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

Narrative

Independent commentary treats government sovereignty claims with more scepticism than trade coverage does. Rodger Cuddington's Substack post "Sovereign in name only" sets the UK's £500 million Sovereign AI Fund, launched in April 2026, against France's €109 billion of committed AI investment announced at the Paris AI Action Summit in February 2025 and anchored by Mistral's €1.7 billion Series C round that September, concluding the UK programme is roughly a hundredth the scale of what it is compared against. Hamish Low's Cambrian Research newsletter pushes the argument further, contending that compute access and supply chains, not model weights or contractual sovereignty language, are the actual constraint on any government's AI independence. Coverage of Cohere's April 2026 acquisition of Aleph Alpha, read on the hightechinvesting Substack as "the quiet sale of Germany's AI hope," reinforces the same scepticism: Europe's national sovereign-model champions are consolidating into transatlantic vendors rather than staying nationally rooted, a pattern corroborated by trade press including CIO and Computerworld reporting the same transaction.

On deployment share, independent bloggers are actively contesting each other's numbers rather than converging on one. OpenRouter's State of AI report, an empirical analysis of 100 trillion tokens on its own platform, found open-weight models approaching a third of token volume by late 2025, with Chinese open-weight models around 30 percent of weekly volume. Mozilla.ai's blog and the independent Digital Applied blog both push back on treating that as representative, arguing OpenRouter's traffic is self-selected developer usage that undercounts locally run small models and says nothing about regulated-enterprise behaviour. That caution is sharpened by Menlo Ventures' December 2025 enterprise survey, which found open-source share of enterprise LLM API spend falling from 19 percent in 2024 to 11 percent in 2025 among the enterprises it surveyed, the opposite direction from OpenRouter's platform trend. The independent analyst site luizneto.ai treats this divergence as evidence that licensing terms and support costs, not raw model capability, now decide enterprise adoption.

Ben Thompson's Stratechery DeepSeek FAQ, published in January 2025, remains the reference point independent writers cite when arguing that open-weight releases diffuse capability regardless of export-control regimes; a LessWrong post making the case "against export controls" reaches a similar conclusion from a policy-analysis angle. Applied to regulated sectors, an aiforhealthcare Substack post on NHS-LLM, built on Meta's 13-billion-parameter Llama weights via the OpenGPT instruction-tuning framework, and an aetosky Substack piece on air-gapped sovereign LLMs paired with smaller edge-deployed models in the public sector, both describe on-prem, weights-in-hand deployment as the default once data cannot leave a controlled perimeter. In both cases the parameter counts used are deliberately modest, not frontier-scale, sized to hardware regulated buyers can actually field rather than to benchmark leaderboards.

Across this lane, the clearest fault line is measurement, not architecture. OpenRouter's dataset and Menlo's survey point in opposite directions on whether open-weight adoption is rising or falling, and neither is dispositive for financial services, defence or healthcare specifically since neither isolates those sectors. Independent bloggers converge more on the sovereignty question: government agreements are consistently characterised as directional signalling rather than measurable infrastructure shifts, with Cuddington's funding-scale comparison and the Aleph Alpha sale both cited as evidence that European sovereign-AI positioning has outpaced its financing.