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

Narrative

The dominant quantitative source for this lane, Menlo Ventures' survey of nearly 500 US enterprise decision-makers, found open-source and open-weight models' share of enterprise LLM workloads falling from 19% to roughly 11-13% between 2024 and 2025, even as total enterprise generative AI spend tripled to 37 billion dollars. That single data point undercuts the 2023-2024 venture narrative, visible in early a16z and Sequoia commentary, that open weights would commoditise the model layer. A16z's own 2025 CIO survey complicates the picture further: open-weight adoption of Llama and Mistral concentrates at the largest enterprises, driven by on-prem data-security and fine-tuning requirements rather than cost, suggesting the aggregate decline masks a bifurcated market rather than a uniform retreat.

Regulator-published data gives the most reliable baseline for financial services specifically. The Bank of England and FCA's third joint survey, published November 2024, found 75% of UK financial firms already using AI, foundation models accounting for 17% of use cases, and only 2% of automated decision-making rated fully autonomous. Four of the top five perceived risks were data-related, and firms cited data protection and outsourcing rules, not model choice, as the largest regulatory constraint. No comparable fourth-wave survey had published by September 2026, leaving 2024 as the most recent regulator-measured snapshot for UK banking.

On sovereign providers, the clearest measured event is France's Ministry of the Armed Forces framework agreement with Mistral, notified 16 December 2025 and steered by the AMIAD defence AI agency, which specifies deployment on French-controlled infrastructure and fine-tuning on defence-specific data, a rare instance of sovereignty as a dated, scoped procurement criterion rather than a marketing claim. Cohere's UK arrangement, by contrast, remains a non-binding memorandum of understanding signed around June 2025, corroborated by independent trade press (BetaKit) but with no disclosed contract value. The April 2026 Cohere-Aleph Alpha combination, reportedly valued near 20 billion dollars and backed partly by 600 million dollars from Schwarz Gruppe, drew active sponsorship from Germany's and Canada's digital ministers, evidence that at least two governments now treat sovereign-AI consolidation as a strategic asset rather than a private commercial matter.

The routing-layer evidence is the most likely to mislead if read uncritically. OpenRouter's own published token data, and an independent arXiv analysis of it, show Chinese open-weight models growing from under 2% to a peak of 46% of weekly token share by April 2026, with open-weight models overall carrying most usage but capturing only about 4% of model-layer revenue. Practitioner analysis from Interconnects frames this correctly as a developer-cost story, open models run roughly six times cheaper per call at near-parity capability, rather than an enterprise-procurement signal, since OpenRouter's traffic is price-sensitive individual-developer usage and is not representative of regulated-enterprise deployment decisions.