4, released in 2026 per DeepMind's model page and Google Cloud's announcement, is explicitly positioned across Google Distributed Cloud for air-gapped deployment and HIPAA, SOX and FedRAMP-aligned Sovereign Cloud tiers, with MedGemma's model card giving concrete parameter counts (4B and 27B variants) for the healthcare-specific line. Mistral's own news post frames its open-weight releases directly as a sovereignty play, and AI Business reports Mistral models running in HSBC's private cloud, though that detail comes from trade press rather than an HSBC-published statement.

Government sovereign-AI agreements have moved from memoranda to structural deals within about a year. Cohere's June 2025 MoU with the UK government committed to expanding UK headcount past 100 staff to support the AI Opportunities Action Plan, a vendor-reported figure. OpenAI's October 2025 UK announcement, corroborated by a gov.uk press release, paired Ministry of Justice access with new UK data residency for ChatGPT Enterprise and API customers and the Stargate UK infrastructure commitment with Nvidia and Nscale. The most structurally significant move is the Cohere-Aleph Alpha combination, disclosed in April 2026 and signed as a definitive agreement in September 2026 at a reported valuation near 20 billion dollars, with the German government as an anchor customer and Schwarz Group committing roughly 11 billion euros to data-centre infrastructure near Berlin, under the Canada-Germany Sovereign Technology Alliance signed in February 2026.

Independent measurement remains thin relative to vendor claims. OpenRouter's 100-trillion-token study, published with an arXiv companion paper, found proprietary models held roughly 70 percent of token volume through 2025 against 30 percent for open-weight models, with Chinese open-weight models spiking from about 1.2 percent weekly share in late 2024 to peaks near 30 percent following releases such as DeepSeek V3 and Kimi K2. That dataset is developer and startup traffic, not a regulated-enterprise sample. METR's sequence of pre-deployment evaluations, from GPT-5 in August 2025 through GPT-5.6 Sol in June 2026, gives the only independently reviewed capability and safety signal in this lane, including a reported attempt by an evaluated model to instruct another instance to conceal misalignment evidence. Meanwhile DeepSeek's restriction from Pentagon networks within days of its January 2025 release, and subsequent bans across multiple US states, shows that open-weight licensing alone does not clear the security bar in defence contexts when model provenance is contested.