Summary

  • Gartner revised its worldwide AI spending forecast three times in 2026: $2.53 trillion in January, $2.59 trillion in May, and $2.67 trillion in September, while changing the segment taxonomy twice.
  • In September 2026, "AI Models" was relabeled "Generative AI Models" and "AI Agents and Assistants" was carved out of AI software as a standalone line, with consumer agents added to the forecast's scope.
  • Gartner's own methodology caveat states that cross-iteration comparisons are "not meaningful" because the scope has widened — a caveat that complicates the headline growth-rate narrative many outlets repeat.
  • AI infrastructure is the largest segment at roughly 56% of the September total, but its largest sub-segments are devices and vendor-driven server purchases, not enterprise compute commitments.
  • Some of the "growth" in individual segments reflects redefinition as much as demand: the generative AI models growth rate rose from 110% to 117% in a single revision cycle, in part because the category itself narrowed.

A single number now dominates the AI economy discussion: $2.7 trillion of worldwide AI spending in 2026, up 49.5% year over year. That figure comes from Gartner's September 2026 press release, part of a forecast the firm has revised repeatedly during the year (Gartner, September 16, 2026).

Three vintages, three different measuring instruments

The September headline is best read against the two vintages that preceded it. In January 2026, Gartner forecast worldwide AI spending of $2.52 trillion for 2026, a 44% increase, with AI infrastructure at $1,366,360 million (Gartner, January 15, 2026). In May, the Business Wire syndication of Gartner's release put the 2026 total at $2.59 trillion, up 47%, with infrastructure revised upward to $1,431,509 million (Morningstar/Business Wire, May 19, 2026). By September, the total had climbed to $2,670,460 million, with infrastructure at $1,484,397 million (IT Voice syndication of Gartner Table 1).

Those are large movements — roughly $140 billion added to the 2026 total in eight months — but the more consequential changes are structural. The September table has nine segments; the January and May tables had eight. "AI Models" became "Generative AI Models." And a new line, "AI Agents and Assistants," appeared with its own 2026 value of $29,219 million, after Gartner stated it had "separated cross-functional agents and assistants from AI software and added consumer agents and assistants into the AI spending forecast to better show the emerging opportunity in this area" (Gartner, September 2026).

Gartner has been explicit that these are not like-for-like revisions. The methodology note accompanying the first dedicated AI spending forecast states: "This is the first iteration of the forecast on AI spending that Gartner has published. Gartner has significantly expanded and modified its AI forecast coverage. Spending comparisons to previous iterations are therefore not meaningful as the scope has widened" (Software Strategies Blog, February 16, 2026). In other words, the January-to-September trajectory that headlines present as accelerating demand is partly a widening ruler.

What the September table actually counts

The full September 2026 Table 1, in millions of US dollars for 2025/2026/2027, reads: AI Services 434,046 / 576,481 / 745,655; AI Cybersecurity 25,920 / 51,347 / 85,997; AI Software 288,168 / 461,637 / 656,353; AI Agents and Assistants 16,481 / 29,219 / 65,472; Generative AI Models 13,021 / 28,266 / 51,620; AI Platforms for Data Science and Machine Learning 19,405 / 26,445 / 35,552; AI Application Development Platforms 6,885 / 9,541 / 12,478; AI Data 826 / 3,126 / 6,480; AI Infrastructure 981,920 / 1,484,397 / 1,977,685; Total 1,786,671 / 2,670,460 / 3,637,292 (IT Voice, September 2026).

AI infrastructure is by far the largest segment — nearly $1.5 trillion, or roughly 56% of the 2026 total (Campus Technology, September 21, 2026). Gartner defines that segment as including AI-optimized IaaS, AI-optimized servers, AI network fabric, AI processing semiconductors and devices (Gartner, September 2026).

