What if the most valuable company in AI right now isn’t the one building the smartest model, but the one that turns other people’s models on?
That’s roughly the bet investors are making on Modal Labs. According to TechCrunch, the company is closing in on a $750 million funding round led by Accel, at a valuation of $15.75 billion including the new money. Four months ago, after a $355 million raise, Modal was valued at $4.65 billion. So this round would more than triple the price tag in about a third of a year.
If you follow AI casually, you’ve probably never heard of Modal. You’ve heard of the model makers. You’ve heard of the chatbots. Modal sits in the plumbing, and the plumbing is suddenly where the money is going.
What inference actually means
There are two phases in an AI model’s life, and the vocabulary trips people up constantly, so let me make it plain.
Training is the expensive schooling. You feed a model enormous amounts of data, it adjusts itself millions of times, and eventually it knows things. This happens once per model version, costs a fortune, and gets all the press.
Inference is everything after that. It’s the model doing its job. Every time you ask a chatbot a question, every time an AI agent reads your calendar and drafts a reply, every time a support bot decides whether to escalate your ticket, that’s inference. One training run, then billions of inference calls.
Modal Labs is in the second business. Companies bring their trained models; Modal runs them, handles the servers, scales up when traffic spikes, scales down when it doesn’t.
Why agents make this the interesting part
This is where the story gets relevant to anyone following AI agents, which is most of why I write here.
A chatbot answers one question with one model call. An agent does not. An agent plans, tries something, checks the result, adjusts, tries again, maybe calls a second model to verify the first one’s work. A single task an agent completes on your behalf might involve dozens of model calls chained together.
Multiply that by every company shipping agents, and the demand curve for inference does not look like the demand curve for chatbots. It looks steeper. Investors appear to have noticed, and TechCrunch’s sources describe appetite for AI inference infrastructure as accelerating.
Think of it like the shift from people occasionally buying a physical newspaper to everyone streaming video all day. Same underlying internet, wildly different load.
The part that should make you squint
I try not to write breathless things about funding rounds, so here’s the tension in this one.
TechCrunch’s reporting notes that this sector runs on thin margins. That’s not a small detail. Inference providers buy or rent expensive compute, then resell access to it. The cost of the hardware underneath sets a hard floor on how cheap they can go, and customers are extremely price-sensitive because inference is a recurring bill, not a one-time purchase.
So investors are paying a very high multiple for a business in a space where the economics are known to be tight. That’s either a smart read on where volume is heading, or a bet that gets uncomfortable if growth slows.
I’ll also flag something honestly. One third-party tracker lists Modal at roughly $6.3 million in annual recurring revenue against a $1.1 billion valuation, from an earlier period. I can’t verify that figure or its date against the TechCrunch reporting, and private company revenue estimates are frequently wrong. Treat it as a question worth asking rather than a fact to build on. But it does frame the shape of the debate: are these valuations tracking revenue today, or expected revenue several years out?
What it means if you’re not an investor
A few practical takeaways for anyone building with or buying AI tools.
- Inference cost is your cost. If you’re paying for an AI feature, a chunk of that price is somebody’s inference bill. Competition among providers like Modal tends to push that down over time.
- You probably don’t need to pick a provider. Most people encounter inference infrastructure indirectly, through whatever product they’re already using. It matters to you mostly through pricing and reliability.
- Watch the boring layer. When capital rushes into infrastructure rather than applications, it usually signals that investors expect a lot more usage coming, not just more excitement.
The model makers get the headlines because the models are the visible magic. But a model nobody can afford to run at scale is a research paper. The companies figuring out how to run them cheaply and reliably are quietly deciding how many AI products actually ship.
A $15.75 billion valuation says a lot of investors now believe that’s the more durable position. I’d watch whether the margins ever catch up to the price.
🕒 Published: