I looked into CoreWeave and the abyss gazed back

Hello, my friends. Have you been feeling too sane lately? Have I got something for you! It is a company called CoreWeave.

You may not have heard of it because it’s not doing the consumer-facing part of AI. It’s a data center company, the kind people talk about when they say they want to invest in the “picks and shovels” of the AI gold rush. At first glance, it looks impressive: it’s selling compute, the hottest resource in the industry; it’s landed a bunch of big-name customers such as Microsoft, OpenAI and Meta; and its revenue is huge — $1.4 billion in the third quarter this year, double what it was in the third quarter of 2024. The company has almost doubled in share price since its IPO earlier this year, which was the biggest in tech since 2021. So much money!

But as I began to look more closely at the company, I began feeling like I’d accidentally stumbled on an eldritch horror. CoreWeave is saddled with massive debt and, except in the absolute best-case scenario of fast AI adoption, has no obvious path toward profitability. There are some eyebrow-raising accounting choices. And then, naturally, there are the huge insider sales of CoreWeave stock.

After I unfocused my eyes a little, I realized CoreWeave did make a horrible kind of sense: It’s a tool to hedge other companies’ risks and juice their profits. It’s taking on the risk and the costs of building data centers that bigger tech companies can then rent while they build their own data centers which may very well wind up competing with CoreWeave. What’s more, it’s part of a whole stable of companies that are propping up demand for the behemoth of the AI boom: Nvidia.

I don’t think CoreWeave’s weaknesses are a secret. It just seems like a lot of investors are ignoring them. Whether that’s because of AI fomo or a sophisticated game of chicken — getting as much money from CoreWeave shares before the inevitable collapse as possible — I can’t really say. I just know that after spending several days parsing its financials, talking to analysts and other experts, and trying to understand AI infrastructure, here’s my life now:

CoreWeave went public in March, at a share price of $40, and at its peak this year was worth $187 a share. That was back in June. Today, the shares opened at $75.51. Some of the price decline is due to CoreWeave announcing, along with its third quarter earnings, that delays on a data center mean it’ll make less money this year. The delay underscored some of the difficulties CoreWeave faces in becoming a profitable business.

CoreWeave is “the poster child of the AI infrastructure bubble,” wrote Kerrisdale Capital, an investment manager, in a September note announcing it was shorting the stock. “Strip away the noise, and CoreWeave remains an undifferentiated, heavily levered GPU rental scheme stitched together by timing and financial engineering, not lasting innovation.” Kerrisdale thinks the fair value stock price of CoreWeave is $10.

CoreWeave, of course, does not agree with this assessment. “CoreWeave is the essential cloud for AI, a software solution, built on cutting-edge physical infrastructure and designed from the ground up to provide the most effective and efficient super computers to those whose workloads demand the most computing power,” said Lia Davis, CoreWeave’s head of global communications, in a written statement. “We are seeing strong support from existing customers, who continue to extend and expand their agreements with us.”

All right. Time for the rabbit hole.

I originally thought it might be fun to use CoreWeave as a glimpse at the economics of AI infrastructure. It’s recently public, it doesn’t have a second business line, and so whatever’s going on there might give me a direct sense of what to expect out of the data center boom. I was not expecting what I found.

Let’s start with some very recent history. CoreWeave is a data center company that pivoted in 2022 from crypto. (In 2021, CoreWeave made its money by… mining Ethereum.) Essentially, CoreWeave is a landlord for compute: companies pay for the use of its server racks for AI projects.

CoreWeave first came to my attention because it innovated in something that surprised me: using GPU as collateral for $2.3 billion in loans at an effective interest rate of 15 percent in the last quarter, according to the company’s most recent quarterly filing. It turns out this was a pioneering innovation — companies such as Crusoe, Fluidstack, and Lambda have also taken out similar loans. Even CoreWeave itself took out a second loan, for $7.5 billion, using more GPU for collateral. This time, the company got better terms — only 10 percent interest, as of the third quarter. It also amended the loan to draw another $400 million — getting an effective 9 percent rate in the process. The third loan was even better: 9 percent interest in the third quarter. (All three loans have floating interest rates.)

