Picture a typical contract for AI chips on a chart, with the number of chips on one axis and time on the other. You get a rectangle: a fixed block of graphics processors, reserved for three years and often starting three to six months after the deal is signed. Erik Bernhardsson, cofounder and CEO of the AI infrastructure company Modal, has argued that this shape is the biggest downside of today's GPU shortage, bigger than the difficulty of finding chips at all. One of his posts on the subject came up on the Latent Space podcast. It said a "three year rectangle" is not what demand looks like for most users.
Modal's business depends on that mismatch. The company signs the long, rigid contracts itself and rents the capacity out to customers in small pieces that can grow or shrink quickly. On October 14, 2026, Modal announced a $750 million Series D round led by Accel at a $15.75 billion post-money valuation. It also said it had passed $1 billion in annualized run rate, a company measure that projects its current revenue pace over a full year. The announcement says usage tripled in the five months since the company's Series C and grew more than fifteenfold year over year. In the episode, published October 22, Bernhardsson said he did not remember the exact figures from his previous appearance two years earlier, but guessed Modal was "at least like 50x" bigger in revenue, users and usage.
What Modal sells
GPUs, or graphics processing units, are the chips that do most AI computation. Much of what Modal's customers do with them is inference: running a trained model to answer requests, generate images or write code. Bernhardsson described Modal as a compute and storage platform with a software kit that lets companies such as Decagon, Suno, Runway and Cognition build AI applications while Modal manages "all the capacity underneath."
He said customers rarely need a flat block of chips. Their workloads "scale up and down throughout the 24-hour cycle," grow over time and have burst periods. A company that buys enough GPUs for its peak pays for idle hardware the rest of the time, and one that buys less risks missing demand. Suno, the AI music company, is a case in point. In a 2024 customer story, Modal describes song requests jumping around Christmas and Valentine's Day, with Suno's GPU use scaling into the thousands. The same account says Suno's founders wanted to avoid committing to three-year GPU reservations.
Modal's answer is to pool demand. Bernhardsson said that because the company has "many, many thousands of customers," their individual spikes and dips average out: "the demand gets very smooth in aggregate." A smooth total is something Modal can cover with long contracts, while each customer gets on-demand access. If a customer needs 100 GPUs, he said, Modal can "typically give them that, like, immediately," without a call to a salesperson. He called this, loosely, a "duration transformation": long commitments go in on one side, and short, flexible use comes out on the other. Asked whether that sets Modal apart, he said it is a "massive differentiator."
He split Modal's value into two parts. The first is developer experience, meaning people can build things without thinking about infrastructure. The second is capacity. "We're aggregating the world's compute capacity," he said, and Modal is "very good at slicing it out, like, very granularly" because it has invested in scaling up and down quickly.
Other companies attack the same rectangle in other ways. SF Compute, for example, lets buyers order compute time with a set start and duration and offer unused hours for resale to other buyers. That makes reserved capacity tradeable. Modal keeps the long contracts on its own books and sells usable slices of them.
Where the chips come from
Bernhardsson said the way Modal buys capacity has changed a lot. For its first few years, almost all of it came from the big cloud providers: AWS at first, then Google Cloud and, over time, Azure, plus Oracle early on because Modal wanted A10 and A100 chips and got good rates there. Most of it was spot capacity, meaning spare cloud hardware rented on the spot that the provider can take back. That arrangement, he said, "sort of works in aggregate" because Modal drew on hundreds of regions at once. Until what he called the end of last year, most of its capacity was spot or one-year reservations.
That stopped being enough as Modal grew and the cloud market changed. The company now works with "almost 25 clouds," Bernhardsson said, and often has to sign deals of three years or more, paying "a meaningful amount of money up front" and locking in prices for the full term. Modal's own announcement puts the count at more than twenty clouds. To manage the shift, he said, Modal set up a procurement desk of four people who work full time calling neoclouds, the newer providers that specialize in renting out GPUs, doing deals and vetting capacity. A separate team of about 10 engineers manages bare-metal machines around the world, dealing with network problems, bad drives and "GPUs falling off the bus."
Software built for one global pool
Asked why Modal had to build its own autoscaling tools, Bernhardsson said every new era of computing runs into the limits of the old tools. Kubernetes and Docker, the standard software for packaging and running cloud applications, were built for steady backend workloads with an even stream of requests. Those systems often run at low utilization because CPUs are cheap. AI is different: GPUs are expensive, and large models have to be loaded over the network onto the chips. So Modal built its own file system, CPU and GPU snapshotting, autoscaler and scheduler. He said you "can't build like a wrapper on top of the old stuff."
