Why buying GPU like it's a rectangle is the mistake every AI founder still makes

Erik Bernhardsson, CEO of Modal, describes a problem every founder running agents in production has already felt in their pocket: cloud contracts are fixed, but AI demand never is. The gap between these two formats turned into a $16 billion business.

The contract every founder buys, but no one should defend

When you sign a GPU capacity contract today, the format is almost always the same: a fixed number of cards, for a fixed period. Plot that on a capacity-over-time graph and you get a perfect rectangle. That's how Amazon, Microsoft, and most dedicated GPU providers have sold compute for years, and that's how most technical founders learned to buy.

The problem, as Erik Bernhardsson, CEO and co-founder of Modal, puts it in an interview with the Latent Space podcast published on October 22, 2026, is that no real AI workload behaves like a rectangle. It grows, spikes, and drops again over the course of a day. Buying for the peak means paying for idle hardware most of the time. Buying less than the peak means losing demand exactly when it shows up. It's not a technical detail: it's the budget hole in almost every AI startup that scaled too fast for a contract that doesn't track real usage.

Modal's bet: turning duration into elasticity

Modal's thesis, which Bernhardsson calls a kind of duration transformation, is simple to state and hard to execute: aggregate demand from thousands of customers, guarantee capacity underneath, and let each customer use exactly what they need, when they need it. The smoother the aggregate demand gets, the more efficient the pool becomes, because the peaks of some offset the valleys of others.

The idea isn't new in infrastructure, but the timing is. After a $5 billion Series C about five months ago, Modal reports having crossed $1 billion in annualized revenue, raised a $750 million Series D valuing the company at $16 billion, and has already launched close to three billion sandboxes, ephemeral environments where models and agents execute code. These are numbers that confirm one thing: the market is paying a real premium to avoid buying rectangles.

What sandboxes reveal about demand that didn't exist two years ago

The most revealing part of the conversation isn't the valuation, it's the origin of the product that now sustains this growth. Modal launched sandboxes after Bernhardsson and co-founder and CTO Akshat Bubna noticed that language models generated code, and someone would need a place to run that code. Initial adoption was so modest that the team even discussed discontinuing the product.

What changed the game were coding agents, background agents, and reinforcement learning, workloads that create an environment, run code, inspect the result, and try again, repeatedly. Bernhardsson says sandboxes now account for about a third of Modal's revenue, a result he himself didn't predict at launch. For those building agent products, this is the real demand curve: no fixed-capacity contract was designed for an agent that decides, on its own, how many times it will try again.

In short: the product that was almost canceled for lack of use became the strongest argument in favor of the elastic model, because it's exactly the kind of workload no rigid contract anticipates.

The risk Modal takes on in place of the customer

There's a financial side to this equation that Bernhardsson doesn't hide. If a provider locks in capacity at today's prices and the price of GPUs falls afterward, it ends up holding expensive supply. He says he doesn't expect an imminent collapse, but he treats the possibility as something that weighs on how the business is built, not as a remote scenario to ignore.

This is the point every founder should understand before migrating a workload to an aggregated compute provider: the elasticity you're buying has an embedded price, which is the market risk the provider took on on your behalf. When it works, you only pay for usage and sleep soundly. When the provider gets the bet wrong, or simply doesn't survive its own business model, your workload depends on another company's financial health, not your own.

The counterargument that deserves to be taken seriously

Not every workload benefits from elasticity, and it would be shallow to pretend otherwise. Training large models generally runs at high utilization over long, predictable periods: that's exactly the scenario where reserved capacity, bought at a discount, ends up cheaper per GPU-hour than any pooling layer. That's also why Modal coexists, rather than competing alone, with providers like Evan Conrad's SF Compute, which bets on making capacity tradeable with physical settlement, instead of slicing it into elastic pieces.

The difference in approach matters:

None of the three is universally right. The question every technical founder should ask isn't "which one is better," it's "does the shape of my demand look more like a rectangle or like the electrocardiogram of an agent running in a loop."

What changes for those building with agents in Brazil

Most Brazilian startups putting agents into production today don't have the scale or bargaining power to negotiate a reserved-capacity contract with a meaningful discount. In practice, that should work in favor of the elastic model, not against it: it's precisely the founder who can't predict their own product's usage peak who most needs a provider that absorbs that variance.

But the lesson from Bernhardsson's comments on GPU price risk counts double for anyone outside the United States: if the compute cost underlying your agent depends on an American provider's financial bet on the future price of GPUs, it's worth understanding, before signing, what happens to your operation if that bet goes the other way.

A case like Mage's, a customer that took Modal from $100 in monthly revenue in November 2022 to $30,000 in January 2023, according to Bernhardsson, is the kind of disproportionate growth that only an elastic model can keep up with frictionlessly. It's also the kind of dependency every founder should map out before relying on it.

My take, after listening to the whole conversation: for an agent product in growth stage, elastic compute isn't a luxury, it's the only way for infrastructure cost to scale alongside real usage instead of ahead of it. But a vendor's elasticity doesn't replace founder due diligence, especially when the provider absorbing your GPU risk is, itself, betting against the market.