Comparative LLM Usage Across Sectors

Narrative

Enterprise spending on large language models roughly doubled in the first half of 2025, rising from $3.5 billion at end-2024 to $8.4 billion by mid-year, according to Menlo Ventures' survey of 150 technical leaders. Market concentration remained high: Anthropic held 32% enterprise share, OpenAI fell to 25% from a 50% peak in 2023, and Google's Gemini claimed 20%. Only 11% of enterprises switched model providers, while 66% upgraded within their existing vendor family. Closed-source models powered 87% of enterprise workloads at mid-2025, with open-source usage declining from 19% to 13% over six months as performance gaps widened.

Bloomberg Intelligence's June 2026 Generative AI Outlook raised its market forecast to $2.3 trillion by 2032, a $500 billion upward revision from its March 2025 estimate, driven by accelerating token consumption, the rapid expansion of coding and customer service agents, and a faster-than-anticipated shift from training to inference. The report noted hyperscaler capital expenditure approaching $750 billion in 2026 alone. Bloomberg also introduced its own agentic AI tool, ASKB, to its Terminal in early 2026, signalling that financial data providers are embedding multi-step AI agents directly into professional workflows rather than treating LLMs as peripheral add-ons.

The cost picture is structurally paradoxical. Blended enterprise token prices fell 67% year-on-year from Q1 2025 to Q1 2026, yet total AI bills rose substantially: Ramp's transaction data showed average monthly AI token spend across its customers up 13x since January 2025 and total token usage up more than 1,000% between January 2025 and April 2026. The FinOps Foundation's 2026 State of FinOps survey, drawn from practitioners responsible for over $83 billion in cloud spend, found that 98% now manage AI costs, up from 63% in 2025 and 31% in 2024. The shift from seat-based SaaS pricing to volatile per-token and per-agent-step billing has broken traditional forecasting tools, creating a new discipline of real-time AI cost governance. The median business on Ramp used nine models in April 2026; the average used 16.5.

DeepSeek's January 2025 R1 release disrupted assumptions about cost and openness. The model, reportedly developed for around $6 million, achieved performance comparable to leading US frontier models and was released under an MIT licence. Amazon confirmed thousands of enterprise customers deployed DeepSeek-R1 through Bedrock within days of its availability. Despite geopolitical concerns, 80% of surveyed engineers said they would consider using DeepSeek, according to Kong's enterprise survey. However, the broader trend ran against open-source: enterprises cited compliance, support, and total cost of ownership as reasons to favour closed providers, and open-weight model share of production workloads continued to contract through mid-2025.