The chips bring me to the main thing: CoreWeave simply isn’t possible without Nvidia. The company said it owned more than 250,000 Nvidia chips, the infrastructure necessary to run AI models, in documents CoreWeave filed for its initial public offering. It also said it only had Nvidia chips. On top of that, Nvidia is a major investor in CoreWeave, and owned about $4 billion worth of shares as of August. Nvidia made the March IPO possible, according to CNBC: when there was lackluster demand for CoreWeave’s shares, Nvidia swooped in and bought shares. Also, Nvidia has promised to buy any excess capacity that CoreWeave customers don’t use.

“These investments are not circular; they are complementary,” CoreWeave’s Davis wrote. “This is about an entire ecosystem all rowing in the same direction to accelerate the AI economy. There’s nothing circular about it. Rather, these partnerships are about accelerating innovation and adoption. We are, collectively, defining the next-generation operating system for civilization.”

Okay, well, first of all: I notice that AI so far can’t so much as turn on the lights — and remember, when it comes to future claims, everything is vaporware until it ships. But CoreWeave isn’t alone in its big bets on AI. So let’s talk about the customers who are defining the “next-generation operating system for civilization.” They’re big names, sure, and they’re paying a lot of money. But I’m not sure they’re reliable long-term clients — in fact, they’re more likely potential rivals.

CoreWeave’s big buyer is Microsoft, which accounted for 71 percent of CoreWeave’s income in the second quarter and 67 percent in the third. That makes CoreWeave more dependent on Microsoft than it was in 2024, when Microsoft accounted for 62 percent of revenue that year. But it kinda seems like Microsoft may have soured on CoreWeave. In March, the Financial Times reported that Microsoft “has walked away from some of its commitments” with CoreWeave “over delivery issues and missed deadlines.” (CoreWeave disputed the reporting.) Around the same time, Microsoft chose not to buy $12 billion of CoreWeave capacity, even though it had the option, according to Semafor. Now, this may have more to do with Microsoft and OpenAI rejiggering their partnership rather than CoreWeave’s service, but it nearly derailed the CoreWeave IPO.

Beyond any immediate conflicts, Microsoft is building its own AI chips and says it wants to rely on them in the future. It has its own data centers, and is opening more. It is not difficult to imagine a future where CoreWeave’s biggest customer suddenly becomes its biggest competitor; that possibility is listed in CoreWeave’s risk factor section of its own filings.

OpenAI entered its own five-year, $11.9 billion agreement with CoreWeave in March, and invested $350 million in CoreWeave to boot. That deal was expanded in May, and again in September, for what CoreWeave says is a total value of $22.4 billion. This isn’t an airtight deal — OpenAI can terminate part or all of it if CoreWeave repeatedly doesn’t deliver on schedule. There are two other things that make this deal risky. First, OpenAI doesn’t make money. Second, OpenAI is investing in its own data center project, Stargate, and says it will supply 75 percent of the compute OpenAI needs by 2030. So that’s a five-year contract that might not renew.

Naturally, I asked CoreWeave how they planned to compete, if necessary, with these customers who are also building their own data centers. I am going to give you the written response verbatim:

Astute readers will notice that this does not answer my question. The dodge suggests to me that CoreWeave doesn’t actually have a good answer.

From the IPO documents, there is an unnamed fourth customer that, after Microsoft, is its second biggest. (None of CoreWeave’s other customers represented 10 percent or more of CoreWeave’s revenue in 2024, according to the S-1.) This is Nvidia, which “agreed to spend $1.3 billion over four years to rent its own chips from CoreWeave,” according to The Information.

In September, Nvidia signed a $6.3 billion contract that lets CoreWeave interrupt delivery, “giving us the ability to redirect capacity toward other opportunities without compromising overall commitments,” Davis said in her written statement. “As a result, this contract allows CoreWeave to sell shorter-duration capacity to smaller companies that today are not positioned to sign the long-dated contracts we typically sign with our customers without sacrificing our disciplined approach to success-based capex.”

There are some upsides to CoreWeave that aren’t yet on its balance sheet: remaining performance obligations (RPO), or revenue CoreWeave expects that hasn’t yet been paid. CoreWeave projects that almost half of its current $50 billion in RPOs will arrive by the end of September 2027, according to its most recent quarterly filing. There’s also some revenue backlog.

