Menlo Ventures' fourth annual enterprise AI report puts 2026 deployment spend at $176.8B, up 5.4x in a year, with agents consuming 80% of tokens on OpenRouter. The headline is the least useful part. The useful parts are the budget-overrun data, the cache economics, and the three conditions that decide whether an agent project ships.
Last Updated: November 9, 2026
Menlo Ventures' fourth annual State of Enterprise AI estimates enterprise AI deployment spend at $176.8B as of September 2026, up 5.4x from $33B, with agents driving the inflection. The headline multiple is a VC's bottoms-up model layered on a restated base year — treat it as directional. The operationally useful findings are buried further down: enterprises spent an average of 4.7x last year against a planned 1.9x, the median came in at only 2.2x, and cache reads now make up 95%+ of an agent's tokens. If you are budgeting agent work, those three numbers matter more than $177B.
Menlo Ventures published its 2026 enterprise AI report on November 3, and it has been syndicated into every AI newsletter since. The number everyone repeated: enterprise AI deployment spending grew from $33 billion to $176.8 billion in a year, a 5.4x increase against 3.9x the year prior. Growth accelerating, not decaying.
That number will show up in your next board deck and in three vendor pitches this week. So it is worth being precise about what it is, what it is not, and which parts of the underlying report actually change how you build.
The budget overrun is the real story
Buried in Part I is the finding that should worry anyone who owns an AI line item. Menlo reports that enterprises planned to roughly double AI spend in 2026 and instead landed at an average of 4.7x the prior year — about 2.5x what they budgeted. The explanation is structural, not sloppiness: agents run in parallel, keep working after the human closes the laptop, and therefore scale consumption with work performed rather than with seat count. Spend per enterprise nearly quintupled while the share of enterprises adopting AI moved only from 55% to 65%.
But notice the gap between mean and median. Median spend came in at 2.2x against a 1.8x plan. Menlo attributes the spread to a "fat head" of power users. For a normal enterprise, that means the honest planning anchor is somewhere near 2.2x, with a variance tail that can take you to 4.7x if one team ships an autonomous workflow that runs continuously.
| Measure | Planned 2026 | Actual 2026 | What it implies for your plan |
|---|---|---|---|
| Average enterprise | 1.9x | 4.7x | Skewed by heavy agent adopters; do not use as a baseline |
| Median enterprise | 1.8x | 2.2x | The realistic planning anchor for most organisations |
| Market-wide deployment spend | — | 5.4x ($33B → $176.8B) | Modelled estimate, not an audited aggregate |
The practical consequence: seat-based forecasting for agentic workloads is dead. If your FinOps model still multiplies licences by users, it will miss by a multiple, not a percentage. Budget on cost-per-completed-task with a hard ceiling per workflow, and alert on tokens-per-task drift rather than monthly totals.
Cache economics now dominate your unit cost
Why agent tokens are cheap — until they aren't
Menlo reports that cache reads can be 95% or more of an agent's tokens, and that Opus 5.5 prices cache reads at $0.20 per million against $4 for fresh input and $20 for output. At a mix of roughly 3% fresh input and 2% output, a session blends out to about $0.70 per million tokens. The model you pick matters far less than whether your prompt prefix is stable enough to hit cache.
This is the single most actionable number in the report for engineering teams. A workflow that rebuilds its system prompt, reorders tool definitions, or injects a timestamp near the top of context destroys the cache prefix and moves those tokens from $0.20 to $4 per million. On a high-volume agent, that is the difference between a defensible unit cost and an incident review.
It also explains why blended token prices keep falling even as buyers trade up the model ladder — Menlo notes the most popular models shifted from Claude Sonnet 4.5 and GPT-5 a year ago to Claude Opus 5.5 and GPT-6 Astra today. Teams are buying more expensive intelligence and paying less per token, because the mix shifted toward cached agent traffic. Anyone citing falling token prices as evidence that AI is getting cheap is reading a composition effect.
The three conditions are a qualification checklist
Menlo's explanation for why coding ($34.0B) and enterprise chatbots ($25.0B) make up 76% of the $77.3B application market is the most transferable framework in the report. Agents concentrate where three conditions hold: an environment they can act in, complete context they can learn from, and results they can verify. Coding scores on all three — compilers and tests verify outcomes for free.
The inverse holds too. Human resources grew 2.5x to $0.9B and marketing just 1.8x to $1.2B, categories where success is subjective and verification is expensive. Customer success, where playbooks document steps and resolutions define success, grew 4.3x to $2.7B.
The agent-fit test, before you scope anything
Are the systems of record reachable through APIs or connectors with write scope, or is the agent reduced to generating text a human re-keys?
Both the tacit knowledge (intent, local conventions) and the procedural knowledge (how the work is actually done) need to exist in retrievable form. If it only lives in people's heads, you are funding a documentation project first.
A deterministic check — a test, a reconciliation, a schema validation. Without one, you cannot build an eval, and without an eval you cannot safely remove the human.
That is the integration work — and it is why the app-layer opportunity Menlo describes is mostly plumbing, not prompting.
Run every candidate workflow through this before costing it. We use a version of this in automation scoping, and it kills more projects at the whiteboard than anything else — which is the point.
Your infrastructure assumptions are what break
Menlo reports that 74% of builders now use agents to manage workflows and execute tasks, up from 27% of startup deployments in 2025, with 32% already running multiple agents together. Adoption of the supporting primitives moved just as fast: code sandboxes from 11% to 40% of builders, search from 5% to 30%, browser infrastructure from 4% to 14%. Databases are shifting from long-lived provisioned systems to ephemeral ones that agents spin up, use and discard.
