Erik Bernhardsson, chief executive of the AI infrastructure company Modal, keeps buying computing power. "I'm extremely aggressive, like buying capacity every day and like raising money to buy," he said. What worries him more than a shortage is a collapse in prices.
Bernhardsson explained his reasoning on Latent Space in an episode published October 22, 2026. Modal rents graphics processors (GPUs), the chips used to train and run AI models, and sells their time to its customers. The publisher's episode notes say Modal reports it has passed $1 billion in annualized run rate and has raised a $750 million Series D at a $16 billion valuation. Modal's business is pooling customers' changing demand on top of fixed, long-term GPU contracts. Here Bernhardsson dwelt on the financial risk of holding those contracts.
Renting rather than owning
"We don't own capacity," Bernhardsson said. Modal rents all of its GPUs, typically on three-year deals. He said he is not "mega dogmatic" about this. Renting is capital efficient. Buying brings operational work: vendors, suppliers, power and cables. It also raises the question of how to pay. A company can raise a billion dollars in equity and buy hardware, but he called that a very dilutive way to do it, because selling new shares shrinks existing owners' stakes. Debt is hard too: to cover the interest, he said, a borrower usually wants "locked in long-term revenue."
His preference is to remain "a very lean software company." He compared hardware to an office: most software companies, he said, shouldn't own one. Still, "in order to get the capacity we want, like, you know, everything's on the table." Securing the capacity Modal needs for 2027, he said, will take work down "a lot of different avenues." In the long run Modal might own some hardware, rent some, or try joint ventures or complex SPVs, the special purpose vehicles that hold an asset and its financing apart from the parent company.
Asked whether this meant hiring a lot of finance people, Bernhardsson said he has had to spend a lot of time on the problem himself. He called the overall picture "kind of staggering": in his view AI investment is becoming several percentage points of US GDP, and US interest rates had been "shooting up in the last few weeks." He described neoclouds, the newer cloud companies built around renting out GPUs, as "very, very sophisticated financial engineering companies," and said Modal might have to do similar things.
Why rising prices don't scare him
Asked whether supply would stay constrained for a while, leaving buyers no choice, Bernhardsson disagreed. By his account NVIDIA is delivering several hundred thousand of its Blackwell chips a month, and anyone willing to outbid others can get capacity, even with Anthropic and OpenAI buying as much as they can. "I don't necessarily think it's like a fully cornered resource," he said.
He did say prices are rising, and that this is unusual. The normal state of hardware is "kind of like gravity," he said: eventually prices always fall. The current rise has put the market in "a weird sort of period of time," and he could not say how long it would last: "could be another six months, could be another 36 months." He doubted that companies really must rush to buy, comparing the impulse to "the great toilet paper shortage" of the pandemic, or a run on a bank.
Higher prices don't hurt Modal much, he said. Scarcity lets Modal charge customers more, and its reservations are locked in at fixed rates, which makes the company "effectively like long GPU prices": in trading terms, it gains when prices rise. "I do worry more about actually prices collapsing," he said. He doesn't expect that soon, but a collapse would hurt anyone who has locked in a lot of supply, because they would have paid too much. "There's a risk of buying at the top."
He put it in economic terms. Tight capacity creates a "scarcity rent," extra profit that goes to whoever controls the scarce thing. He said it goes mostly to NVIDIA and partly to neoclouds, and that anyone who owns capacity has better economics today. Over the very long run he doubts that advantage fully holds. Owning GPUs may always earn a small premium, he said, but as for whether it is worth holding such expensive assets on the balance sheet, "I'm not so sure."
Why reselling won't save you
The conversation then turned to a question many buyers ask: if demand for intelligence is basically infinite, why not buy as much as possible and sell whatever you don't need?
Bernhardsson said Modal couldn't use everything it might buy today. The resale plan fails because of correlation. If something goes wrong in the market, prices fall, utilization falls, and the market for reselling capacity "vanishes" at roughly the same moment. "I'm very bullish," he said, "but like, I do want to point out and like, you know, like, there are tail risks here. And when those tail risks happen, like, it could get very bad."
Part of his caution, he said, may be age. He had his first job in 1999, lived through the dot-com crash and the 2008 financial crisis, and wants Modal to be around for "20, 30, 50 more years." Modal could buy "an insane amount of capacity" and try to corner the market, he said. Then he brought up the Hunt brothers, "a legendary story in trading": they tried to corner the silver market, and "it failed spectacularly." His plan is to keep ramping capacity aggressively, at a pace that "basically matches our demand growth" and Modal's run-rate goal for the end of next year.
