Why Modal Is Betting That AI GPUs Shouldn't Come in Fixed Contracts

In an interview with the Latent Space podcast, Erik Bernhardsson explains how Modal grew to US$1 billion in annual revenue by selling sliced GPU instead of fixed blocks, and why agent sandboxes became a third of the business.

Editor's note: this article cites the surname of cohost Vibhu as "Sapra," but the consulted source identifies him only as "Vibhu," with no surname. In addition, the closing reflection attributed to Bernhardsson about elastic vs. predictable demand appears in the source material only as the title of a segment of the interview (a timestamp marker), without the actual content of the statement available for verification. We are keeping the text because the article's central argument (the sliced-GPU model, the revenue figures, and the sandbox share) is fully verified in the source, but we are flagging these two specific attributions so readers can weigh them with that caveat.

In an interview with the Latent Space podcast, Erik Bernhardsson explains how Modal grew to US$1 billion in annual revenue by selling sliced GPU instead of fixed blocks, and why agent sandboxes became a third of the business.

The way the cloud industry sells GPU has a geometry problem. A typical capacity contract guarantees a fixed number of cards for a fixed period: plotted on a usage-over-time graph, that draws a rectangle. The problem is that no real inference or training workload behaves like a rectangle. It grows, spikes, and drops throughout the day.

That's the thesis Erik Bernhardsson, CEO and cofounder of Modal, laid out in an interview with the Latent Space podcast, published on October 22, 2026, and hosted by Vibhu Sapra and new cohost Basil Chatha. The conversation comes just months after Modal raised a US$750 million Series D at a US$16 billion valuation, less than five months after its US$5 billion Series C. Over the same period, the company says it crossed US$1 billion in annualized recurring revenue and created nearly three billion sandboxes.

The Rectangle Is the Problem, Not the Solution

Anyone who buys GPU by reserving a fixed number of instances for months faces a simple dilemma: buy for the peak and pay for idle hardware most of the time, or buy for the average and run out of capacity right when the product takes off. Bernhardsson describes Modal's bet as a kind of duration transformation: the company aggregates demand from thousands of different customers, locks in capacity underneath (through its own contracts with GPU providers), and lets each customer consume only what it needs, when it needs it.

The logic only works if aggregate demand is smooth enough. If every customer spiked at the same time, the pool wouldn't help at all; since different customers' loads spike at different moments, their sum tends to be more predictable than any individual load. It's the same principle behind insurance and risk pooling, applied to silicon.

Bernhardsson acknowledges the risk on the provider's side: if Modal locks in capacity at today's prices and the price of GPU falls later, the company is left holding expensive inventory. He says he doesn't expect an imminent collapse in GPU prices, but treats that possibility as part of the business's risk calculation, not as a scenario too remote to consider.

The Sandbox That Was Almost Discontinued

Modal launched sandboxes, isolated environments where an agent can run code generated by a model, after realizing that LLMs were generating code and that someone would need a safe place to run it. Initial adoption was modest enough that the team debated internally whether it was worth keeping the product alive.

What changed the game was the arrival of coding agents, background agents, and reinforcement learning training, which started demanding exactly this kind of ephemeral environment at scale. Today, according to Bernhardsson, sandboxes account for about a third of Modal's revenue, a result he says he didn't predict when the product launched.

This growth changes what infrastructure needs to deliver. A typical coding agent creates an environment, runs code, inspects the result, and tries again, often many times per task. In that cycle, the latency of creating and restoring the environment stops being an operational detail and becomes part of the experience the end user feels when using the agent.

Fast Startup as an Infrastructure Feature, Not an Accident

Elasticity only works if capacity arrives fast enough to matter. In Modal's technical material on serverless GPU, the company describes the machinery behind this: a buffer of already-available GPUs, a proprietary file system, process checkpointing, and GPU memory snapshots. Together, these pieces reduce the work needed to bring another replica of a model online.

It's a concrete example of how an infrastructure decision shows up, at the end point, as an application's responsiveness and cost. A model that takes a few extra seconds to spin up from a cold snapshot, multiplied by thousands of agent runs per day, turns into a visible line item by the end of the month.

Early in the interview, talking about why he built so much from scratch instead of using Kubernetes and Docker, Bernhardsson summed up the logic behind these engineering choices:

My thesis is like, every new sort of era of technology kind of runs up against limitations of tools, whether that's the web, the cloud, mobile, PCs, or AI. Tools exist, but they're often not particularly good, and you often need to rethink those tools in order to reap the benefits and really turbo charge and start building for this new world.

Erik Bernhardsson, CEO and cofounder of Modal

What Changes for Those Building Agents in Brazil

For those running agents or inference outside the United States, the central argument has a direct practical effect on the cost spreadsheet. Contracting reserved GPU for months, in the classic cloud model, requires nailing down in advance a usage volume that no one can predict precisely for a growing AI product. Paying for elastic usage (whether through Modal or any competitor adopting the same logic) trades that capital risk for a unit-price risk, usually higher per GPU hour, but without the long-term commitment.

The math changes especially for teams building products with coding agents or tool execution, where usage patterns come in short, unpredictable bursts rather than steady inference loads. For that profile, paying for a fixed cluster tends to mean idle GPU most of the time; the elastic pool logic was designed precisely for that case.

One caveat is worth noting: Modal's elasticity depends on aggregating demand from customers like Suno, Runway, Cognition, and Decagon, mentioned by Bernhardsson in the interview. Teams with very predictable, constant load (a mature production model running 24 hours at the same volume) may find it cheaper to negotiate reserved capacity directly with a cloud provider or neocloud, because in that case the premium paid for flexibility stops making sense.

The question that remains, according to Bernhardsson himself in the interview, is how far AI compute demand keeps growing elastically, and at what point it starts settling into more predictable patterns, similar to those of traditional backend workloads.