But that’s the future, and CoreWeave’s current debt is substantial. Unless there’s a sudden huge uptick in CoreWeave’s revenue over the next year, CoreWeave will likely need to borrow more money, Forbes reports. Which brings me to the thing that hangs over the company…

What worries me most about CoreWeave is that I can’t figure out how on Earth it outruns its debt.

According to the three major ratings agencies, CoreWeave’s credit rating isn’t investment grade, which means they think there’s a likelihood CoreWeave won’t repay its debts. Junk bond status makes it more expensive for CoreWeave to borrow; junk bonds are more politely known as high-yield bonds, because lenders ask for extra money, ostensibly to compensate them for extra risk.

Don’t worry, the gap between interest and income is getting bigger. In its third quarter earnings, reported earlier this week, we’ve got $51.9 million in operating income — and $310 million in interest expense. As of the end of September, CoreWeave was $14 billion in the hole with current and long-term debt, about $3 billion more than the quarter before.

CoreWeave has several kinds of debt, but the main vehicles are the delayed draw term loans (DDTL), a type of loan that can be structured to have particularly complex requirements for the borrower to receive the loan payout. At the end of July, CoreWeave announced its third DDTL, to help finance services for OpenAI. All three loans are with CoreWeave subsidiaries, called special purpose vehicles, which make it easier for the loans to be chopped up and securitized. The loans are secured by CoreWeave’s Nvidia chips and data centers.

I’d like to linger on the special purpose vehicles for a minute. These SPVs aren’t subject to the same kinds of regulations as the parent company. They allow funding for projects before a contract is made. They may have tax benefits. They can allow third-party partners to participate in financing. And, depending on how they are structured, they can protect assets in the event of a bankruptcy.

In CoreWeave’s case, they also allow for a lower cost of capital, Davis wrote. In the OpenAI loan, an SPV exists for GPU, high-speed fiber optics, and other assets that support the deal. The debt from the DDTL is secured by the assets and the OpenAI contract. OpenAI doesn’t have a debt rating; a deal done this way for an investment-grade customer would likely be even lower.

CoreWeave’s SPVs “look pretty standard for de-risking incoming capital to an existing company,” says May Hen, a fiscal sociologist at the University of Cambridge’s Faculty of Law, in an email. CoreWeave chief executive officer Michael Intrator, a former hedge fund manager, is “probably ensuring each part of the influx of capital is protected separately as an investment vehicle.”

CoreWeave’s loans also have some notable conditions. The first two DDTLs, as noted in the S-1, have variable interest rates, based partly on its customers. Investment-grade customers, for instance Microsoft, mean CoreWeave can borrow at lower rates. The third and most recent loan requires CoreWeave to bill or receive 85 percent of the revenue it projected in that month and the two before it. So if a big customer, for instance OpenAI, doesn’t make its payment, that’s a potential default.

This isn’t even all the debt. There’s a $650 million revolving credit agreement with JPMorgan Chase, which expanded to $1.5 billion in May, and expanded again to $2.5 billion on Nov. 12th. (As of September, CoreWeave had drawn down $700 million of that money.) There are senior notes, a loan from Magnetar, and vendor financing. Each of these types of lending have different structures, says Gil Luria, an analyst with DA Davidson.

“They have to keep borrowing more and more because they spend more money than they can get, structurally,” Luria says. “They have to continue to borrow to pay interest on the last loan.”

Surely CoreWeave is running a tight ship on its complex financing operations, right? Actually, CoreWeave has already violated some of the terms of its $7.6 billion loan, triggering a technical default, which it had to ask Blackstone to waive. (Even though CoreWeave was paying its loans on time, violating some of the terms means lenders can demand repayment in full.) Fortunately, Blackstone didn’t even ask for a penalty payment, according to the Financial Times. Still, this — in combination with the disclosure of “material weaknesses in our internal control” in the S-1 — seems significant.

Those weaknesses, according to the third quarter report, have not yet been fully remediated. That makes it possible that CoreWeave could screw up its financial statements, though there is no evidence that has happened. “We believe we are making progress,” the quarterly statement notes. But the company can’t conclude it’s fully fixed the problems until they’ve finished all their planned changes and evaluated them.

