CoreWeave: Are Data Centers Going the Way of the Mall?

CoreWeave, which builds data centers and leases compute power to AI companies announced a $14 billion contract with Meta this week, and last week announced a contract expansion with OpenAI with a total value of $22 billion.

CoreWeave is in turn pouring billions into concrete, steel, and power infrastructure, and investors cheer every new megawatt as if it automatically increases the company’s value. CoreWeave’s stock price was up 15 percent Tuesday, the day the Meta deal was announced.

But the market—which focuses so much on the energy component of data centers—is ignoring an extreme risk when it comes to data centers. The buildings may last for decades, but the hardware inside does not.

At a valuation of more than $61 billion with zero profits, CoreWeave is essentially a bet that GPUs will hold value longer than their real economic life.

CoreWeave: The Big Short?

DISCLOSURE: I’m not short, nor have I ever owned CoreWeave Stock (CRVW) in any capacity…yet. For now i’m just fascinated by the topic.

GPUs can lose 20 to 30 percent of their value every year while chip innovation cycles accelerate. Just this week, Huawei announced a ramp-up of AI chip production on par with NVIDIA, adding even more depreciation pressure. As The Financial Time’s Bryce Elder wrote in March: “Six years in AI is an eternity. Nvidia’s server-grade V100 GPU cost around $10,000 in 2019 and can now be picked up for a few hundred dollars.”

This makes CoreWeave less of a durable infrastructure company and more of a speculative vehicle. They raise money to build facilities, secure power and buy chips. But if depreciation accelerates further, if new technologies like efficient photonic compute or cheap ASIC inference chips reduce the demand for large GPU clusters, or if the AI bubble pops and demand dissipates, the economics could collapse quickly.

These are capital intensive, long term businesses trying to make money by slicing it into fractional short term flexibility.

Sound familiar? That was the WeWork business model.

What happened to WeWork when competition saturated the market and the economy took a dive? Bye bye predictable revenue stream…hello fixed costs.

When Infrastructure Acts Like Hardware.

I’m not saying data centers aren’t needed, they are essential to the economy. I’m just saying the business model appears to be nuts. Markets treat them like longer term assets while the value inside quickly depreciates.

Hardware has always been the most difficult part of tech, full of companies that once looked untouchable and then failed. Even Apple came close to collapse in the 1990s.

Today’s data centers may not produce chips, but they are exposed to the same forces that once wiped out hardware giants.

NVIDIA’s Game.

To me, CoreWeave and its peers feel less like the future of computing and more like shopping center operators who, after dining with tech and energy friends, spun up a business that NVIDIA was happy to fund and eventually sell to. Just look at Brookfield, which has committed $30 billion to building data centers in France and Sweden over the next 5 to 15 years.

NVIDIA is doing the same thing with OpenAI today, round-tripping a $100M investment that simply inflates its own sales.

Does this seem like a healthy business model?

For investors looking past the AI hype cycle, CoreWeave does not look like the future of data centers. It looks like the big short hiding in plain sight. But hey, what do I know about data centers anyway?

***PS: I would NOT recommend shorting this stock in the short term. The mania and momentum in the AI sector outweigh conventional logic. Consider this piece an exploration of today’s AI bubble, not investment advice.

If you liked this piece, subscribe to my Substack.