State of Enterprise AI 2026 — the run-cost reckoning
240 enterprises. AI ROI, run-cost benchmarks, and the bills nobody saw coming.
- 01 Median enterprise inference spend grew 4.1× year-over-year in 2025.
- 02 38% of enterprises have a formal model risk function; only 11% have one for GenAI specifically.
- 03 Top-quartile programmes ship in 90 days; bottom quartile take three quarters or are cancelled.
- 04 Sovereign deployment is a procurement requirement in 7 of 11 surveyed jurisdictions.
Methodology
Survey of 240 enterprises across BFS (n=82), government (n=54), healthcare (n=38), retail (n=30), telecom (n=22), and energy/utilities (n=14). All respondents were CIO/CTO/CDO grade with budget authority over AI spend.
Run-cost reckoning
Inference cost is now the second-largest line item in enterprise AI budgets, after talent. The median programme spent 41% of its 2025 budget on inference — up from 14% the previous year. We expect the share to peak in 2026 as routing and caching disciplines mature.
Time-to-production
The gap between top and bottom quartile is widening. Top quartile programmes ship in under 90 days because they pre-built the platform; bottom quartile programmes are stuck in capability acquisition. The differentiator is operational maturity, not model choice.
Governance state of play
38% of enterprises have a formal model risk function. Only 11% have one specifically tuned for GenAI. The gap will close in 2026 as regulators move from guidance to enforcement.
Recommendations
Three priorities for 2026: (1) build an inference-cost dashboard before the bill becomes board-level, (2) extend model risk management to cover GenAI, (3) bias toward shipping over assessing.
Related reports.
Get the full report.
PDF available to enterprise subscribers. We will route the request to the right partner.