Translate that into your environment and the list of things that break is specific: service-account sprawl when every agent needs its own identity; audit trails that assume a human actor per action; network egress policy that never contemplated a browser-driving agent; and database provisioning workflows with a ticket-based lead time measured in days against an agent that needs a schema for ninety seconds. None of this is an AI problem. It is an integration and governance problem that agent adoption surfaces all at once.
What the report is carefully not saying
Menlo is a venture firm, and the report is transparent about it — portfolio companies are asterisked throughout, including Anthropic, OpenRouter, Fireworks, Modal, Mercor, Cursor, Factory, Legora and Lovable. The evidence chain runs through assets the author owns. That does not make it wrong; it makes it a prospectus-adjacent document.
Far from being a zero-sum game, AI is creating new value across the economy, expanding markets far more than it’s displacing existing revenue.
— Menlo Ventures, 2026: The State of Enterprise AI
Four things to hold in mind when someone quotes the headline at you:
The base year was restated. Menlo's methodology notes that prior-year figures were reclassified under this year's taxonomy, moving data-labelling vendors like Mercor from deployment to development. So 5.4x is an estimate divided by a revised estimate.
The 80% agent figure is one channel. It comes from OpenRouter token volumes, and Menlo states plainly that its users "skew more toward individual developers and startups than the enterprise buyers" in the survey. Directionally convincing, not an enterprise measurement.
Spend is an input, not an outcome. The report sizes dollars, not realised value. There is no churn, retention or ROI data here. The $1.70 of deployment revenue per $1 of training spend is a ratio of two modelled aggregates, not anyone's unit economics.
The survey is real but narrow. Roughly 400 US enterprise decision-makers responsible for more than $4B in AI spending, fielded by Morning Consult over three weeks in September 2026. Solid for a market report; not a census.
Where it does cross-check well: Menlo puts Anthropic at 55% of enterprise LLM API spend, and Ramp's AI Index independently showed Anthropic at 60.5% of business AI token spending in the week ending October 23, 2026, with US business AI adoption at 55.9% in September. Two different methodologies pointing the same way is worth more than either alone.
Budgeting 2027 against 4.7x because a report said so
The 4.7x average is dragged upward by a small group of heavy agent adopters. Most organisations landed at 2.2x. Planning to the mean either inflates your ask until finance kills it, or — worse — gets approved and sets an expectation of transformation you have not scoped the integration work to deliver.
The concentration risk nobody is pricing
Anthropic and OpenAI together now account for 86% of enterprise LLM API spend by Menlo's estimate, up from 67%, while Google fell from 21% to 9%. Open-source handles roughly one in five tokens but only one in 25 dollars. The obvious hedge — post-train an open model for your domain — is not yet an operational escape hatch for most teams: adoption rose only from 6% to 10% of AI builders, and 91% of those who do it rely on outside partners.
So the realistic posture for 2027 is not provider independence. It is provider portability: keep prompts, tool schemas and evals in your own repository rather than a vendor console, abstract the model call behind one interface, and maintain a scored fallback model you actually test against monthly. That costs a few engineering weeks now and buys you negotiating position later. If you want help structuring that, start here.
Frequently Asked Questions
Is the $177B enterprise AI spending figure reliable?
It is a credible directional estimate, not an audited figure. Menlo built it bottoms-up from a Morning Consult survey of nearly 400 US enterprise decision-makers, public filings and third-party sources, and reports it as annualised run rate. Prior-year numbers were also reclassified under a new taxonomy, so the 5.4x multiple compares an estimate to a restated estimate. Cite it as a market indicator, not a measurement.
Do agents really consume 80% of AI tokens?
On OpenRouter, yes — Menlo reports agents at 80% of tokens, roughly seven times human-driven usage, with agent traffic up 37x since overtaking humans in February 2026. But OpenRouter is a single model-agnostic gateway, and Menlo acknowledges its users skew toward individual developers and startups rather than large enterprises. The direction is well supported; the exact enterprise share is not established by this data.
How much should we plan to increase AI spend next year?
Use the median, not the average. Menlo found median enterprise spend reached 2.2x the prior year against a 1.8x plan, while the average hit 4.7x against 1.9x planned because a small group of heavy agent adopters skewed it. Plan your base case near the median and model agentic workloads separately with per-workflow spend ceilings, since agent cost scales with work performed rather than headcount.
Why does prompt caching matter so much for agent costs?
Because cache reads can be 95% or more of an agent's token volume. Menlo cites Opus 5.5 at $0.20 per million for cache reads versus $4 for fresh input and $20 for output, blending to roughly $0.70 per million for a typical session. Any change that invalidates your cached prefix — reordering tool definitions, injecting timestamps early in context — can multiply your bill without changing a single model choice.
Should we switch to open-source models to cut costs?
Not as a default. Menlo found open-source handles about one in five tokens but only one in 25 dollars of model spend, and that buyers keep trading up to frontier models for agentic work. Post-training adoption rose only from 6% to 10% of builders, and 91% of those use external partners. The more durable move is portability — own your prompts, tool schemas and evals, and keep a tested fallback model.
Research digest
AI Research Briefing
Honest insights on AI agents, Small Language Models, and local RAG. No hype. Only when we have something worth sending.
- No hype, just measurable outcomes
- Read by 2,400+ engineers
- Unsubscribe anytime