Generative AI Model Providers (Gartner)

Gartner published its first Magic Quadrant for Generative AI Model Providers on 1 September 2026. The report is credited to Birgi Tamersoy, Leinar Ramos, Chirag Dekate, Mike Fang and Radu Miclaus.

The market it defines is narrower than the phrase suggests. Gartner covers vendors that develop and offer foundation generative AI models for enterprise use, across text, image, audio and video, made available through APIs, SDKs, platforms, marketplaces and licensing agreements. That is a definition about first-party models and the businesses that sell them, not about the infrastructure underneath them and not about the applications built on top.

Alibaba Cloud reports that it was named a Leader. That placement was announced on 18 September 2026, and it is the only placement this article could verify from public sources. The remainder of the vendor field and the other quadrants are not named here, because the material available does not establish them and a placement invented to fill a paragraph would be worse than a gap.

The inaugural Magic Quadrant for Generative AI Model Providers, and the market it declares

An inaugural quadrant is a claim that a market has become a purchase in its own right.

The claim is easier to make here than in most categories, because the stack around it has already been carved up. Gartner published an inaugural Magic Quadrant for Cloud AI Infrastructure in July 2026, where Alibaba Cloud is also a Leader. It maintains a Magic Quadrant for AI Application Development Platforms, updated at mid-year 2026. OpenAI reports a Leader placement in the 2026 Magic Quadrant for AI Coding Agents. Compute, models, development platforms and coding agents are now four separate scorecards with four separate buyer conversations.

What is left for a model provider to be graded on, once the infrastructure and the tooling have their own reports? The answer is the thing the definition protects: the model as a product, developed by the vendor, sold or licensed to enterprises that want the model itself rather than a place to build with it.

That is a real purchase, and it is a narrower one than the attention the category receives. A company that wants to call a frontier model through an API from inside its own application is buying exactly this. A company that wants a platform to build agents on is buying something else, and Gartner has a different document for it.

This is the half of the record the site did not have

The other half is already on this site at AI Foundation Models, which covers Forrester's evaluation of the same ground.

That page records a 2024 Wave that ranked large language models directly, under the title AI Foundation Models For Language, Q2 2024. It then records what happened next: the category broke, and Forrester's next edition moved to AI Platforms with the Q3 2026 Wave, where the scorecard grades the environment around the model rather than the model.

So for two years this site has carried a Forrester-only account of how models get bought, and an account built on the premise that the model stopped being the unit of evaluation. Gartner has now published the opposite structure for the same subject.

Both documents can be right at once, and understanding why is most of what a buyer needs from this pair.

Forrester declared the category broken, then gave the model a different job

The Forrester position is captured in the title of its own research note introducing a successor frame: a model is not a business model.

That is not a claim about model quality. It is a claim about where the money and the risk sit. A model on its own does not carry the governance, the data connections, the evaluation harness, the cost controls or the deployment path that an enterprise needs before anything can be put in front of a customer. Forrester's conclusion was to stop scoring models as a category and start scoring the platforms that make them usable.

Gartner's quadrant does the opposite and grades the model business directly. The two positions look like a disagreement and are closer to a division of labor. Forrester is describing the purchase most enterprises actually make, which is a platform. Gartner is describing the purchase a smaller set of enterprises make, which is model access itself, and which has enough volume behind it to support a quadrant.

The tell is in the inclusion logic on both sides. Forrester's platform Wave can be entered by a vendor that hosts other companies' models, because the buyer is purchasing a place to work. Gartner's model quadrant cannot, because the buyer is purchasing the model.

What the report grades is a business, not a benchmark

Look at what Alibaba Cloud cites in its own announcement and the evaluation criteria become visible without reading the report.

The company leads with distribution: more than 460 models open sourced, which it says have produced over 300,000 derivative models and more than three billion downloads. It cites model detail next, naming Qwen3.8-Flash as an open-weight multimodal mixture-of-experts model with 125 billion total parameters and six billion activated per token. It then cites serving reach, at 106 availability zones across 31 regions, and Apsara Stack for hybrid and on-premises deployment.