One early choice shaped the rest. Modal went multi-region very early, he said, and treated its capacity as "one big pool over the entire world" rather than separate clusters in each region. He called that "a fateful decision in hindsight."
Modal's engineers have described how that pool is managed. A 2025 post explains a service that uses linear programming, a standard optimization method, to decide which machines to add or drop as cloud prices and availability change. It weighs customer demand, acceptable GPU types, region limits and a reserve of spare capacity kept for fast starts. One example in the post is buying H200 chips for less than the going price of older H100s. A 2026 engineering post covers the other half: making a new copy of a model ready to serve quickly. It does this with a buffer of idle GPUs, lazily loaded files, and restoring saved CPU and GPU state instead of initializing from scratch. In more than 10,000 test starts with a small Qwen model, mean startup fell from about 96 seconds to about 14 with the vLLM serving engine, and from about 84 to about 17 with SGLang. The post notes that snapshots do not remove the cost of loading model weights, and that GPU snapshots worked only for single-GPU workloads at the time.
Small blocks in odd places
The conversation turned to how Modal competes for chips with OpenAI and Anthropic, which would seemingly buy everything available. Bernhardsson said rising prices partly reflect that buying. He thinks the big labs prefer "tried and true" suppliers, partly because their training runs need very large blocks of chips. Modal's pool can absorb scattered leftovers instead. "You can get 32 nodes in Armenia, you can get 128 nodes in Indonesia," he said. "We love that kind of shit."
He added that NVIDIA, the chipmaker, seems to prefer spreading its allocation across many neoclouds and alternative clouds rather than a few huge buyers. He described what follows from that as "more speculation": smaller, riskier clouds have become more of an opportunity for Modal.
Vetting a warehouse full of GPUs
Asked whether some of these suppliers are just a "janky warehouse," Bernhardsson said that sometimes it feels that way. Modal sends each vendor a spec sheet covering storage speed, CPU benchmark scores and network bandwidth. It usually runs acceptance tests on a single node first. Even so, he said, the process involves "a little bit of a leap of faith": Modal sometimes signs commitments of "tens or hundreds of millions of dollars" with a vendor it still does not fully know.
He compared Modal's experience-based judgment to ClusterMAX, the GPU cloud rating from the research firm SemiAnalysis, saying Modal almost has its own "Modal max" score internally. Vendors with good reliability get more deals, and so do vendors that respond quickly to support requests and fix outages fast. SemiAnalysis's latest ClusterMAX report covers 77 providers and makes a similar point. It tests reliability by injecting hardware errors and timing the recovery, because fast benchmark runs alone say little about how a provider operates or supports customers.
Functions, and the pull toward selling rectangles
Bernhardsson said Modal has two main products. Sandboxes are isolated environments for running code, many of them for AI agents. Modal Functions, launched in 2021, let developers deploy code as Python functions that Modal runs in containers, scaling them up with demand and down to zero when idle. Functions found product-market fit in 2023 and now make up about two-thirds of revenue, he said, "mostly inference."
Modal has never had a dedicated GPU product as such, and Bernhardsson said he never thought of it as a neocloud. He still sees "a risk of getting sucked in that direction": selling customers a fixed number of GPUs for three years. In his view, Modal's strength is the opposite case, where a customer deploys something "very cyclical or, like, bursty" and Modal manages the capacity.
Moving chips between models
Bernhardsson applies the same pooling logic to AI models themselves. Another of his posts, read during the interview, argued that inference providers have a structural reason to exist separately from model builders: the best model keeps changing, and a provider with smooth overall demand can move resources between models.
He pointed to open models from Chinese labs. Each time a lab releases a stronger model, demand shifts toward it, though not overnight, because prompts and products built on older models create "a certain stickiness." A lab holding its own capacity would see demand surge after a launch and fall when a rival released something better. Modal, he said, might run several thousand GPUs for one model today and need to move half of them to another tomorrow. "No one even needs to tell us to do that," he said. "It just magically does that."
That does not make the model builders commodities, in his view. Asked whether it does, he reached for Taylor Swift, while saying he was not sure it was the best analogy. She releases a big album and tours, has a quieter year while she records, then releases another, and her brand lets her keep a lot of margin. Labs could work the same way, he said, as long as they keep launching new models and earn enough on each one to pay off the cost of training it.