There’s a fun phrase that occurs twice in the CoreWeave IPO filing: “We have a proven track record of securing power.” This track record means a deal with a company called Core Scientific (no relation) “for more than 500 MW of capacity” as of the end of 2024.

Core Scientific is a formerly bankrupt Bitcoin miner. As part of its Bitcoin mining operations, it made long-term deals with power companies. Those long-term contracts are worth “many orders of magnitude more” than actually mining Bitcoin, because utilities have finite capacity, says Luria. So Core Scientific is now selling power to its only customer, CoreWeave. “These Bitcoin miners stumbled into a goldmine, completely accidentally,” Luria says.

CoreWeave has tried and failed to buy Core Scientific twice now. Sure, buying it would have given CoreWeave a gigawatt of data center capacity, but most of that capacity is already contracted out. And while it also would have meant that CoreWeave could stop paying Core Scientific for its services, I’m not really sure that matters. Core Scientific has its own problems, primarily that it doesn’t make any money, either.

“Core Scientific is spending a dollar and five cents to get a dollar from CoreWeave,” says Luria. “And CoreWeave is spending a dollar to get 95 cents of revenue.” Also, it’s not like CoreWeave would be getting revenue from other Core Scientific customers, because Core Scientific doesn’t have any other customers. So by cutting out the middleman, CoreWeave would be spending $1.05 to get 95 cents, Luria says.

That deal fell through because Core Scientific shareholders thought their company was worth more than what was offered. “The outcome of the acquisition had no impact on CoreWeave’s ability to execute against its roadmap,” wrote CoreWeave’s Davis. In the third quarter, CoreWeave’s capacity increased more than 600 MW. “This increase in contracted capacity in the third quarter exceeds the entirety of our relationship with Core Scientific.” She characterized Core Scientific as “nice-to-have, not need-to-have.”

But CoreWeave has nonetheless been on an acquisition spree. This year alone it’s acquired Marimo, Weights & Biases, Monolith, and OpenPipe, all of which make assorted computing tools. “We’ve really been buying and building the AI cloud,” Intrator said on Bloomberg TV.

As near as I can tell, this is meant to differentiate CoreWeave from just being a data center company. “CoreWeave has built an AI cloud platform that includes both infrastructure and software,” Davis wrote. Be that as it may, I don’t think Microsoft, OpenAI or Meta are there for software tools — they make their own. And I’m not convinced that the small companies that might need these tools are enough to keep CoreWeave’s lights on.

So how is CoreWeave going to get all the compute and power it needs? Part of the answer is that it’s building its own data centers. But building is expensive. “CoreWeave has been at the tip of the spear of raising capital for the AI buildout,” Intrator noted. “Our growth continues to rage along. It’s amazing just how fast we’re growing.” He went on to say that CoreWeave would continue to raise capital to support its buildout as was necessary.

Building is slow. As Intrator told Bloomberg, CoreWeave can’t control every aspect of the build. And construction delays may seriously threaten CoreWeave, because there’s one more expense lurking that isn’t yet on the company’s balance sheet: $34 billion in scheduled lease payments. Some of those payments are for “data centers and office buildings that have not yet begun to operate or bring in revenue,” writes Forbes. So if CoreWeave can’t build out fast enough, and customers cancel their contracts, it’s particularly vulnerable.

When I asked about CoreWeave’s plans for meeting its contractual obligations, Davis said that all the contracts in CoreWeave’s revenue backlog are matched to specific, existing data centers.

But building fast seems particularly crucial to the company. Remember that if CoreWeave repeatedly fails to meet its obligations to OpenAI, OpenAI can yank some or all of its money. If that happens, CoreWeave goes bankrupt, Luria says. And CoreWeave has already made investors nervous. In its third quarter earnings, the company disclosed that a data center partner hit a delay. According to Intrator, that means a delay of revenue as well; he declined to tell Bloomberg which customer had been affected. CoreWeave’s share price plunged. Even CNBC’s Jim Cramer, usually a CoreWeave bull, seemed upset.