How neoclouds pre-sell what they haven't bought
Asked what Modal could learn from neoclouds such as Nebius, Bernhardsson described how they grow so fast. A neocloud goes to Anthropic, OpenAI or another large buyer and offers, say, five years of chips at a good rate. With the signed contract it goes to Wall Street, shows the commitment, perhaps $10 billion worth of chips, and raises the money. It then buys hardware from Supermicro, Dell or NVIDIA, installs it and delivers. "You can basically sort of pre-sell things that you haven't even bought," he said.
He called Nscale "the most crazy example." In his loose description, it went from nowhere to hundreds of billions of dollars in its pipeline while barely having any capacity. He called CoreWeave and the other big neoclouds "very, very savvy" about raising money and using GPUs as collateral for loans. CoreWeave's 2026 proxy statement shows what that looks like. A CoreWeave subsidiary took out a $2.3 billion delayed-draw loan to finance GPU servers and infrastructure for customer contracts, secured, with exceptions, by substantially all of that subsidiary's assets. Under a later facility, how much the company could borrow was limited by the purchase price and depreciable cost of the GPU servers.
Modal can't easily copy this, Bernhardsson said. Its growth has been product-led, with developers signing up and paying for what they use, so it lacks the same big long-term contracts to borrow against. That makes it harder to scale as fast, though he expects Modal to "get to the same point" eventually.
Can GPU time trade like oil?
The discussion moved to efforts to make GPUs a financial asset class. Bernhardsson named startups such as Silicon Data, Compute Exchange and Ornn that are building price indexes and trading derivatives. The difficulty, he said, is that compute is "definitely not as homogenous as other asset classes": networking, location, disks, CPUs and reliability all vary. "That being said, like, you can say the same thing about oil," which markets still turned into tradable commodities.
Silicon Data's H100 rental index addresses the problem he described. It publishes a daily standardized price per GPU-hour, adjusts observations for specifications, rental terms, cluster size, interconnect, performance and geography, and reports hyperscaler prices separately from neocloud and marketplace prices.
Bernhardsson said he is glad people are working on this, because at today's scale the industry needs to get smarter about hedging and sharing risk, including some form of insurance. He is less sure GPUs will become as commoditized as Henry Hub natural gas or Brent crude, two benchmark energy prices. The episode notes contrast him with Evan Conrad of SF Compute, whom Latent Space interviewed separately: Conrad focuses on making capacity tradeable, Bernhardsson on making it usable in smaller, elastic slices.
Looking beyond NVIDIA
Asked whether Modal buys only NVIDIA hardware, Bernhardsson said it is looking at alternative accelerators. For five or maybe 10 years, he admitted, he has predicted that next year would be the year of non-NVIDIA chips, "and then I'm always wrong." This time he said "2027 might be the year of non-NVIDIA accelerators, finally."
He said AMD chips are getting good scores on some tasks and Google's TPUs seem to work well inside Google. He knows less about Amazon's Trainium but said the team is very good. Hardware is hard and software is easier, he said, so fixing the software layer for these chips could unlock a lot of demand. Deeper bottlenecks remain at the chipmaker TSMC and the lithography-equipment maker ASML. "Monopolies don't always last," he said, though he added that he loves NVIDIA and called it an amazing company.
He also mentioned the chipmakers Cerebras and Groq, and setups that split the two stages of answering a prompt across different hardware. In the first stage, called prefill, a model processes the whole prompt at once, which takes heavy computation. In the second, decode, it produces the answer token by token, a stage limited mainly by memory bandwidth. Bernhardsson described people running NVIDIA for prefill and Cerebras for decode. Cerebras's own announced example pairs Amazon's Trainium for prefill with Cerebras hardware for decode. He found these mixed setups exciting, but said "it starts to get very complex when you're deploying these things in practice."
Where the cash goes
Asked how Modal would spend new money, Bernhardsson said much of the cash it needs goes to capacity prepayments, the upfront payments that lock in GPU supply. To him that is "the best use of cash today." Whatever deployment model Modal chooses later, he expects some kind of down payment to remain. Salaries and offices are "basically a rounding error" in comparison. Modal has 200 employees who "cost a lot of money," he said, "but, you know, the GPUs cost a lot more money."
Asked about the company's top priorities, he named "capacity and the sort of financial complexity" first: pricing and packaging, the tension between owning and renting, and a question he has not settled. "Should we go in a more NeoCloud direction?"