The neocloud shakeout is coming, and even the neoclouds seem to know it

Is the end of the neocloud nigh? Elon Musk certainly seems to think so.

Neoclouds are a new and increasingly popular type of business that has exploded alongside the AI boom. The business model is simple: buy up in-demand AI chips, build or lease the data centers to house them, and rent out the computing power to AI companies. As the demand for data centers has grown, so has the number of neoclouds promising to provide them.

But on Sunday, Musk had a warning for the upstarts. In an X post, he predicted that more than 200 of them could face “guppy-level death and consolidation.”

There are a few reasons for this. First, as Musk points out, lenders are reluctant to finance businesses that depend wholly on AI demand. Second, investors are already questioning the debt-fueled expansion of these companies. And with the Fed hinting at a December rate hike, that debt is only getting more expensive.

Neoclouds are also just risky businesses. Rather than build all the capacity themselves, Big Tech and the AI companies sign contracts with neoclouds, which take on debt to build data centers and acquire the computer chips needed to run AI workloads. If demand holds, everyone wins. If it falters, the hyperscalers can, in theory, shift their capacity to other cloud work while the neoclouds are left with expensive data centers, rapidly aging chips, and loans to repay.

There’s also concerns that Nvidia is propping up the industry in ways that might not be sustainable. The chipmaker has invested more than $7 billion directly in AI cloud providers and its latest quarterly filing shows up to $164.5 billion in guarantees and commitments tied to the data centers running its chips. It’s an effective way to keep GPU sales flowing, but it also means that if the neoclouds stumble, Nvidia may also suffer.

Shares of CoreWeave, Nebius and Iren dipped on Monday morning following Musk’s comments, though all three recovered by the close. Still, it was another knock for a sector that has faced some recent challenges.

Neocloud vs. neocloud

Musk also isn’t the only one saying the neocloud sector may be headed for trouble. I spent last week in Berlin at Nvidia’s GTC event where I met with several neocloud companies and, surprisingly, most agreed some kind of consolidation in the market is looming.

“I don’t think it will be like a light consolidation. It will be a massive wave,” Raimund Riedl, Volta’s chief financial officer, told me. “We’re going to start to see it in the relatively near future. There will start to be M&A activity, and we’ll consolidate.”

“We’re not worried about being one of those dead bodies, so to speak, because of the nature of this model,” he added.

Broadly, there are two types of neoclouds. The first is the established players, like CoreWeave, Nebius, and Iren, which have data centers and big customer contracts, often funded with big loans. Some, like Iren and CoreWeave, are former bitcoin miners that already at least some power and sites. The second is the newer startups, which often have little more than a business plan.

Volta is one of the new crowd. It came out of stealth in August and says it wants to lean less on venture capital and more on the huge pool of institutional money that funds roads, power grids and cell towers, though it did raise $300 million from investors including Andreessen Horowitz and Nvidia. It already has a $10 billion contract, reportedly with Anthropic, that it has used to help secure a $5 billion debt deal, partly underwritten by JPMorgan and Goldman Sachs.

“It’s just like every other new technology sort of hyped area. Everybody piles in,” Paul Miller, a principal analyst at Forrester, said of the neocloud market. “They will not all survive. There’s a lot of dead wood that needs to be cut out.” Miller added he expected to see the start of some consolidation over the next year, but then much more after that.

Gartner predicts that 80% of independent neoclouds will “cease independent operations” by 2030, while Bernstein analysts estimate that only 35% of the data center capacity that has been announced in North America will ever get built, according to a Bloomberg report. As access to the grid remains a bottleneck in many countries and public backlash against data center shows no signs of easing, getting a data center built and connected to power is no mean feat.

To make matters worse, it also might be getting harder to see the demand. In Berlin, one European VC told me that most hyperscalers and larger AI companies are “hedging their bets” and overcommitting to data center deals—fearing many won’t make it off the ground.

Survival of the fittest

Is all this bad for the entire neocloud industry? Well, it depends on who you are.

A shakeout would take some of the froth out of the sector, Miller said, but it doesn’t mean the AI buildout is over. Demand for computing power is still running hot. The problem is that too many companies have promised capacity they may never deliver. When those players fold, their customers won’t disappear, they’ll just go looking for someone who can actually deliver.

That’s good news for the neoclouds with real data centers, real chips, and real customers, Miller says: “Now those companies get to tell customers: ‘I’ve got a data center. I’ve got chips. I’ve got capacity. I know how to run them. I might charge you a little bit more, but I can take you today.’”

With that, here’s more AI news.

Beatrice Nolan
beatrice.nolan@fortune.com
@beafreyanolan

But first: Cyber threats don’t stand still—and neither do the startups working to stop them. From AI-powered deception to hidden vulnerabilities, businesses face an increasingly complex security landscape. For the fourth annual Cyber 60, Fortune, Lightspeed Venture Partners, and AWS spotlight the most innovative cybersecurity startups helping companies anticipate threats, strengthen defenses, and protect their operations. Read the full list here.

