How are companies using AI?

How are companies using AI?#

People love to talk about their pets, the weather, a recent vacation. But there’s one thing that people love to talk about even more: what they’re using AI for.

• I used ChatGPT to build me a workout plan
• I vibe coded a NYT cooking inspired recipe generator
• I had AI write me a song about the Python 3.14 release notes

It’s not just people using AI anymore; corporations want to vibe code a recipe app customer support platform too. According to Menlo Ventures’s “2025: The State of Generative AI in Enterprise”, AI spend in 2025 was $37 billion, up 3.2x from the year before. $18 billion of that spend was on something called AI infrastructure.

AI infrastructure spend#

I happen to work at one of these providers called Modal. Thankfully, we can go deep with our data in a way you can’t with an industry report:

• Actual (not self-reported) spend over time at the company level
• We can infer detailed use cases from customer conversations and sales notes

In the spirit of not letting AI replace me, I went through each one of our top 100 or so customers and bucketed them into one of the following use cases:

• LLM inference: running custom or open-source LLMs to build chatbots, document parsing, or low-latency code generation
• Computational biology: predicting molecular structures or protein folding (thus concludes all the comp bio buzzwords in my arsenal tyvm)
• Audio inference: transcription, text-to-speech, generative music
• Image / video inference: running custom or open source models to generate ghibli images of yourself and things like that
• Coding agents: vibe-coding apps and tools

This methodology requires human judgment and isn’t perfect. It doesn’t take into account customers with multiple use cases. I’ve also taken some liberty in excluding some outlier customers that skew the data in an unhelpful way. Here's what I learned...

Coding agents are (apparently) properly rated#

Menlo’s report describes coding as AI’s first “killer use case”, and we’d have to agree. We’re seeing the fastest spend growth among coding agent companies like Cognition and Lovable.

We’re also seeing other customers use the same technology that powers coding agents (sandboxes) in new use cases like reinforcement learning. We expect this category to continue to grow.

LLM inference is increasingly self-hosted#

At Modal, we think of LLM inference as three different types of workloads:

• Offline - batch, analytical processing, asynchronous
• Online - interactive, low-latency, synchronous
• Semi-online - mix of the above two

Choosing OpenAI or another closed source provider is like choosing a Toyota Camry; a middle-of-the-road option where you may not want to think too hard about mpg or top speed or acceleration. You just want to get the job done reasonably well.

But if you’re building a business, customizing performance of your offline / online / semi-online workload matters more and more. These companies might choose the “OS stack”: pairing an open-source like Qwen with an open-source inference engine like vLLM that gives them that level of customization.

As you can see above, the spend we’re seeing from LLM inference companies using the OS stack has skyrocketed over the last 6 months. Reducto is solving the PDF extraction problem at scale (they parsed the Epstein files). Decagon is using online, low-latency voice chat in their core AI customer service product.

Science is still alive#

The US may not be a measles-free country for long, but computational biology startups aren’t discouraged. A small, but steadily growing share of our spend is towards science use cases like drug discovery. And to think that all this time, I thought running reproducible experiments was only relevant in e-commerce (that’s what you get when you spend 10 years as a data scientist in tech).

B2B is so back (even though it never left) #

Even though Modal has thousands of paying customers, we are but a small slice of the AI infrastructure pie. Still, slicing spend by the categories above can give us a uniquely detailed view into how some of the fastest growing AI companies are spending that sweet sweet AI money.

Overall, the AI market has matured. It’s moved from generating AI spongebob memes to writing 30% of a $32 billion dollar company's code. If 2025 was AI’s experimental college phase, 2026 will be the year it’s going to start looking for jobs. Something tells me it’s going to ace the interview.