CoreWeave, the largest of a new breed of companies driving the artificial-intelligence boom, has watched $33 billion of value vaporize in six weeks.
The share price plunge comes as investors worry about a possible AI bubble, the fallout from a failed merger and public criticism from high-profile short seller Jim Chanos, known for predicting the collapse of Enron.
But some of the high-tech company’s biggest problems began with a very low-tech nuisance: unexpectedly turbulent rainstorms in North Texas.
Over the summer, heavy rains and winds caused a roughly 60-day delay at a construction site in Denton, a small city north of Dallas, preventing contractors from pouring concrete for a major AI data-center complex, according to people familiar with the matter.
As a result, the completion date for the huge data-center cluster, consisting of about 260 megawatts of computing power that CoreWeave plans to lease to OpenAI, has been pushed back several months. There were additional delays caused by revisions to design plans for some of the data centers a partner is building for CoreWeave in Texas and elsewhere, according to filings.
Share-price performance since the end of September
The slowdown was compounded by mixed messaging from CoreWeave’s chief executive, Michael Intrator, which spooked investors and accelerated the company’s share-price decline at a particularly vulnerable moment for the AI trade.
CoreWeave’s business model involves using high-interest debt to buy thousands of advanced AI chips from Nvidia, installing them in server racks inside data centers that it leases from third-party landlords, then renting access to the chips to AI companies.
As capital spending on AI infrastructure has intensified, CoreWeave, which is 7% owned by Nvidia and backed by hedge funds such as Magnetar Capital and Coatue Management, has become the standard-bearer for both the promise and the risk of the AI boom.
Some critics point to the high levels of debt it has taken on to finance its data-center build-out, while others worry that the company depends on just a handful of large customers, such as OpenAI, Microsoft and Meta, for the bulk of its revenue. CoreWeave saw sales more than double in the most recent quarter to nearly $1.4 billion, from $583 million a year earlier, but the company is unprofitable and lost $110 million in its most recent quarter.
In early November, before the construction delays were widely discussed, Intrator played down fears of an AI bubble at a Wall Street Journal event in Northern California.
“If you’re building something that accelerates the economy and has fundamental value to the world, the world will find ways to finance an enormous amount of business,” he said, adding that the high number of buyers of data center computing services had convinced him that there is not a bubble inflating.
One of those major buyers is OpenAI, the company behind popular AI models ChatGPT and Sora. CoreWeave plans to rent the chips at its Denton data-center cluster to OpenAI once it is built. The builder is Core Scientific, a former crypto-mining company based in Austin that has emerged as one of CoreWeave’s biggest landlords.
On Nov. 10, Intrator spoke to investors during a quarterly earnings call and delivered a confusing, contradictory message about the construction problems.
At one point in the call, he attempted to quash a string of questions about the delays and their impact. “There was a problem at one data center that’s impacting us, but there are 32 data centers in our portfolio, all of them are progressing to one extent or another,” he said.
He said that “this one data center will catch up, and then we will move forward from there.” That statement was inaccurate, however, and Intrator’s chief financial officer, Nitin Agrawal, quickly corrected the CEO and said that the delays were concentrated at “one data-center provider,” rather than just one singular data center, suggesting a more widespread problem.
At another point, Intrator described construction delays as “systemic challenges” that are “very frustrating for our clients,” and said the company was trying to diversify its supply base of data-center builders to soften the impact of inevitable delays.
The comments unnerved investors. The next day, Intrator gave an interview to CNBC’s Jim Cramer, once again repeating the “one data center” message before correcting himself after being prompted by the host. CoreWeave’s stock, which hit $105.61 the day of its earnings report, fell by 16.3%, closing at $88.39. The skid has continued into December.
CoreWeave’s gyrations highlight a broader issue that affects the entire AI industry: The rapid pace of growth has raised questions about how and when major capital investments are going to produce healthy profits. They also underscore the intensity of the frantic rush to build enough computing infrastructure to satisfy demand.
Billionaire Elon Musk has been roaring ahead with construction of a data center and energy project known as Colossus, built on 114 acres of swampland in western Tennessee, that’s set to house 200,000 Nvidia chips meant to power models made by his xAI startup, including Grok. Rivals across the AI chips industry are taking on Nvidia for a chance at a piece of the AI computing pie, citing “insatiable demand” for chips, data centers and power.
At the same time, concerns are mounting around the promised timelines for building out computing capacity across the industry, with construction delays and supply-chain bottlenecks threatening to push back hundreds of billions of dollars’ worth of spending already priced into valuations. Shares in cloud provider Oracle and custom-chip designer Broadcom are both down by double-digit percentages after they said some spending would come later than investors had hoped in recent quarterly earnings reports.
In late October, Core Scientific shareholders voted overwhelmingly to reject a $9 billion takeover attempt by CoreWeave after hedge fund Two Seas Capital, the landlord’s biggest active shareholder, publicly opposed the merger, writing in a letter that if the deal went through, Core Scientific’s shareholders would be “exposed to the high volatility of CoreWeave’s share price” and vulnerable to “substantial economic risk.” CoreWeave’s shares fell more than 6% after the deal fell apart.
Core Scientific has been flagging delays to its collaborations with CoreWeave since at least February, when it reported that it had pushed back certain construction timelines in order to make “design enhancements to further optimize GPU performance.” The company flagged weather-related delays in August.
Another major concern is the financing market for AI infrastructure players such as CoreWeave. Last week, after Oracle reported unexpectedly high capital expenditures on chips, networking equipment and other infrastructure, the bond market convulsed, raising the cost of capital for many large tech companies. CoreWeave’s cost of insuring against default on its debts soared to 7.9 percentage points.
Last week, CoreWeave priced a $2.25 billion convertible bond offering—a type of financing that has a lower interest rate than the asset-backed financing the company typically uses to pay for new data centers, but also comes with a risk of lowering the company’s stock price by diluting shareholders.
Sina Toussi, founder and chief investment officer of Two Seas, the hedge fund that wrote the August letter opposing CoreWeave’s acquisition of Core Scientific, said the deal made sense for CoreWeave, because it would have reduced its borrowing costs for the construction of new data centers, but the structure Intrator offered was deficient and the price undervalued Core Scientific.
“They’re exceptional at getting large workloads up at maximum utilization and replacing underperforming nodes rapidly without interrupting workflow,” said Toussi, whose fund also owns shares in CoreWeave. “But right now the market is concerned about the long-term value of AI.”
Gil Luria, an analyst with D.A. Davidson, said that CoreWeave has the “ugliest balance sheet in technology, by far.” Luria argues that CoreWeave’s operating margins of about 4% are less than half of what the company pays in interest on most of the debt it uses to deploy computing power for customers, making it hard to see how the company will generate profits going forward.
“The bull case is that they’ll scale into it, and that a lot of companies have low margins to start, but this is a company at scale. There is no scaling going on here,” Luria said.
Comments