Gartner's new dataset puts GenAI model spending at $28.3 billion for 2026, a figure that sits awkwardly next to the trillions of dollars of datacenter and hardware investment now on the books for the next two years. The gap is the story. Annualized run rates at OpenAI and Anthropic suggest a boom; audited spending on models suggests something considerably more modest.

The Next Platform, which analyzed the Gartner numbers, frames the problem bluntly. OpenAI's annualized revenue run rate stands at roughly $25 billion, up about twofold since January. Anthropic reached a $47 billion run rate in May, up 5.2 times from December 2025. These are indicators, not invoices. As the analysis notes, ARR is "not really a good measure of general health," comparing it to a single blood pressure reading.

What do the Gartner numbers actually show?

The dataset tracks spending on AI models and platforms, a narrower cut than Gartner's January forecast that bundled AI PCs, AI servers and AI-enhanced applications. The historical baseline: GenAI model spending was $1.4 billion in 2023 and $5.7 billion in 2024, growth of 320.4 percent.

The forward path is less dramatic:

• Growth ran at triple digits in 2024
• The 2025 growth rate is less than half that pace
• The 2026 forecast decelerates further, to $28.3 billion

That 2026 figure is nowhere near what the OpenAI and Anthropic run rates alone would imply. The presumed scope covers foundation model licensing plus API subscription access.

Where is the growth concentrated?

Domain-specific and specialized GenAI models are growing twice as fast as general-purpose foundation models. The Next Platform expects such models could eventually represent half of all GenAI model revenue. The analogy is enterprise software: ERP, SCM and CRM systems are heavily customized, and those modifications carry much of the installation and maintenance cost. Models, it argues, will follow the same pattern.

Meanwhile, the AI platform market — SageMaker and Bedrock from Amazon, Google Vertex AI, Microsoft Azure Machine Learning, plus offerings from IBM, SAS and H2O — is already larger than the model market itself. It is also growing more slowly. One reading is that enterprises want to license their own models and control their own GenAI fates, which is precisely what hyperscalers, cloud builders and model builders are doing for themselves.

Who can afford to give models away?

The open-weight question hangs over everything. Chinese models — Alibaba Qwen, DeepSeek R1, V3 and V4, and Moonshot Kimi K3 — now approach the capability of closed models from OpenAI, Anthropic and Google. If European and American enterprises choose open weights, they will weigh free Chinese options against costly American ones, and the revenue base of US model makers shrinks accordingly.

The analysis is blunt on who survives free. Because Nvidia earns so much from dominant AI hardware market share, it is "really the only company that is rich enough to give away its Nemotron 3 foundation models. No one else can afford this." Meta Platforms, by this account, has stopped work on its open source Llama models and now emphasizes Spark Muse, which will almost certainly stay closed.

Nvidia's generosity carries legal risk. Giving away models while competing with its own largest customers could, at some point, invite a US or EU antitrust suit alleging that monopoly power in AI hardware underwrites the free Nemotron line. The Wall Street Journal reports that model makers are displeased with open weight and open source models generally — a security concern on paper, a competitive threat in practice.

Why the imbalance matters

The structural issue is simple arithmetic. Trillions in committed capital — datacenters, hardware rental and now outright acquisition — must be recovered from a model and platform market that Gartner sizes in the tens of billions. The model makers bearing those costs, Anthropic, OpenAI, Moonshot and DeepSeek among them, are all reportedly eyeing public listings on the premise that licenses, support and API access will scale.

The Gartner figures do not yet support that premise. ARR may keep climbing, and the January Gartner forecast, broader in scope, painted a larger picture. But actual model revenue, growing at a decelerating rate toward $28.3 billion in 2026, remains skinny against the capital deployed. As the analysis concludes, the actual revenues for GenAI models and platforms are "pretty skinny compared to where they need to be given the humongous investments they are making." Whether domain-specific models and platform tooling close that gap before the investment cycle matures will decide whether this was a transformation or a very well-funded experiment.