Brian Venturo and a couple of fellow hedge fund buddies bought their first GPUs as part of an elaborate joke. It was late 2016, and for the fun of playing with something volatile that could be either trash or treasure, they had been staking their pool and fantasy football games with bitcoin instead of cash. When bitcoin prices surged, Venturo says, the group realized, “Maybe we should take this seriously.”
The friends rush-ordered GPUs, also known as graphics chips, for same-day delivery from Amazon to start mining bitcoin while their enthusiasm ran hot. The powerful processors are prized for their ability to crunch the math of crypto-mining at high speeds. Over time, the hobby turned into a business called Atlantic Crypto, which filled a garage and later warehouses with GPUs. Some of the beefy chips were rented out via the cloud to armchair cryptocurrency miners. As coin prices climbed, time was money, so the company’s technicians grew skilled at speedily installing GPUs and pushing their performance to the brink.
When crypto prices cratered heading into 2019, Venturo and team pivoted. They raised money to buy cut-price GPUs from struggling crypto miners, rebranded as CoreWeave, and offered up their tens of thousands GPUs as a specialist cloud service for companies in need of the powerful chips, like visual effects studios and AI startups. GPUs are essential to machine learning projects like training algorithms to process images or text. CoreWeave was unknown in those spheres, but won over clients by dropping everything to address support queries. “We would resolve things in 15 minutes,” a manager from that era says. “The dedication to customers was phenomenal.”
Everything changed after OpenAI debuted ChatGPT late in 2022. A rush of interest and investment in all things generative AI caused a GPU shortage—especially of Nvidia’s prized chips. With its couple of data centers packed with GPUs and a long-standing relationship with Nvidia, CoreWeave quickly became vital to the accelerating AI hype train. The company picked up clients including Microsoft, OpenAI’s crucial sponsor, and image-generation startup Stability AI, according to former CoreWeave employees. Venturo’s biggest regret is not ordering more Nvidia chips back when he could. “If I had unlimited amounts, I could have it all sold in two weeks,” he says.
Things still worked out. As demand for CoreWeave’s GPUs surged through 2023, the company piled its new revenue into buying more of the chips and building out new data centers. A company that began as a beer-money side project now doesn’t touch crypto, and is all in on cloud computing. Over the past year the startup has picked up a $2.3 billion loan—using GPUs as collateral—and nearly $2.2 billion in investment funding, including $1.1 billion announced this week, from investors including Nvidia. With its latest infusion of cash, the startup’s valuation jumped to $19 billion.
By the end of this year, CoreWeave expects to have 28 data center campuses spread across a dozen or more US states, Europe, and potentially Asia—overall more than nine times its footprint at the beginning of 2023. Venturo, who is now CoreWeave’s chief strategy officer, says long-term contracts to supply AI giants justify the buildout. “It’s transformed this small cloud startup into one of the biggest juggernauts,” he says. “We’re lucky enough to be in this position because of this crazy origin story where we had so much expertise building at this rapid, frenetic pace.”
Lauren Goode
Matt Simon
Tess Owen
Charlie Wood
For some of CoreWeave’s 550 workers, that pace has translated into a culture of grueling hours and high-pressure management that feels unsustainable, according to six former workers speaking on the condition of anonymity to avoid retaliation. But with the AI boom continuing, CoreWeave doesn’t look likely to slow down.
GPUs are crucial for operating and training complex machine learning systems such as OpenAI’s GPT-4. When startup Databricks recently set out to train an open source AI model that might challenge GPT-4, the project monopolized 3,072 Nvidia H100 chips, costing $10 million over three months. These expenses are rapidly consuming the investments received by generative AI companies last year, which PitchBook estimates at $29.1 billion.
In the quest to secure GPU resources, many are following in the footsteps of CoreWeave’s successful 2023 strategy. Cryptocurrency miners including Applied Digital, Hut 8 Mining, and Iris Energy are transitioning into AI by offering their surplus GPU capacity to software developers. As the monetary rewards for bitcoin mining decrease, more GPUs are likely to be allocated to AI research.