“Some people might think it’s one complex, but when I go over the numbers, we’re talking about multiple places,” CNBC’s Jim Cramer said, noting that delays were in Texas, Oklahoma, and North Carolina. “And it just so happens that the places are all connected to an outfit called Core Scientific that you tried to buy.” An uncomfortable-looking Intrator didn’t confirm whether the partner in question was Core Scientific.

CoreWeave isn’t alone in its complex finances. Meta took on debt, using a SPV, for its own data centers. Unlike CoreWeave’s SPVs, the Meta SPV stays off its balance sheet. Elon Musk’s xAI is reportedly pursuing its own SPV deal. “Financial engineering is back in style, making some analysts wonder if all the commitments will be easy to spot,” wrote Bloomberg.

There are other massive debt commitments too — Oracle sold bonds, Google parent company Alphabet has dipped into debt markets twice this year, and something called TeraWulf sold junk bonds for a New York data center. “Morgan Stanley estimates private credit markets could supply over half the $1.5 trillion needed for the data centre buildout until 2028,” Reuters wrote.

CoreWeave isn’t going to be the only one of these companies that hits delays in their buildouts, either. It’s just more vulnerable than they are. “We and many others have been quite forceful in making the case that there is a role in government in helping us,” Intrator told Bloomberg, suggesting permitting processes could be made easier. This echoes remarks made by OpenAI’s Sarah Frier, suggesting the government should “backstop” the debt for the AI buildout.

Asking for government help does not strike me as a bullish sign. And as a taxpayer, I have no interest in helping private companies build data centers so they can keep the profits — if, indeed, there are any.

Perhaps you looked at the second quarter statement and you did some very basic math. CoreWeave had $1.2 billion in revenue, with only $313 million as the cost of revenue. That suggests a margin of almost 74 percent. The third quarter looks similar: $1.4 billion in revenue, with just $369 million as the cost of revenue — also a margin of about 74 percent.

Compare that to Oracle, another major AI data center provider. In October, co-CEO Clay Magouyrk said he expected its AI data center business to “eventually generate a gross profit margin of 30 percent to 40 percent,” according to The Information. “Oracle has found it challenging to generate a gross margin of more than 25 percent from renting out Nvidia chips that came out one or two years ago,” The Information added, citing an internal document.

So let’s consider CoreWeave’s 74 percent margin. Is the company really executing that much better than Oracle? I don’t think so, and neither does Luria. This is an accounting choice, and one I’m not sure I like.

The biggest cost for a data center is the depreciation of assets, which happens fast. Chips get beat up, new chips get released, etc. CoreWeave is reporting this as “technology and infrastructure,” which apparently also captures some other expenses too, and is considered an operating expense. But usually, this goes into the “cost of revenue” figure, Luria says. I checked 10-Q filings for Microsoft, Alphabet, and Oracle for a “technology and infrastructure”category, and didn’t find it. Meta writes directly about putting depreciation of its data centers in its cost of revenue.

So let’s look at the second quarter again: Add the $559 million in depreciation to the cost of revenue, and suddenly, that’s an estimated 28 percent margin, which isn’t that far off from Oracle. The third quarter is worse. In it, we see $747 million in depreciation, which, added to the cost of revenue, gets us an estimated margin of 20 percent.

I am calling this an estimate because when I asked CoreWeave about the decision, Davis told me “the substantial majority of our depreciation and amortization is within technology and infrastructure, with a small fraction in our cost of revenue.” So the figures I’m giving here might not be quite right. She declined to comment on why CoreWeave was using the “technology and infrastructure” line, or why depreciation and amortization had been categorized this way.

Most of CoreWeave’s contract prices are fixed, though the company can get bonuses if it delivers compute early, Davis wrote. Operating costs aren’t fixed in the same way. If, say, chips get more expensive, that knocks the margin even lower. It can’t keep costs down by using its own chips, the way some companies, like Amazon, can. If the chips wear out faster, that’s a problem, too. Companies like Oracle or Meta have other, higher-margin revenue streams. While CoreWeave sells some software, such as that it acquired from Weights & Biases, that’s not its main business.

Even if we ignore CoreWeave’s debt load, its margins are fragile, and dependent on customers renewing or expanding contracts.