Read that list as a set of things a model provider is actually graded on and none of it is a benchmark score. Distribution through open weights, licensing flexibility, inference efficiency expressed as activated parameters per token, multimodality, and the infrastructure to serve the model wherever the customer needs it.

That is the shape of a market where capability has converged enough that the differentiators moved elsewhere. When every serious provider can field a capable multimodal model, the questions that decide a deal become how widely the model is used, what it costs to run, what the licence permits, and whether it can be served inside a regulated environment. All four are business questions, which is why Gartner graded the business.

Alibaba Cloud's chief technology officer, Feifei Li, described the model market in the announcement as being at an inflection point. The placement data supports a narrower reading: the market has reached the point where a quadrant can be drawn, which is itself the change.

The layer with the most attention is not the layer with the most revenue

The sizing figures Gartner publishes, and that Alibaba Cloud's announcement cites, are the most useful corrective in the document.

The total market for generative AI models was 13.0 billion dollars in 2025, of which 11.4 billion was first-party foundation models. Gartner projects growth at a 59.9 percent compound annual rate to 139.2 billion dollars by 2030, with 105.4 billion of that in general foundation models and 33.9 billion in domain-specific models.

Set the model layer against the infrastructure spending and the application spending that surround it, and the ratio is startling. The part of the stack that generates the most public argument and the most executive anxiety is not the part with the most revenue in it today.

The growth rate is the interesting number, because a 59.9 percent compound rate is a forecast that the model layer becomes a market rather than an input cost. Two futures are encoded in it. If models commoditize, that growth lands with whoever serves them most cheaply, and the differentiators listed above, distribution and efficiency and licensing, are the whole competition. If models continue to differentiate on capability, the growth lands with whoever is furthest ahead on the frontier, and the market splits into a small number of premium providers and a long tail of cheap ones.

There is a third signal in the split between general and domain-specific models. Gartner projects roughly a third of the 2030 market in models built for a specific domain rather than for general use. That is a forecast that specialization becomes a real business, and it is worth weighing against the open-weight strategy that Alibaba Cloud leads with. A domain-specific model is usually a model someone tuned on their own data, which is closer to Forrester's framing of the platform purchase than to Gartner's framing of the model purchase. The two structures may converge again.

What to ask before you commit to a model provider

What does the licence permit, and what happens when the model is retired? Open weights and commercial licences carry very different exit costs. Ask what the terms allow you to do with outputs, whether you can self-host, and what notice you get when the model you built on reaches end of life.

What does the model cost at your volume, not at the demo volume? Activated parameters and token pricing are the real unit economics. Ask for a cost model built on your own traffic, including the reasoning-heavy requests that consume far more tokens than a benchmark prompt.

Can it be served where your data has to live? Regional availability and on-premises options are a selection criterion in regulated industries more often than capability is. Ask for the deployment topology rather than a general statement about sovereignty.

Which parts of the stack are you buying from one vendor? A provider that also sells the cloud and the platform can offer a coherent stack and a single dependency. Ask what is portable if you later change the model but keep the platform, and the reverse.

Is your workload a general one or a domain one? The forecast that a third of this market will be domain-specific by 2030 is a hint that the general model may not be the right purchase for a narrow, high-stakes task. Ask what tuning and evaluation support the provider offers before assuming the general model is the cheaper path.

Gartner does not endorse any vendor, product or service depicted in its research publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner's research organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this research, including any warranties of merchantability or fitness for a particular purpose.

The platform half of Forrester's record is at AI Platforms, the Q3 2026 Wave, which scores the environment built around a model rather than the model itself. Read against this page, the pair answers a question neither document settles alone: whether the model being bought is the product or an input to one.

Where a provider serves a model from is a selection criterion in its own right, and the site covers one regional market separately at AI Platforms In China. The presence of a regional provider on this quadrant's Leader rung makes that split concrete: availability zones and licence terms decide which buyers can use a model at all, before capability enters the conversation.