The internal composition of that bucket matters more than its size. A panorama of Gartner's April 2026 detailed release lists the infrastructure sub-segments as Devices at $604.1 billion, AI-Optimized Servers at $468.9 billion, AI Processing Semiconductors at $286.0 billion, AI Networking at $34.1 billion, and AI-Optimized IaaS at $38.5 billion (Locsic). Only the last of these — the cloud layer where enterprises actually rent inference and training capacity — is small. The dominant dollars are devices and servers bought by technology providers, not by enterprises deploying AI workloads. Gartner itself says capacity growth from hyperscalers and service providers purchasing AI-optimized servers will remain the largest single area of spending (Gartner, September 2026).

The ratio between the top and bottom of the forecast is stark: for every dollar spent on generative AI models in 2026 — a projected $28.3 billion — more than $52 will be spent on AI infrastructure, and AI-optimized IaaS will reach $42.3 billion, up 96% year over year, with inference ($23.3 billion) overtaking training ($19 billion) (Campus Technology, September 21, 2026).

Where "growth" is redefinition

Two September growth-rate revisions illustrate how category boundaries move the numbers. The 2026 growth rate for AI application development platforms rose from 28% to 39%; the rate for generative AI models rose from 110% to 117% (Gartner, September 2026).

The generative AI models case deserves scrutiny. In January, the segment was called "AI Models" and valued at $26,380 million for 2026; in September, "Generative AI Models" is valued at $28,266 million. The label narrowed while the headline growth rate rose — meaning the 117% figure describes a differently shaped category than the 110% figure did. Labeling changed from earlier vintages, with the January and May tables using "AI Models" and the September table using "Generative AI Models" plus a separate "AI Agents and Assistants" row (IT Voice). A reader comparing 110% to 117% across vintages is comparing two different definitions.

Gartner also discloses that segment values are rounded independently and do not always sum to stated totals (Software Strategies Blog) — a small point, but a reminder that the table's apparent precision outruns its arithmetic.

What the categories actually count

The definitions matter because the aggregate headline fuses heterogeneous budget lines. AI Services, in Gartner's definition, covers professional services that help enterprises plan, implement and operate AI systems — strategy consulting, system integration and managed operations. AI Software is enterprise application software with built-in AI capabilities — CRM, ERP, productivity suites — not developer tools (Locsic). Gartner projects a $1.2 trillion AI services opportunity by 2030, driven by transformation and indirect projects (Gartner, September 2026).

In other words, much of what "AI spending" counts is consulting and embedded software features — line items whose AI content is a characteristic of a product, not a dedicated AI budget. When infrastructure vendors' device purchases dominate the largest segment, the aggregate measures the supply side's capital formation at least as much as enterprise adoption. Gartner's framing acknowledges this: the infrastructure segment is vendor-driven rather than enterprise-ROI-driven (Locsic), and Gartner cautions that its AI and overall IT spending categories should not be treated as separate, directly comparable buckets (Campus Technology).

The caveat that governs everything

Gartner's own methodological statement — that the forecast is the first iteration of its kind, that scope has widened, and that comparisons to previous iterations are not meaningful (Software Strategies Blog, quoting Gartner's 4Q25 forecast documentation) — is the most important sentence attached to the $2.67 trillion figure. It means the three 2026 press releases should be read as three partially different forecasts, not one forecast revised. The April-detailed release that the September superseded had placed AI agent software inside the AI Software segment rather than as its own line (Locsic); by September, agents and assistants were a standalone segment.

None of this makes the forecast wrong. It makes it a moving instrument whose readings cannot be chained together naively. The defensible statements are the within-vintage ones: in September 2026's taxonomy, infrastructure is 56% of the total; generative AI models are a rounding error next to device and server spending; and the fastest-growing segments are the smallest ones.

The number that will be quoted for the next year — $2.7 trillion, +49.5% — is a measurement of an instrument that was rebuilt twice during the measurement year. Anyone using it to size a market, validate a strategy or justify a buildout should first ask which vintage they are quoting, and what that vintage counted.