After I noticed this, I got a little more interested in the 10-Qs. In the third quarter, there was a drop of almost $1 billion in net cash, over this period last year. This was partly attributed to an increase in “accounts receivable,” which are basically customers’ as-yet-unpaid bills. An increase in receivables means there’s less actual U.S. American dollars on hand. Sometimes it means a business is stressed — because customers that delay payments may be under financial strain themselves.

Elsewhere in the 10-Q, CoreWeave notes that its customers usually have to pay within 60 days, but some can wait as long as 360 days to pay. I asked if the increase in receivables meant some customers were having difficulty paying. “CoreWeave does not have any issues with late payments from customers,” Davis said.

This brings us to the final problem: we don’t know if companies will keep buying the main thing CoreWeave has to sell.

It’s possible they will. “What would allow them to dig out of the debt hole is that they have developed superior software to the rest of the market that makes them a magnet for leading AI providers, and the fact that they’re able to charge hyperscaler-like prices while remaining flexible and able to offer early access to the latest GPUs,” says Brendan Burke, an AI industry analyst. Davis, the CoreWeave spokesperson, noted that CoreWeave “recently earned SemiAnalysis’ highest honor (again).”

That seems to confer, at least, a pricing advantage. CoreWeave charges $68.80 per hour for a cluster of Nvidia B200s, according to Cloud GPUs, a compiler for pricing. The next-most expensive company, Cirrascale, charges from $38.32 to $47.92. The CoreWeave premium reflects that “not every data center is created equal,” Burke says. “There’s a wide spread in reliability between different data centers on the market.” If CoreWeave maintains its engineering advantage, Burke thinks CoreWeave will continue to attract large customers. That’s not a given in the long term. And if CoreWeave has to compete on price, that’s a problem.

There’s also the question of demand. Right now, compute is one of the hottest commodities in AI. But the demand is mostly for training models, says Alex Hanna, the director of research at the Distributed AI Research Institute. Inference — basically, customer use, a more predictable ongoing expense — doesn’t come close. Chirag Dekate, an analyst at Gartner, concurs.

This might work out for CoreWeave in its best-case scenario. Let’s say companies require more and more compute to train ever more powerful models, and customers basically immediately adopt AI and start using it heavily. In this scenario, no one ever has to stop building data centers, and there’s always going to be more demand for compute than Google, Amazon et al. can provide. If CoreWeave can somehow stabilize its debt or charge more for compute, maybe things turn out fine. Or maybe CoreWeave can just keep borrowing and building forever.

Obviously if AI flops, CoreWeave is fucked along with everyone else. What happens if it’s important but basically normal technology? Let’s just assume AI isn’t going to surpass human intelligence in 2030 and upend everyone’s job like Sam Altman’s saying. Let’s say it’s more like online shopping. Online shopping, of course, is now ubiquitous. But its adoption curve left behind a graveyard of companies from the first dot-com boom. Take Pets.com, which shut down in 2000, and was, at the time, the mascot for dot-com failure. In 2017, Chewy, which was basically the same idea, was bought by PetSmart for $3.35 billion, which at the time was the largest-ever ecommerce acquisition. The main difference between Chewy and Pets.com is that in the 17 years between Pets.com’s failure and Chewy’s massive sale, people got comfortable with buying online.

Or consider streaming video. Pseudo.com, an early audio and video streaming service founded by Josh Harris, filed for bankruptcy in 2000. Now, 25 years later, video streaming is eating both television and movie theaters. The difference is that… among other factors, people aren’t on dial-up anymore. Harris’ idea was right, but the tech wasn’t ready yet.

So if AI is a normal technology, it’s going to take time for people to adopt it and companies to work out the kinks. A lot of demand assumptions have presumed that AI will flow out to low-tech industries, but so far its use is concentrated in fairly sophisticated ones, Burke says. The physical world is still relying on old computing techniques, and not every business can afford the investment it would take to shift to AI, even if it were a less risky technology than it is now.

“The idea that this is a complement to our existing systems rather than a fundamental thing suggests we don’t need to exponentially increase the volume of computing,” Burke says. Big companies like Microsoft don’t have to bet their entire company; they can just bet a portion, leaving them flexible to adapt to whatever the future holds. But companies like CoreWeave that are all in on AI have taken on a huge debt load that ties them to one specific future, Burke says.

