If you’re not familiar with CoreWeave by now, you need to remedy that situation posthaste, because they’re the most important player in the business of renting the kind of compute you need to compete in the AI race.

The story’s pretty funny. The company’s founders were commodities traders as recently as eight years ago. They stumbled into the GPU business when they bought a couple of Jensen Huang’s chips and repurposed them to mine the crypto they were using to bet on fantasy football. During 2018’s crypto winter, they started buying up more chips from failed mining operations. One thing led to another and then to another after that and by the time ChatGPT was released to the world in 2022, CoreWeave was set up to sell at-scale access to Nvidia hardware. Huang took a stake in CoreWeave the following year.

They went public in March in a downsized offering (markets were nervous at the time amid Donald Trump’s escalating tariff rhetoric). Huang held the deal together, buying $250 million worth. Since then, the shares have done pretty well. Headed into earnings this week, the stock was up 160% from the IPO, but down nearly 45% from the June highs.

The stock was indicated sharply lower on Tuesday following the disclosure of a data center development “delay,” which forced the company to cut its full-year revenue guide to $5.05 billion to $5.15 billion. The company had tipped sales of up to $5.35 billion.

I’m not sure what to make of that, actually. Maybe it’s nothing. CEO Michael Intrator described the hiccup as “temporary” and said the value of a contract impacted by the snafu isn’t affected. In an interview with Bloomberg, he said, “[T]he data center provider is frustrated, we’re frustrated, the client is frustrated — Everybody’s frustrated.”

That seems like a lot of explaining for a glitch that amounts to a comparatively small $200 million reduction to the top-end of a revenue guide. I mean, it’s all about forward-looking sales when it comes to high-growth tech, and CoreWeave’s position as the leading “neocloud” means the market’s especially sensitive to any disappointments that might be an early indicator of waning demand. But this doesn’t sound like that.

As far as the rest of the numbers, the company’s net loss (“Of course we aren’t profitable, why would we be?”, to make the obvious joke) narrowed to $110 million from $359 million during the same period a year ago. Q3 sales were $1.36 billion, up more than 130% YoY. The backlog stood at almost $56 billion, right in the middle of the consensus range.

Most importantly, CoreWeave signed up Meta for a $14.2 billion deal. One criticism of CoreWeave says it’s far too dependent on a single client: Microsoft. The company went on to say OpenAI made a new $6.5 billion commitment and management alluded to a new contract with an unnamed “leading hyper-scaler.”

I’m reluctant to gamble on stocks like this for a bevy of what I assume are obvious reasons, but I gotta say: If the pre-market losses hold, CoreWeave will trade down almost 50% from the highs, albeit still up more than 100% from the IPO price. Again, they’re in pole position when it comes to wiring up data centers to warehouse AI hardware. And as Sam Altman put it, “I do guess that a lot of the world gets covered in data centers over time.”

I remain very skeptical that the technology will develop along that arc. As I’ve put it over and over again, technology gets cheaper and smaller over time, not more expensive and expansive. But what do I know, right? If it’s all about massive data centers, then CoreWeave’s a pure play. There are more caveats. Read on.

CoreWeave’s operating income margin was just 4% in Q3. That missed consensus by… well, by quite a lot. Wall Street was looking for 6.5%. A 250bps margin miss when analysts were only expecting 650bps in the first place is pretty rough, and suggests spending’s unbridled.

Further, this business is highly levered. The model “naturally” entails borrowing against the company’s Nvidia chips to fund capacity buildout. That’s fine as long as the market for AI compute remains supply-constrained. The second it’s oversupplied — i.e., the moment demand stumbles — boy, oh boy would that be bad for a company that borrowed against chip collateral to build warehouses for those same chips.

Complicating this situation further is the distinct possibility that the chips will become obsolete at faster and faster rates. If true, that’s doubly problematic: It means rent deflation (i.e., lower revenues) on the older chips and, by the same token, it presumably means the value of the collateral backing a lot of the company’s financing will be subject to haircuts of some kind to account for their obsolescence.

Finally, it’s not entirely obvious why the largest hyper-scalers need a middleman, or if it’s obvious now (and I guess it is, because after all, CoreWeave counts at least three of the hyper-scalers as clients) it won’t be soon enough. If I’m one of the largest, most dominant companies on Earth, I’ll just buy my own chips, build my own data centers and hire my own engineers, thank you very much. In fact, I might even buy your engineers. Or maybe your whole company. (That latter possibility may be the best bull case.)

The good news for CoreWeave investors is that everyone involved has already thought of, about and through all of that. Of course, recognizing potential problems and having a viable plan to address them are two different things, but to anyone pondering the stock I’d say this: These issues are so glaringly obvious, and CoreWeave’s important enough to this ecosystem, that you can be reasonably confident management and the company’s key stakeholders are actively engaged in around-the-clock efforts to mitigate the risks.

In other news, SoftBank sold its entire stake in Nvidia for more than $5.8 billion. Don’t worry: They’re plowing the proceeds back into AI bets.