Will Companies Build Or Buy Their GenAI Models?

One of the biggest questions that enterprises, governments, academic institutions, and HPC centers the world over are going to have to answer very soon – if they have not made the decision already – is if they are going to train their own AI models and the inference software stacks that make them useful or just buy them from third parties and get to work integrating AI with their applications a lot faster.

While the big model makers like OpenAI, Anthropic, xAI, Google, Alibaba, and DeepSeek are spending enormous funds training AI models with ever-growing data token counts and parameters to make their models smarter, most organizations have very specific needs and fairly limited amounts of specialized data relevant to their own operations, which means we can expect for spending on generalized models as well as on specialized domain-specific models.

This is certainly a pattern we have seen with other kinds of enterprise software over the decades, with general back office accounting in the mainframe era evolving into a sophisticated stack on industry-specific suites that encapsulate the business logic of a manufacturer, a distributor, a retailer, and so forth across the myriad industry segments that comprise a modern economy.

This is something that the prognosticators at Gartner expect to happen with generative AI models, although admittedly it will take time for the differentiated models to be created and adopted.

According to the data just released by Gartner, its analysts expect for end user spending on GenAI models to hit $14.2 billion in 2025, nearly a factor of 2.5X higher than spending on commercial GenAI models last year, which added up to $5.7 billion in spending.

We have added the GenAI model spending to a previous analysis of overall IT spending we did back in January and GenAI spending based on Gartner data that we did back in April of this year. Take a look:

As you can see, spending on third party GenAI models was pretty thin back in 2023, at only $1.4 billion, but grew by a factor of 4.2X in 2024. It is hard to say how it might grow in the coming years, but we think there are some pretty wide error bars that anyone has to use as they make their prognostications. A lot will depend on how the future of GenAI models are architected, and if some other approach takes the market by storm, as many predict it will.

If mixture of expert models with multimodal capabilities turn out to be more accurate than humongous one-shot, blurty models as were the norm until fairly recently and can be trained on relatively small clusters and inferred on similar iron, provided that their creators license them at a reasonable price and provide industry-specific – if not organization-specific – versions of their foundation models, then the market for third party GenAI models should take off.

But if training gets a lot cheaper and models can be run on even smaller clusters, then organizations might just train their own models and set up their own inference farms.

And if training models takes so much infrastructure that most companies cannot afford to do it – or even get their hands on the GPUs or XPUs to do it – then companies will spend a lot of money on third party AI models and the model makers will get rich and recoup at least some of the enormous sums of money they have put into creating models over the past five years.

What Gartner has said definitively is that by 2027, half of the models used by enterprises will be domain specific, up from 1 percent of models (as gauged by revenue we presume) in 2024.

What we can see from the table above also is that the share of GenAI software spending accounted for by the GenAI models themselves is growing as the years go by, rising from 25.2 percent in 2023 to an expected 38.2 percent in 2025.

Of course, all of this money counting by Gartner misses how much money is being spent by the model makers today to create their models, and how much will be spent by other organizations if they start building their models. A very large portion of enterprise software over the decades was created in house and not by third parties, and this huge investment never gets counted by the Gartners and IDCs of the world. And yet, there it is.