Meanwhile, whether training models keep sucking up more and more power is anyone’s guess. Remember several months ago, when everyone panicked over DeepSeek because it was so effective and so cheap to train? People have calmed down since then, but 90 days ago, Luria says that major companies told him that compute supply was catching up with demand. A few weeks ago, these companies reversed themselves: demand surged again. What will happen next week?

There’s been lots of talk about how many Americans use AI daily — a fifth of Americans, if Menlo Ventures is to be believed — but many of these products are available for free, which may not be sustainable forever. The highest rate of use is among students, Menlo found. Whether people will pay for ChatGPT seems like it’s still an open question. I am skeptical that AI is really a consumer technology — it’s not clear the new web browsers and “AI agents” are actually all that useful. The fact that OpenAI has released a slop social network doesn’t seem like a great sign. What’s more, most surveys of consumer behavior suggest that people don’t really trust AI. There’s good reason for that!

There are probably real enterprise uses for AI, since that category is significantly broader than fact-challenged chatbots. The people I know who are most excited about AI say that it makes them significantly faster at coding. But 95 percent of businesses are getting zero return on their AI investments, according to the MIT Media Lab. What’s more, enterprise-grade systems “are being quietly rejected” by organizations because they tend to, well, fail.

That could change. But it might change slowly, and it might mean specialized small AI programs, rather than big general ones. If the demand for AI grows slowly, the breakneck pace of data center expansion will also slow. “There are a million ways CoreWeave can fail, but people believe in this very rosy scenario where we build data centers forever,” Luria says. “I talk to a lot of investors, and they tell me it doesn’t matter. As long as there’s insatiable demand for compute, CoreWeave will continue to grow.”

Luria told me that many of the big players in the space are willing to overbuild, because other parts of the business can use those centers. All they really have to do is pause spending for a couple years. CoreWeave doesn’t have that luxury. If the current data center frenzy results in supply meeting or outstripping demand, it’s fucked — because it’s the overflow facility for the big players.

“It’s noteworthy that people who have a good view on that business are cashing out,” says Leevi Saari, a fellow at the AI Now Institute. If you thought CoreWeave was going to keep growing in a once-in-a-lifetime fashion, wouldn’t you want to hold onto those shares and see if you could out-rich, like, Larry Ellison or Elon Musk? Of course, if you think CoreWeave’s stock price is a flash in the pan, you might want to get out while the getting is good.

“Stock sales of this kind are common and routine for newly public companies,” CoreWeave’s Davis wrote. “The recent selling has primarily come from early investors. Any founder sales were made under non-discretionary pre-scheduled 10b5-1 trading plans for standard tax and wealth planning purposes and represent a small portion of their overall holdings.”

Despite their sales, CoreWeave founders retain control of the company through Class B shares. Much like Meta, the dual-class structure means that founders effectively get extra votes through these shares, which then convert to regular Class A shares for sales. Davis declined to say how much voting power the founders had after the most recent sales.

So some people have gotten very rich from AI, largely from selling shares in speculative investments. “There’s a bubble, and CoreWeave is likely to fall in the very near future,” says Saari. “I think it is retail that will be left holding the bag.”

I started kicking the tires on CoreWeave because I thought it might be a good place to think about the economics of AI data centers. But the company seemed like such a weird outlier that I, uh, got kind of derailed.

It seems to me like CoreWeave is a great idea for everyone except CoreWeave. AI’s actual end users, the real clients, are coming to Microsoft, OpenAI, and the other big tech companies. CoreWeave lets these companies take on more demand without splashing out on depreciating assets by supplying extra data center capacity until their new centers get built. Once that happens, they may not need CoreWeave at all.

CoreWeave doesn’t have meaningful lines of business besides selling compute. It is dependent on a few big contracts. People who are bullish on CoreWeave are essentially saying that we are going to have to build data centers forever and ever, amen. That seems unlikely, even if there is massive demand for AI, which isn’t necessarily a given.