Leading cloud entities are heavily investing in expanding their AI-focused chip offerings in the cloud, including GPUs. Amazon envisions spending nearly $150 billion on data centers in the following 15 years; Microsoft discusses tripling its own data center expansion later this year; and Google recently publicized a $3 billion investment to advance and develop two U.S. data centers.
Among only seven select cloud partners recognized by Nvidia as “elite” within the U.S., CoreWeave stands unique and promising, despite cloud behemoths like Amazon not featuring on this exclusive list. This positioning might provide CoreWeave with an upper hand in obtaining GPUs in the face of their scarcity. Its relentless drive and adaptability appear to be its strongest assets, resulting in minimal obstacles it hasn’t yet found a way around. “We survived due to our hustle and our creativity,” notes Venturo.
Lauren Goode
Matt Simon
Tess Owen
Charlie Wood
Fulfilling the world’s surging demand for AI chatbots and image generators depends on mundane components as much as glitzy GPUs. Server cabinets, heavy-duty metal enclosures to store GPU systems, have at times been a crucial bottleneck.
As it scrambled to build up its facilities, CoreWeave once ordered 1,400 of the wrong cabinets. It was a costly mistake because supply chain backups have delayed new shipments by months. “You’re moving so fast, and somewhere along the line a process fails—and you don’t realize until you have 17 tractor trailers with cabinets outside the door and you have to turn them all away,” Venturo says.
But in an example of the shrewdness that’s been key to rapid expansion, Venturo’s team in that crisis quickly put aside frustration and decided to buy used cabinets off what he called the gray market. The move prevented a significant delay. “This was just one of the instances of challenges we faced and overcame to make sure we delivered for our partners,” he says.
To keep things moving, CoreWeave has turned to the “gray market” for more than just cabinets. It’s bought networking switches and routers from eBay to sidestep waits as long as two years for new gear, a former employee says. The security and reliability of used parts can be questionable, but amid the urgency of the AI boom, some conventional practices had to be brushed aside, the person says. Venturo says CoreWeave doesn’t turn to eBay anymore.
In Plano, CoreWeave successfully equipped four 1-megawatt data center halls in less than three days each, a task that usually requires weeks. Venturo remarks that their approach allowed them to build swiftly and efficiently.
Another instance of CoreWeave’s inventive problem-solving emerged when an internet provider lagged in setting up a broadband connection at a new site. To circumvent potential project delays, three senior executives convened over a phone call and agreed upon a single solution: securing satellite internet via Elon Musk’s SpaceX’s Starlink service, a convenient but pricey alternative until the fiber provider could establish a connection. Such a move negated weeks of potential holdup. Venturo emphasizes their necessity for flexibility.
Their early projects served as valuable lessons, shaping CoreWeave’s current standard procedures. They opted to pay extra for customized manufacturing of numerous fiber-optic cables as its unique design allows for an installation time of just an hour, compared to the standard ten. After encountering a delay with an essential shipment via US customs, the company started routing orders through multiple alternate ports. They now order surplus parts as a contingency measure, expecting the excess can be shipped to any upcoming project that may need them.
However, this urgency has sometimes resulted in unforeseen drawbacks. For instance, CoreWeave’s data center in Las Vegas reportedly still has a lingering smell of burning plastic. This issue occurred when the company overloaded with GPUs upon establishment, causing some electrical components to burn out.
Lauren Goode
Matt Simon
Tess Owen
Charlie Wood
CoreWeave’s operations heavily rely on its data center technicians, especially the highly skilled ones who function much like a special operations unit. These technicians are often on the move, setting up new data centers and are not confined to a single site. Although the CEO, Venturo, does not disclose the distances these technicians have travelled, he shares that around 6,000 miles of fiber-optic cabling were installed last year. He works closely with this team due to their integral role in the company.
However, a few ex-employees of CoreWeave have expressed that the work environment can be unnecessarily stressful. The employees are expected to be constantly available with their personal phone numbers displayed on their company Slack profiles. One ex-employee confessed that this led to receiving texts regarding work issues at odd hours.
To alleviate such situations, a few years back, Venturo introduced weekly office massages for the employees at the headquarters in New Jersey. The initiative began when he noticed his recurring back and neck aches. Though one ex-employee revealed that only executives felt comfortable availing this service.
