There has been a lot of fuss in recent weeks that open weight AI models from China such as Alibaba Qwen, DeepSeek R1, V3, and V4, and Moonshot Kimi K3 are getting up there in terms of capability compared to the closed models such as OpenAI GPT, Anthropic Claude, and Google Gemini, which hail from the US of A.
The fuss, which started last January with the DeepSeek R1 model, is justified because if companies on the American and European continents decide they want open weights – they generally cannot have open source at this point because models are too expensive to make – then they are going to consider Chinese models over the closed weight models that cost a lot more. That will in turn downshift the amount of money that AI model makers in the United States can generate as revenue, and that will perhaps not be sufficient to cover the enormous rental and now acquisition costs of AI hardware that OpenAI, Anthropic, Google, and others are bearing as they look to the GenAI future they are creating in their own vision and that the rest of us will have to cope with.
There is trillions of dollars of investment in datacenters and hardware on the books over the next two years to underpin those GenAI aspirations. And the growth in annualized revenue run rates for OpenAI (about $25 billion, up around 2X from January) and for Anthropic ($47 billion in May, up 5.2X from December 2025) have people holding their breaths a bit but in an optimistic way that somehow the revenues will catch up and surpass the investments.
Well, ARR is not actual revenue, even if it is an interesting indicator of potential future sales levels. But like a point in time temperature or blood pressure measurement, it is not really a good measure of general health. Here at The Next Platform, it’s money that matters, as Randy Newman so accurately observed, and that’s why we are taking a keen look at some new data from Gartner on actual spending on AI models and platforms. This is a new dataset that Gartner has been developing, so let’s take a look:
This is a much more restricted view of AI spending than the one that Gartner put out back in January, which had everything, including the kitchen sink, in the data, from AI PCs and AI servers to AI-enhanced applications for 2025 and forecasts for 2026 and 2027. We shall see how this all pans out. I don’t quite buy it all yet, even though the hyperscalers, cloud builders, and AI model makers have to buy it because their whole superb growth story is dependent on us all believing that this GenAI boom will not just be another monstrous bubble.
I mean, that’s how bubbles work in a Pangean economy, right? Ask John Law, who might be spinning in his grave at 7,200 RPM like an old disk drive. . . . We are all stitched together by capital and goods flows here on Earth, whether we like it or not, and our 401(k)s are all dependent on this GenAI boom being a real transformation that will also not wreck that economy and the lives of people.
In any event, we wanted to drill down on AI model and platform spending, which is so relevant as Anthropic, OpenAI, Moonshot, and DeepSeek are all looking to go public on the idea that they are going to sell tons of licenses to software and support or API access to their respective models.
In the table above, Gartner pegged GenAI model spending in a 2024 research report at $1.4 billion in 2023 and $5.7 billion in 2024, with a growth rate of 320.4 percent. We rejiggered the numbers with some more significant digits to make the rounding and the percent change work, which is why the numbers are in blue bold italics. Estimates by me based on ratios and hunches are shown in bold red italics as usual.
What is interesting here is how fast the growth rate in GenAI model revenues is slowing. It was triple digits in 2024, and the growth rate is less than half in 2025 and has decelerated a bit here in the forecast for 2026. In fact, that 2025 GenAI model figure is a bit lower than it was in the January numbers, but that may be because it also included traditional AI machine learning models. In any event, the revenues for 2026 at $28.3 billion, are nowhere what you would expect based on the ARRs for just OpenAI and Anthropic alone. We presume that this Gartner number for GenAI model revenues includes foundation model licensing as well as API subscription access to foundation models.
Go figure.
What is also interesting – and useful – is the breakout of domain specific and specialized GenAI models, which are growing in popularity and which are growing twice as fast as revenues for broader foundation models that can do anything. I have no problem believing that such models could represent half of GenAI model revenues in time, and that their popularity will increase, not level out. Think about back office applications. These ERP, SCM, and CRM systems are built, but they are also heavily customized to match the needs of specific customers, and the mods made by customers represent a lot of the cost of installing and maintaining such systems. (Which is why they are not necessarily less costly than building your own apps from scratch.) AI models should be no different.
What is also interesting is that the AI platforms that are used by companies to run their models and develop applications that use GenAI functions are already a larger market than for the AI models themselves. This suggests an accelerated maturity, and proves that GenAI is just an extension of AI machine learning and even HPC in some sense. (The taxonomy for these AI platform categories is here at Gartner.) The AI development platforms include ones available from IBM, SAS, H2O, and others as well as Amazon SageMaker and Bedrock, Google Vertex AI, and Microsoft Azure Machine Learning, among several many others.
But the other thing is that the market for such GenAI platforms is growing much slower than for the models themselves. That could be because the AI models are API heavy in terms of revenues, or it could be because companies want to license their own models and control their own GenAI fates. That’s the smart thing to do, and it is what the hyperscalers and cloud builders and now the AI model builders are doing for themselves. You are underwriting their experimentation and deployment of AI tools they will use themselves or to enter new markets.
When it comes to GenAI models, I think fully open source models – including code as well as weights – is the real big threat here for those selling access to closed foundation models. Because Nvidia is so rich from selling AI hardware that has dominant market share in the world, it is really the only company that is rich enough to give away its Nemotron 3 foundation models. No one else can afford this. Full stop. This is why Meta Platforms is not working on its open source Llama models anymore and is emphasizing its Spark Muse models, which almost certainly will not be opened up.
The fact that Nvidia is the only one that can afford to give away its foundation models (and maybe the Chinese model makers who want to make mischief with the US model makers) has all kinds of implications for the long-term growth of GenAI model licensing and subscriptions. If you have to choose free over very expensive when the hardware is crushingly expensive, which way are you going to go?
Oddly enough, giving it away and taking on its largest customers may get Nvidia in trouble with its biggest customers and even the law at some point. A lot depends on the appetite of the US and EU governments for a lawsuit that might accuse Nvidia of using its monopoly power in AI hardware to underwrite its free Nemotron models. It is hard to say what the Trump administration will and won’t do, with the model makers and Nvidia cozying up here and there with the White House to try to sway. The AI model makers, according to a post in the Wall Street Journal, are certainly not pleased with open weight and open source models out there, which are no doubt a security threat but which are also a competitive threat to Anthropic, OpenAI, and Google, who want to be able to charge for foundation model use.
The future of their businesses depends upon this. And from what we can see, 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.