Source Report 10

Enterprise AI Spend and Model API Revenue 2026

Full research prompt

What are current enterprise gen AI spend and foundation-model API revenue for 2026, versus Menlo Ventures' $37B and $12.5B for 2025?

From Is there an AI Bubble?

Jon Sinclair using Luminix AI
Jon Sinclair using Luminix AI Strategic Research

Menlo Ventures’ November 3, 2026 report (“2026: The State of Enterprise AI”) provides the primary recent benchmark, estimating enterprise AI deployment spending at an annualized $176.8 billion as of September 2026—roughly 5.4x the restated $33 billion for 2025 (or the original ~$37 billion figure from their prior report).[1][2][2]

This reflects a surge driven primarily by deeper usage among existing adopters (adoption rose only modestly from 55% to 65% of enterprises) rather than broad new entry, with agents accelerating consumption as token usage scales with work performed instead of seats or headcount. Enterprises overshot budgets significantly (actual spend averaged ~4.7x prior-year levels vs. a planned ~1.9x, or ~2.5x plan overall).[3][2]

Foundation-model and inference spending (the closest proxy for foundation-model API revenue) reached ~$94.2 billion in 2026, up 7.5x from $12.5 billion in 2025.[1][1]

This category dominates infrastructure spend ($99.5 billion total infrastructure out of $176.8 billion deployment), far outpacing the $12.5 billion 2025 API baseline referenced in the query. The 2025 total spend figure was restated downward to $33 billion in the new report due to reclassification of items like data labeling into development spending.[1]

Deployment spending breaks down as infrastructure (~$99.5B, heavily weighted toward frontier model inference) and applications (~$77.3B, with coding at $34B and AI assistants/productivity tools at $25B comprising 76% of apps).[4][2] Development/build spending (training compute + data/services) added another ~$102.5B.[4]

Frontier labs (Anthropic, OpenAI, Google) captured 69% of the $176.8B deployment total ($122.7B), up from 50% in 2025, with their first-party apps adding substantial share within applications. Within enterprise LLM API spend specifically, Anthropic rose to 55% (from 40%), OpenAI to 31% (from 27%), and Google fell to 9% (from 21%); open-source models held ~19% of tokens but only ~4% of spend.[3][2]

These figures are Menlo estimates derived from a survey of nearly 400 U.S. enterprise decision-makers (responsible for >$4B in AI spend) plus bottom-up modeling informed by public filings and third-party data; they represent annualized run rates as of September 2026 rather than full-year audited totals.[1][2]

Company-level revenues (e.g., Anthropic at a ~$67.8B run-rate or OpenAI at ~$50B as of late 2026) reflect broader (including non-enterprise) activity and are not direct substitutes for the enterprise deployment or API categories.[5][6]

For competitors or new entrants, the data highlights a market where capability (especially agentic/long-horizon performance) drives outsized spend concentration at the frontier, while new budgets fund the majority of growth rather than pure displacement of existing SaaS.[7][2]

  • Implication for positioning: Differentiate on verifiable agent workflows in high-spend categories like coding or productivity, where labs’ own apps already dominate but independent players can still capture meaningful share (e.g., the non-lab coding market grew 2.9x).[2]
  • Implication for pricing/monetization: Token-based/agent consumption outpaces seat-based models; focus on usage that scales with output rather than users. Open-source or lower-cost alternatives capture usage but far less revenue.[2]
  • Implication for go-to-market: Target power users and net-new budgets; incumbents retain most software-budget shifts, so pure displacement plays face headwinds.[7]

No other comprehensive, post-March 2026 independent estimates of these exact 2026 enterprise metrics appear in recent sources; all major coverage converges on the Menlo November 2026 report. Later full-year 2026 actuals or competing forecasts (if released after mid-November) could refine these run-rate figures.

Sources

Get Custom Research Like This

Start Your Research