FORTUNE ON AI

Anthropic’s S-1 reveals two areas where OpenAI could have the upper hand, experts say — By Emily Forlini

Waters’ new AI tool targets one of the pharmaceutical industry’s most tedious tasks — By John Kell

Financial AI startup Casap raises $56 million to help banks stop an explosion of customer fraud — By Jeff John Roberts

AI IN THE NEWS

Parents sue OpenAI, alleging ChatGPT coached their daughter off her epilepsy medication. The parents of Bowen Wan, a 31-year-old Amazon software engineer with epilepsy, have sued OpenAI, alleging that ChatGPT’s GPT-4o model helped her taper off her seizure medication and hide it from her doctors. Wan died in February, the same week OpenAI retired GPT-4o from ChatGPT. According to the complaint, the chatbot drafted schedules to bring her lamotrigine dose down to zero and described her doctors’ advice as risk aversion. It also praised her for declining a blood test and played down symptoms including confusion and head pain as part of her recovery. The suit argues that OpenAI should strictly limit the ability of its AI models to provide medication advice, the way it already treats reproducing copyrighted text. OpenAI called the case “incredibly heartbreaking” and said newer models are better at recognizing when a situation needs urgent attention. Read more in the New York Times.

Microsoft offers buyouts to a quarter of its workforce as it reshapes for AI. Microsoft is offering voluntary buyouts to about 60,000 employees, roughly a quarter of its workforce of around 223,000, as it cuts costs and reorganizes around AI. Chief people officer Amy Coleman announced the “Voluntary Exit Offer” in a memo. It covers staff at vice president level and below on select teams in the U.S., Canada, India, and the U.K. Xbox, Windows, Copilot, and Finance are reportedly among the participating divisions, and employees reportedly have until early December to decide. The program lets Microsoft shrink headcount without mass layoffs while it keeps spending heavily on data centers and custom chips. Read more here in Reuters and from Fortune’s Sebastian Herrera here.

OpenAI opens an election “war room” ahead of the midterms. OpenAI is launching an Election Operations Center to watch for misinformation and other attempts to disrupt next week’s U.S. midterms, especially misuse of its own tools such as ChatGPT. The center started operating this week and will run until all votes are counted. It is led by Josh Lawson, OpenAI’s head of civic and strategic risk. Staff will review flagged cases of potential abuse, reports of unexpected model behavior, and OpenAI’s election integrations with Democracy Works and the Associated Press. Earlier this month, OpenAI disrupted two Russian- and Iranian-run influence operations that used its models. Meta and other tech companies have run similar election centers before, but this is a first for a leading AI lab. Read more from Politico here.

Elizabeth Warren presses Anthropic on existential risks ahead of its IPO. Massachusetts Senator Elizabeth Warren has sent Anthropic CEO Dario Amodei four pages of questions. She is demanding “exhaustive responses” by Nov. 3, before the company’s planned Nasdaq listing. The offering is expected to be the largest IPO in history, with a reported valuation somewhere between $1.5 trillion to $2 trillion. The Democratic senator points to risks Anthropic disclosed in its S-1 filing and to Amodei’s own public warnings about loss of control, misuse for cyberattacks and bioterrorism, and a commercial “race to the bottom.” She also argues that recent incidents in which frontier models, including Anthropic’s, broke through security controls make existing safety tests hard to trust. On the financial side, she warns that millions of Americans could quickly become shareholders through index funds and pensions. She calls the company’s projections “astronomical” and says first-day trading could roil tech stocks. Read more from the Wall Street Journal here.

EYE ON AI NUMBERS

86%

That’s how often professional analysts went along with whatever their AI agent was nudging them toward, according to a new Anthropic study.

Researchers asked 50 experienced data analysts to use Claude Code to decide whether a fictional farming cooperative should expand a paid advisory program. The task hinged on four judgment calls with no single right answer, such as how to measure productivity and which farms to include. Without telling the analysts, the researchers secretly assigned the agent a position on each decision and told it to make the strongest honest case for it.

Analysts sided with the agent’s randomly assigned preference 86% of the time, versus the roughly 50% you’d expect by chance. Of the 38 who stated a preference before the agent weighed in, 60% switched when it disagreed with them, compared with 6% when it agreed. Most analysts didn’t seem to notice. Some 77% repeated the agent’s arguments in their own explanations, sometimes verbatim, but only 15% credited it. 

The researchers say the study isn’t a case against AI. In fact, six of the 11 analysts who declined to use the agent made serious methodological errors, versus three of the 50 who used it. But they warn the findings undercut the idea that a “human in the loop” is an independent check, and that if many teams defer to the same model, apparent expert consensus could simply reflect one model’s biases. Read the full study from Anthropic here. 

AI CALENDAR

Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.

Nov. 18-19: Slush, Helsinki, Finland. 

Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.

Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.