Venturo believes that the high demands on the workers are not baseless. The company aims at identifying and eliminating obstacles. The teams are encouraged to work quickly and when issues arise, they are resolved collectively.
Tight staffing and equipment supplies have sometimes affected the reliability of CoreWeave’s services as it has grown. The number of server outages started to increase as it shifted away from crypto because CoreWeave until recently had just one engineer focused on maintaining uptime, former employees say. The company sometimes didn’t have enough working GPUs to fulfill contract obligations, according to two of the sources. “They never had an appropriate number of spares,” one says. To compensate, the company skipped some testing of newly installed GPUs, the two sources say. “We don’t have another choice I think,” a vice president said in one internal Slack message seen by WIRED. Venturo says the company’s testing platform has been designed to ensure clients meet their timelines to get online.
CoreWeave’s fast pace has forced some considerations—like accounting for the company’s environmental impact—down the priority list, according to former employees. Venturo says most of CoreWeave’s data centers are operating on 100 percent renewable energy and that the company hasn’t pursued data centers in areas such as floodplains that aren’t insurable. “It’s not so frenetic that we’re put in a position where we have to make irrational choices,” he says. Asked whether CoreWeave would release a sustainability report, as some other cloud companies do, to report its water usage and efforts to reduce supply-chain emissions, Venturo told WIRED in October that it would have more to share soon. It’s yet to release any report.
If you’ve ever used ChatGPT, Microsoft’s Copilot, or other generative AI offerings, or image creator Stable Diffusion, you may have reaped the fruits of CoreWeave’s frenetic labors.
The company has put together facilities for Microsoft, with input from OpenAI, sources say, and is hosting a supercomputer data center for Nvidia in Plano, Texas, according to the chipmaker. CoreWeave’s Plano campus cost about $1.6 billion and runs about 450,000 square feet, around the size of the largest Ikea store in the US. Another space dedicated to a single customer is in Oregon, where it has served business chatbot software startup Inflection AI. These sites tend to have somewhere between 16,000 to 32,000 GPUs, Venturo says.
Lauren Goode
Matt Simon
Tess Owen
Charlie Wood
CoreWeave’s other locations provide GPU access to multiple companies. Customers include Stability AI, a developer of image generators, and Anlatan, a provider of the writing tool NovelAI. These sites serve a broad range of uses by providing access to fairly priced chips, like Nvidia A100s launched in 2020. These sites are often located on the outskirts of key cities—Venturo refers to them as places that host National Football League teams—to reduce delays in delivering customers their AI-generated content. “We’re trying to develop our cloud to be as flexible and dynamic as our customers demand,” says Venturo.
Different CoreWeave locations cover new construction, revival of empty buildings, and leasing fully equipped data centers from others, including a bitcoin miner. Such a nimble strategy sets CoreWeave in a favorable business position. Raul Martynek, CEO of DataBank, estimates that companies like CoreWeave, which offer GPU access, may have a gross profit margin of about 70 percent by avoiding the costs of establishing their own facilities. Venturo doesn’t comment on gross margins.
Major cloud providers, like Amazon and Google, could potentially force smaller competitors like CoreWeave out of the market by offering lower prices once demands for GPU time stabilize. But Tony Harvey, a senior director analyst overseeing data centers at Gartner, believes a niche provider like CoreWeave might survive. “It’s an interesting game to play, and they’re going to have to work hard at it,” he says.
A more immediate concern for CoreWeave and the entire data center industry is that growth might be hindered by power shortages caused by factors including a lack of new power and transmission construction. Currently, American campuses of cloud giants like Microsoft consume about 11 gigawatts of power. Smaller players like CoreWeave consume another 11 gigawatts, according to JLL, a real estate firm. It projects usage to grow about 20 percent yearly through 2030 to an estimated combined total of 79 gigawatts.
CoreWeave for its part hasn’t developed a data center in the popular region of Northern Virginia because it’s such “a food fight to get power” there, Venturo says. But the company could be just one fresh, clever idea away from getting ahead of the pack.
Updated 5-3-2024, 5:05 pm EDT: This article has been updated with additional comment from CoreWeave.
