On September 16, Gartner revised its 2026 AI spending forecast to $2.7 trillion, a 49.5% increase from 2025. In May the same house had said $2.59 trillion and 47%. The headline moved. The buyer did not. John-David Lovelock, a distinguished vice president analyst at Gartner, said demand for AI infrastructure is still strong and inelastic to higher memory prices, and that hyperscalers and service providers buying AI-optimized servers remain the largest single area of spending.
Read the table under the headline before you take the round number into a board pack. Gartner’s line items for 2026 sum to $2.67 trillion. AI infrastructure is $1.48 trillion of that. AI services are $576 billion. AI software is $462 billion. AI cybersecurity is $51 billion, almost double the 2025 line. Agents and assistants are $29 billion. Generative AI models are $28 billion. Infrastructure is about 56% of the table. The software a mid-market company can actually put on a purchase order is a minority of the story Gartner is telling.
What changed since May
Gartner raised two near-term growth rates in this quarter’s forecast. Spending on AI application-development platforms is now expected to grow 39% in 2026, up from 28% in the prior forecast, because enterprises, software vendors, and services firms want custom applications. Growth in generative AI models moved from 110% to 117%. Gartner called that a small but growing opening for domain-specific models that cost less and fit a job, not a general demo.
It also split cross-functional agents and assistants out of the AI software line, and added consumer agents, so the new table is not a clean restatement of the May pile. Do not subtract $2.59 trillion from $2.7 trillion and call the difference new enterprise demand. Part of the revision is a bigger infrastructure build. Part of it is a new way of counting agents.
If more than half the forecast is servers, chips, networks, and AI cloud, matching “the industry” is not a budget. It is a category error.
Where the company’s check actually lands
Gartner says generative AI is in the trough of disillusionment in 2026, and that enterprises are using simpler AI features embedded in software they already have: faster workflows, customer responses, decisions. Lovelock’s other point is the one a CFO can use. Buyers are going to service firms less often for a transformation program and more often for smaller projects that turn on the AI already inside an incumbent system. Gartner still sees those two kinds of work combining into a $1.2 trillion AI-services opportunity by 2030. That is a vendor forecast. It is also a description of the renewal you will see this quarter: the suite you own, plus an agent SKU, plus a services line to switch it on.
That is a different problem from the $2.7 trillion. You do not buy the hyperscaler’s data center. You do buy the seat, the consumption overage, and the implementation that “just enables” the feature the vendor shipped to stay relevant. Gartner said vendors are embedding agentic AI in existing products to defend against new cross-functional agents. The defense shows up on your invoice.
What to take to the next renewal
Separate three numbers and refuse to let them share a slide. One: industry infrastructure spend, which you do not control. Two: the AI line inside a product you already license. Three: a workflow with a definition of done, and whether that workflow moved.
Ask the vendor which of Gartner’s buckets the quote sits in. An embedded feature in a suite you renew is software. A usage meter on a model is not a seat. A partner “to turn it on” is services. If the quote will not say which, you are being sold the headline.
The May piece on this site still stands on the part that was never about the size of the forecast. A larger industry bill does not create a return on your line. Gartner revised the bill upward in four months. The question for an operator is unchanged: which check on this forecast is yours, and what does it retire?
The post, then the table
Gartner’s own account posted the revision on September 16 and linked the newsroom page: AI spending up 49.5% in 2026. That post is the announcement. It is not the table. The table is on the newsroom page, and it is less round than the post. The line items sum to about $2.67 trillion, not a clean $2.7 trillion. Infrastructure is $1,484,397 million. Services are $576,481 million. Software is $461,637 million. Cybersecurity is $51,347 million, against $25,920 million in 2025. Agents are $29,219 million. Models are $28,266 million. The 2025 total on the same table is $1,786,671 million. Gartner’s 2027 total is $3,637,292 million, which is about 36% above 2026. Use the post to know the day the claim went public. Use the table when someone tries to put “$2.7 trillion” in a cell that is supposed to be your budget.
Two weeks later, on September 30, Gartner posted a different pressure. As enterprises ramp AI spending, the account said, CIOs face growing pressure to deliver more with limited budget growth. The post points at a 2027 CIO Agenda Preview and says the preview reveals four technology-planning disconnects. It does not name the four. This piece will not name them either. What an operator can use without the preview is the collision. The September 16 forecast says the industry bill is still rising, and that more than half of it is infrastructure a mid-market company does not buy. The September 30 post says the CIO is being asked to produce more without a budget that rises in step. Those two sentences are how a renewal gets approved on a headline and paid for out of a flat operating plan. Separate them in the pack: industry spend, the line you will sign, and the workflow that has to move.
The same week’s other major product event sits in a third bucket. On September 28 NVIDIA announced a fence around agents, OpenShell in software and a BlueField watchdog it calls Sentry. That is not a line in Gartner’s infrastructure total unless you are the one buying the chip. It is a candidate for the services-and-software check this forecast keeps hiding under the round number. The write-up is NVIDIA put the fence outside the agent.
Sources
- Gartner on X, AI spending up 49.5% in 2026, September 16, 2026
- Gartner, “Forecasts Worldwide AI Spending to Grow 49.5% in 2026,” September 16, 2026
- Gartner, “Forecasts Worldwide AI Spending to Grow 47% in 2026,” May 19, 2026
- Gartner on X, CIO budget pressure and the 2027 agenda preview, September 30, 2026
- Gartner on X, managing AI cost is becoming as important as deploying it, September 24, 2026
Earlier briefing: AI spend is not the same as AI return. Operator guides: How to tell AI spend from AI return. For one workflow, use token cost to gross margin, then Put an AI cost on one workflow.