1 day Ago
A quirky gaggle of startups that host and rent the chips that power AI tools have gone on a fundraising spree over the last year. By Iain Martin , Forbes Staff W hen Vultr founder David Aninowsky started building his own data centers back in 2014, customers only cared about cloud compute and the processors that supported it. Graphic processing units (GPUs) were the stuff of gamer forums, not corporate boardrooms.
Now those chips are essential to the development of artificial intelligence models, and renting them has catapulted Vultr to a $3.5 billion valuation. The Florida-based startup raised $333 million from chip giant AMD and hedge fund LuminArx Capital earlier this month .
Vultr has built more than 30 data centers around the world over the last decade offering the advanced level of hosting typically sold by AWS, Google Cloud or Microsoft Azure, and renting time on GPUs has become its biggest growth driver. “AI is the fastest growing part of the infrastructure market,” JJ Kardwell, Vultr’s CEO told Forbes . It’s not just Vultr.
Investors have plowed some $20 billion over the last year into 25 companies that rent access to GPUs, according to a Forbes analysis of corporate filings and Pitchbook data. That includes $8 billion in equity and over $12 billion in debt in the form of loans from Wall Street giants like BlackRock, Carlyle and Pimco. It’s a fast emerging sector, but no one is quite sure what to call it yet: “Cloud GPUs”? “GPU factories"? Dylan Patel, analyst with SemiAnalysis, has settled on something else.
“New giants will emerge from the neoclouds ,” he told Forbes . The biggest winner so far has been Coreweave . The New Jersey-based startup had built an armada of GPUs to mine cryptocurrency but after the market crashed in 2018 pivoted to renting them to startups working on AI projects.
In the last year alone, it raised $1.75 billion of equity, and $8.1 billion of debt, attaining a $23 billion valuation.
The clutch of startups playing catch up to Coreweave are a strange crew. They include crypto refugees like Crusoe Energy, which started out mining bitcoin on gas flared from oil rigs, and German-listed miner Northern Data that got a $1.1 billion lifeline from stablecoin giant Tether to reinvent itself as an AI compute powerhouse.
The field also includes older data center outfits like Vultr and France’s OVH that pivoted from cloud compute to AI compute. Then there’s players like Nebius , which emerged from the wreckage of Yandex, the ‘Google of Russia’, with a Finnish data center, $2 billion in cash and a suspended Nasdaq listing. Its shares resumed trading in October, and it raised $700 million in a private placement from Nvidia and venture fund Accel earlier this month, to power its new business — renting GPUs.
A long tail of smaller players include Runpod, which cobbles together smaller clusters of Nvidia chips, and Fluidstack, which finds buyers for data centers with idle chips. Investors might be willing to look past some of the strange backstories of this new clutch of unicorns because of the profits generated by renting GPUs, largely by the hour. “At one point the pay back on buying a GPU was six months.
Now it’s a couple years,” Patel said. The boom for the neoclouds has not just been driven by the scarcity of AI chips but by massively undercutting giants like AWS and Oracle on price. Coreweave sells access to AI chip workhorses like the Nvidia A100 for $2.
21 an hour. AWS charges $5.12 when GPUs are hired by the hour, according to a price analysis from neocloud rival Paperspace.
Though the Amazon-owned cloud giant has steep discounts for long-term bookings. “There’s a misconception that because Amazon is the Walmart of ecommerce they must be for cloud computing,” said Vultr’s Kardwell. “They’re really the Niemann Marcus.
” Neoclouds are able to keep prices lower because they sell “bare metal” GPUs without the bundled software and services typically peddled by giants like AWS. While corporates might appreciate the concierge experience, AI startups are often just looking for the lowest price. “AI factories are not investing in the software that AWS, Google and Microsoft do to make those clouds usable for any kind of compute you need,” said chip analyst Karl Freund.
These stripped-back, smaller challengers also have an edge when it comes to getting their hands on AI’s key commodity: the GPUs themselves. Nvidia has backed Coreweave and Applied Digital while chip rival AMD has backed Vultr. That squeeze on chips has helped make tech giants like Microsoft and Oracle some of the neoclouds’ best customers.
Microsoft warned that “capacity constraints” in its data centers spelled slower growth for its AI business in its last earning calls. And so it turned to Coreweave . Ironically, the selling points of neoclouds for the so-called hyperscalers, who build airport-sized data centers, might be their ability to scale.
Coreweave has doubled its data centers to 28 from 14 over the last year; it has a $10 billion deal with Microsoft over the next five years. Crusoe Energy’s CEO Chase Lochmiller said it brought online a renewable-powered data center in less than a year when it would take a hyperscaler at least three years. And Oracle reportedly signed a $3.
4 billion deal to lease a Crusoe AI data center in Abilene, Texas. Still, there are challenges ahead. Fundraising is one.
Building and running data centers is another, particularly for small teams whose bitcoin mining acumen might not translate to managing finicky GPUs and sensitive customer data. “It’s like building an industrial warehouse versus a five star hotel; they are very different things,” warns Lochmiller. There are also hints that demand for AI chips is softening.
Some startups have slashed prices by up to a third since September, while others have taken to posting “GPUs to rent” ads on AI investors Nat Friedman and Daniel Gross’ Craiglist-style listings page . With a new crop of energy-intensive but powerful Nvidia chips headed to market, the bigger, more established neoclouds may well end up with a winning hand. That’s less certain for the smaller players.
For them, the combination of debt-fuelled expansion and a potential chip glut exacerbated by the expiration of multi-year contracts at a time when more efficient chips are becoming commonplace could be a problem. Said Patel, “Let’s be clear, I think a good chunk of these neoclouds are going to go bankrupt.”.
