Despite CoreWeave’s relentless pursuit of compute, questions remain about AI

When CoreWeave expanded into AI data centers in 2019, venture capitalists weren’t interested in funding the idea of such facilities purposely built for AI. “They were like ‘this business doesn’t make sense. You’re not hardware. You’re not software. We can’t do it,’” Venturo said.

It seemed unlikely that a New Jersey crypto mining company founded in 2017 by a group of commodities traders could beat the cloud providers at their own game.

But the experience with crypto mining, where GPUs are run inside dusty data centers that leave GPUs caked in dirt, were a catalyst for what became CoreWeave’s secret sauce. “You learn how to deal with failures at scale. And you learn how to optimize things,” Venturo said. “If I can run a GPU like a chicken coop in North Carolina, it’s not that hard to run it in a tier four data center.”

Building AI data centers was a relatively new field, and one that CoreWeave’s team initially knew nothing about. But the cloud companies were in the same boat. “The cloud, as it had been configured, wasn’t really the best way to approach building compute for this new use case,” Intrator said.

Large language models like ChatGPT, which were in their infancy at the time, were trained using thousands of GPUs running simultaneously as if they were one computer. If even one of those GPUs failed during a training run, the entire operation would be ruined.

As AI models increased in size, the risk of a GPU failure carried even higher stakes. Training runs could take weeks or months and cost hundreds of millions of dollars.

CoreWeave’s chicken coop days gave it a good foundation for that skillset. It quickly developed a system that could check GPUs for signs that they might fail, pulling them out of service ahead of time, before they could ruin a training run.

The reliability of CoreWeave’s data centers became a selling point in the AI world, and business was picking up. Microsoft, which had made a large investment in OpenAI, hired CoreWeave to provide compute capacity to Sam Altman’s firm.

Then, in November 2022, OpenAI released ChatGPT. Practically overnight, demand for AI compute skyrocketed. Nvidia’s most powerful chips became one of the hottest commodities on the planet, sending its stock soaring.

CoreWeave, which had about 50 employees at the time, was perfectly positioned to capitalize on the opportunity, but it needed an infusion of cash.

Not only did it know how to build AI data centers, it had a unique window into the landscape of AI companies already using high-performance compute. In the meeting with Nvidia at the end of 2022, Venturo said Huang wanted to know about where CoreWeave was seeing growth. “We were almost the front-line view for him of different use cases coming in,” he said.

Though it had Nvidia’s investment, CoreWeave also took on debt to fund the capital costs of building out data centers, secured by long term contracts with companies like Microsoft.

The debt was a major departure from the tech company blueprint of the last two decades. With ample venture capital available, most startups prefer to stay private as long as possible.

From CoreWeave’s perspective, debt made a lot of sense. Compute power was becoming a scarce resource and it had already secured contracts with customers. Once the data centers were built, the company would be able to pay off the debt.

But ahead of the company’s IPO, skepticism mounted about CoreWeave’s unorthodox business model. Critics questioned whether it was using the Nvidia investment to buy more Nvidia GPUs.

In March, news leaked that Microsoft had cancelled a contract with CoreWeave because of falling demand for AI compute.

But according to people familiar with the deal, Microsoft’s decision centered on a rift with OpenAI and internal disagreements within Microsoft related to resources for a new team run by Mustafa Suleyman, who had left Inflection AI to lead Microsoft’s internal AI efforts.

As soon as Microsoft cancelled the deal, OpenAI immediately snatched it up for $11 billion, giving CoreWeave a public vote of confidence.

CoreWeave’s IPO was more successful than even its biggest backers had expected. The company’s valuation hit roughly $100 billion in June, from $18 billion at its IPO.

Since that peak, CoreWeave’s stock price has steadily dropped, in part because of the end of the lockup period that allowed early investors to sell their shares.

But the downturn also reflects a broader skepticism in the market about AI. New AI model releases, like OpenAI’s GPT-5, have shown only marginal gains in capability. New studies show AI may not be giving companies the productivity gains they initially thought.

When Nvidia reports earnings on Wednesday, the market reaction will be a sign on how deep that skepticism goes.

Whether the AI boom is a bubble about to burst depends on what you expected. If you thought superintelligence was going to arrive in 2025 or any time in the next handful of years, then you’re in for major disappointment (or celebrating, depending on your view). If you thought AI was going to upend the economy overnight and lead to massive unemployment, you can breathe a sigh of relief.

But that doesn’t mean an end to massive investment in AI infrastructure or a drop in demand for compute resources.

From the coders burning tokens by the billions to the army of startups using generative AI to build products for large companies to Hollywood studios churning out special effects, there is just no going back.

For every novel use of AI today, there will soon be 20 more that nobody saw coming.

And yet, it doesn’t feel that way. The world has been chasing the high of the ChatGPT moment for close to three years now. But if you zoom out, ChatGPT was not so much a technological breakthrough as it was a cultural realization of the power of neural networks, which came from decades of academic research and experimentation at large tech companies.

In other words, AI isn’t just a parlor trick that led to a bubble. It’s a wave that’s been building for a long time and I wouldn’t bet against it.

Short seller Jim Chanos has accused CoreWeave of an accounting trick by estimating the useful life of Nvidia GPUs at 7.5 years. If it turned out the chips were worthless after a few years, the company would look a lot less attractive to investors.