Comparative LLM Usage Across Sectors

Narrative

The clearest quantitative window into enterprise LLM adoption comes from Menlo Ventures' paired 2025 reports. The mid-year report recorded enterprise LLM API spend doubling from $3.5 billion to $8.4 billion in a single six-month stretch, driven by inference overtaking training as the dominant cost category. By year-end, Anthropic had captured 40% of the enterprise LLM API market, up from 12% in 2023, with OpenAI falling to 27% and Google reaching 21%. Code generation was the catalyst: Claude held 42% developer share in that workload, double OpenAI's figure. These are survey-derived numbers from roughly 500 US enterprise decision-makers, and Menlo is an investor in Anthropic, which warrants scepticism about precision, but the directional claim of a rapid share shift is corroborated by the OpenRouter empirical dataset.

OpenRouter's 100 trillion-token study, published in partnership with a16z in December 2025, offers the most granular observed-behaviour dataset available to independent analysts. The platform, serving over five million developers across 300-plus models, recorded proprietary models handling the majority of tokens throughout the study period, but open-weight models grew to roughly one-third of usage by late 2025. Within the open-weight segment, DeepSeek alone accumulated 14.37 trillion tokens, followed by Qwen at 5.59 trillion and Meta Llama at 3.96 trillion. The market fragmented sharply after what the study labels the "Summer Inflection": DeepSeek's near-monopoly of over 50% of open-source share in early 2025 collapsed as Qwen 3, Kimi K2, and the GPT-OSS variants entered the field. Tool use and agentic workloads showed a parallel shift, with reasoning-model tokens climbing from a negligible slice in Q1 2025 to over 50% of total usage by mid-year.

Self-hosted inference emerged as a mainstream business strategy rather than a researcher pastime. Practitioners broadly distinguish two deployment paths: Ollama for prototyping and single-developer use, and vLLM for production multi-user workloads, with the latter's PagedAttention mechanism cited as delivering substantially higher throughput at concurrent load. The driver in regulated sectors - healthcare, legal, finance, government - is compliance. HIPAA, GDPR, and SOC 2 create hard boundaries around data egress that push organisations toward on-premise or private-cloud deployment of open-weight models, primarily Llama and Mistral variants. By contrast, Menlo's enterprise survey found open-source adoption actually declining from 19% to 11% year-on-year among large enterprises, suggesting that the self-hosting trend is more pronounced in mid-market and developer-led organisations than in top-tier procurement-driven accounts.

The production-deployment gap remains the defining practical problem. The Metadata Weekly Substack cited an MIT figure that 95% of enterprise AI pilots never reached production in 2025, while ZenML's LLMOps Database, which crossed 1,200 catalogued case studies by December 2025, found that successful systems consistently combined LLMs with traditional deterministic rules and classical ML rather than delegating entirely to foundation models. On cost, practitioners writing on Medium and specialist FinOps blogs in 2025 and 2026 document a structural measurement problem: LLM bills arrive as undifferentiated totals that hide per-feature and per-user attribution. The practitioner consensus, drawn from multiple independent sources, treats LLM cost as an architectural discipline imposed at design time rather than a billing line read at month-end. Token prices dropped approximately 80% between early 2025 and early 2026, yet total enterprise spend tripled, illustrating classic Jevons-paradox dynamics that independent commentators noted explicitly.