What happened
Menlo Ventures published *2026: The State of Enterprise AI* on November 3, 2026. The report, by Tim Tully, Joff Redfern, Deedy Das and Derek Xiao, estimates that enterprise spending on deploying AI reached $176.8 billion by September, up 5.4 times from $33 billion in 2025. It divides deployment into $99.5 billion for infrastructure and $77.3 billion for applications. Separately, it estimates $102.5 billion went to developing models: $87.2 billion for training compute and $15.3 billion for training data and services. Menlo summarizes the relationship as roughly $1.70 spent deploying AI for every dollar spent developing models.
The report says agents have become major consumers of models. On OpenRouter, a platform offering access to hundreds of models, agents generated 80% of tokens, or 130.9 trillion per week, according to Menlo’s figures. The report says agent traffic overtook human-driven traffic in early February after the launches of Opus 4.6 and GPT-5.3-Codex, and had grown 37 times since then, compared with five times for human-driven use. Menlo also reports that 74% of AI builders use agents to manage workflows and execute tasks, and 32% use multiple agents together. These figures describe the report’s estimates and the OpenRouter channel, not a census of every AI workload.
Spending on applications is concentrated: coding accounted for an estimated $34 billion and enterprise chatbots $25 billion, together 76% of the $77.3 billion application market. Menlo attributes the growth in these categories partly to conditions that support agent work: access to relevant systems, usable context and outcomes that can be checked. Coding has compilers and tests to help verify results; the report contrasts this with functions where judgment is harder to evaluate. It estimates that coding spend grew 7.7 times year over year, while chatbot spend grew three times.
The report also finds a split between model usage and spending. From September 2025 to September 2026, open-source models’ share of tokens rose from 11% to 19%, while their share of model spending fell from 6% to 4%. Menlo estimates open-source token volume at 2.9 quadrillion per month, up 22.5 times, and says the average open-source token price fell 3.3 times to $0.13 per million. Frontier models gained share too: from Q3 2025 to Q3 2026, their share of spending rose from 35% to 63%, and their share of tokens from 18% to 42%. Menlo reports Anthropic and OpenAI together held 86% of enterprise LLM API spend in 2026.
What the archive adds
New Runtime’s earlier coverage, “There is no single AI adoption rate” (November 2, 2026), compared differently defined estimates of AI adoption. Menlo’s report is a new event with a related but distinct focus: estimated spending, model usage and agent activity, rather than a single adoption rate. Its figures broaden the picture, but should not be treated as directly comparable adoption measures.
Trend signal
This does not yet fit a pattern New Runtime tracks. Menlo argues that spending is increasingly associated with agent activity and work performed, but the report’s spending estimates do not establish cost per verified completed task—the unit needed to substantiate a shift toward completed-work economics.
Why now
Menlo’s estimates cover spending through September 2026 and compare them with 2025, a period in which the report says agent use expanded sharply. The timing matters because companies are budgeting for AI use while the mix of workloads, providers and model types is changing. The report presents its September totals as a point-in-time estimate, not a final account of the full calendar year.
Why it matters
For an AI team, the headline is not simply that spending is rising. Menlo’s estimates suggest the economic question is moving from whether an organization has access to models toward which work it can put them to use on, and at what cost. The concentration of application spending in coding and chatbots is consistent with domains where agents can reach the relevant systems, draw on procedural context and produce outputs that can be checked.
The report also cautions against reading usage share as budget share. Open-source models handle a growing portion of tokens but a much smaller portion of spend, while frontier models gained both usage and spending share. A team comparing providers therefore needs to measure outcomes and total workload costs, not just token prices or popularity. Menlo’s numbers can help frame those questions, but they are market estimates rather than a substitute for a team’s own workload data.
Limits and uncertainty
These are Menlo Ventures’ estimates, not independently verified market totals. The report’s agent-usage figures from OpenRouter describe activity on that platform; the report itself notes that it is one channel. The excerpted figures also do not establish how Menlo defines every spending category or how estimates account for overlap among infrastructure, model providers and applications. Market-wide shares cannot show whether a particular company’s agents save money, improve quality or complete tasks reliably. Nor do token and spending shares alone explain which tasks generate value, how much human review is required, or how results vary by sector and workload.
Reader takeaway
Menlo’s report estimates that enterprise AI deployment now draws more spending than model development, with agents and a few application categories driving much of the story. Treat its figures as directional market estimates, not a purchasing prescription. For a team deciding what to build or buy, the useful next step is to measure cost, verification effort and task outcomes on its own workloads.
Open question
Will rising agent usage translate into measurable reductions in the total cost of completing enterprise work, including human review and correction?