While all of CoreWeave’s clients benefit from CoreWeave eating their risk, there’s a company that benefits most: Nvidia. It is, after all, CoreWeave’s investor, customer, and vendor. Two emails requesting comment to Nvidia’s press contact — and two more to Ken Brown, its head of corporate communications — were not returned.

It makes a certain kind of cynical sense to view CoreWeave itself as, effectively, a special purpose vehicle for Nvidia. Here’s a stylized way of looking at it: Nvidia can spend one dollar — in investment or other payments — to prop up an entity that promptly buys five dollars of Nvidia chips, using money borrowed from other companies. Nvidia directly profits from the sale, and also creates competition for its chips, giving it leverage over the Microsofts and Amazons of the world. All without Nvidia taking on CoreWeave’s debt, because CoreWeave is a separate company.

“If Jensen can keep the competitive fire on, that might discourage hyperscalers from developing their own chips,” says Saari. CoreWeave and other so-called neocloud companies help Nvidia keep its market share and push itself as the de facto infrastructure in the market.

“CoreWeave and Nvidia have a long-standing partnership that’s consistent with Nvidia”s relationships across the industry and important for building out access to AI infrastructure to meet today’s relentless demand,” CoreWeave’s Davis said in a statement. She added that CoreWeave is an independent company that doesn’t receive preferential treatment from Nvidia.

“There is no neocloud that exists without Jensen,” says Saari. That makes neoclouds, in effect, extensions of Nvidia, he says. And none of them make money, so to expand, they must take on debt.

If we look at these as being, metaphorically, Nvidia’s special purpose vehicles, then it doesn’t really matter if the companies are any good or will survive in the long term. Their job is to boost Nvidia’s sales. Even OpenAI, also an Nvidia investment, kind of falls into this category — because the massive data center buildout that OpenAI wants the government to backstop sure involves an awful lot of Nvidia chips.

If you are old enough, or possessed of a certain kind of disposition, you may be thinking, Wait a minute, aren’t you describing Enron? And uh, in some sense, yes! Enron’s whole thing was special purpose vehicles with extremely speculative valuations that were used to take on debt, Luria notes. But Enron lied about what it was doing, and that’s fraud and illegal. (It also got up to other illegal stuff besides.) Nvidia’s relationship with CoreWeave is all happening in plain sight. So are all the relationships with the other neocloud companies. It kind of seems like the tech company version of the GameStop open pump-and-dump.

“It’s not good behavior, and it’s not healthy behavior,” Luria says. “But it’s legal. Any investor can see this. Many are just choosing not to.”

Nvidia’s investments have been accelerating — $3.8 billion in nonmarketable equity securities as of the end of July, a $2 billion increase from the year before. Nvidia made four investments in 2020; it has made 51 as of September 2025, according to CNBC. “The majority of Nvidia’s portfolio companies have some strategic connection to the company’s business,” CNBC noted.

Huang has been selling shares — more than $1 billion since June. To put that in perspective, Huang has sold almost $3 billion of Nvidia shares since 2001, according to Bloomberg. It’s not just Huang. Nvidia insider sellers have been unloading throughout the third quarter. With Huang’s sales, the insider figures stand at $1.5 billion in that quarter. “While many companies have benefited from the AI gold rush, Nvidia stands alone in a virtually unheard-of feat of wealth creation with seven billionaires, including Huang, among its ranks,” Bloomberg wrote of the recent sales.

Maybe I am being unfair. Nvidia has been making a lot of money over the last few years, and generally investors expect companies to do something with that money. One could argue Nvidia’s profits are so outsized it’s exhausted all the usual methods, like share buybacks and acquisitions. But by investing in AI companies with ties to Nvidia itself, Nvidia has added more risk — because if something goes wrong for Nvidia, it goes wrong for all these investments, too.

I’ll tell you something: diving into CoreWeave financials has made me feel fucking crazy. Like, here I am suggesting that a bunch of independent neocloud companies look like Nvidia SPVs. But the real thing I learned from CoreWeave has been that Nvidia is basically propping this company up, with some help from Microsoft, while Huang signs titties or whatever. If that’s indeed what’s happening, then — as the song goes — enjoy yourself, enjoy yourself, it’s later than you think.