AI Data Center Mania Conjures the B-Word. Is It Something to Fear?
Plus, the AI names going to Meta & Musk's new X
The Week in Short
CoreWeave’s leverage raised eyebrows after its stock’s 4X run, while Nvidia broke the $4 trillion barrier. Cautionary notes on the boom came with new data from Ramp that showed an AI spending slowdown from some corporate customers.
Paid subscribers only: Mark Zuckerberg was busy poaching top AI research names from OpenAI & others, while Elon Musk’s launch of xAI’s latest model was overshadowed by the Grok chatbot going full MechaHitler. A series of mega-rounds propped up venture funding totals for Q2. Benchmark-backed Manus AI moved to Singapore, and AI regulation stayed on track in Europe, and maybe California. We’ve got a list of the people Zuckerberg has been poaching + a table of the funding rounds sucking up all the oxygen.
The Main Item
AI Exuberance Brings Debt into Play as Markets Charge Ahead
There have been plenty of signals recently that the AI investment mania may be tipping into bubble territory, or at least nearing a top.
The 4X jump in the share price of data center operator CoreWeave in the months following its lackluster March IPO is one such sign. Nvidia’s run-up to become the first $4 trillion company, though backed up by extraordinary profit growth and margins, is another.
Meta’s success in recruiting an all-star AI team with massive pay packages can also be read as an indicator that the smart money is ready to cash out. Superstar hires including Nat Friedman, Daniel Gross, Alexandr Wang, and a handful of OpenAI researchers apparently judged that they should take the payday while the getting was good and join Zuckerberg’s AI Justice League.
Bubbles, though, are a feature of every great wave of tech innovation, and are not necessarily to be feared. Much depends on where you sit in the cap table.
The AI data center mania in particular evokes comparisons with the telecom bubble of 2000-2001, where over-investment in a revolutionary technology — fiber optics — brought massive financial losses, but also built a lot of infrastructure.
Paul Kedrosky has a good piece on the analogy. First he lays out the structure of Meta’s reported $29 billion data center plan. All but $3 billion of the money would come from debt taken on by private-equity backed special-purpose vehicles, and that debt would cost 2 to 3 percentage points more than if Meta just borrowed the money itself.
The purpose of that is to keep it off Meta’s balance sheet. But with Meta owning a big chunk of the SPVs and the debt backed by its contractual commitments to use the data centers, it’s hardly immune to a shift in sentiment if demand falters, as it did in the early 2000s.
“We are already spending more now than then, driven by the same sort of hyperbolic projections of future usage and spending,” Kedrosky writes. “They may turn out to be correct, but if the returns aren't there quickly, the thin equity cushion above data center SPV debt holders will be speedily wiped out.”
During the telecom bubble, “vendor financing” was the sleight of hand, with companies like Lucent at the peak in effect advancing their customers the money to buy their gear. That served to disguise a collapse in underlying demand.
The “Nvidia economy” of AI datacenters could carry similar types of risk. The huge cost of the facilities is mostly due to the huge cost of Nvidia processors, and companies led by CoreWeave are essentially middlemen sitting between the chip-maker and the AI companies scrambling for access to GPUs. Nvidia, notably, owns 7% of CoreWeave.
That can be a good business as long as demand and pricing are only going in one direction, and analysts say CoreWeave has been shrewd. But it’s not by chance that CoreWeave’s founders are commodities traders; multi-billion-dollar data centers are in some sense a three-way trade involving the cost of the chips, the price they can be rented for, and the structure and cost of the financing.
That’s why skeptics this week were seizing on CoreWeave’s depreciation schedule as a danger signal. It writes down the value of its Nvidia inventory over six years, even though the chip company has been on a two-year product cycle and other firms depreciate the chips more quickly. Slower depreciation makes the numbers look better, but could readily lead to big write-offs later. (Forbes has a good breakdown of CoreWeave’s risks.)
Analysts and investors were already backing off on the company this week after the big spike, with shares declining about 15% in the wake of CoreWeave’s acquisition of data center partner Core Scientific.
A comparison of the current situation with the telecom and dot-com bubbles of 25 years ago offers some comfort. Back in 2000, the entire Nasdaq was trading at a triple-digit price-earnings ratio, compared with a high-but-not-insane 50 or so for Nvidia currently. There were some 476 IPOs in 1999, many involving companies with heavy losses and minimal revenue. Nobody can get away with that today.
Still, the similarities are there. Companies such as Global Crossing were leveraged bets on future demand (and pricing) for broadband services that turned out to be far too optimistic. The more debt in the system, the more vulnerable it is to a forecasting miss.
That’s why new research from fintech provider Ramp this week raised a lot of eyebrows. With purchasing data from its customers, Ramp can see trends in actual business spending on AI, and for the past 18 months it has been steeply up and to the right — until very recently.
The percentage of businesses with paid subscriptions to AI models, platforms, and tools ticked down slightly from 42.5% in May to 42% in June, after exploding upwards from 25% at the beginning of the year, Ramp said.
“We’re seeing more businesses switch to free, sometimes less advanced AI tools as pricing fatigue hits customers,” said Ramp economist Ara Kharazian. He pointed to coding app Cursor’s recent rollback of price hikes as an indicator of the pressures.
Kharazian added in an email that “API spend on AI models continues to increase.”
Still, it’s fair to say that “pricing fatigue” is not a phrase that’s very welcome in the Nvidia economy.
Even though many of the financial risks in AI lie with VCs and private-credit providers, public market investors are very much implicated too.
The CoreWeave run-up was mostly being driven by retail investors and currently has meme-stock characteristics — a valuation disconnected from the underlying business. Nvidia’s surge is more closely related to its results, and projections of continued torrid growth. But any wobble in the momentum could end the party in a hurry.
The trick, of course, is that recognizing that there are bubble valuations out there is not especially helpful in itself. We could still be early in the bubble-inflation cycle, and even a downturn could end up being analogous not so much to the crash of 2001 as the long-forgotten blip of 1997, when the Asian financial crisis tanked markets and early internet mania looked like it could be a fad.
If you take the long view, as economist Viktor Shvets of Macquarie does in a Bloomberg piece, none of this matters much. AI might be a perfect recipe for “anger, frustration, grievances, and hence, deep and persistent polarization, both locally and globally,” according to Shvets, but that doesn’t mean it isn’t a business opportunity for the ages. He argues that the AI cycle is both much bigger and moving much faster than previous tech revolutions, so any downturn now would be in the category of “bumps in the road.”
One lesson from the past though is likely to hold: retail investors are the most exposed, followed by employees and other equity-holders, and then creditors lining up for a slug-fest. Depending on your time horizon, bumps in the road could be pretty painful. Only some will come out stronger.
Newcomer Podcast
Shaun Maguire’s Growth Hack
Tom and Madeline have rejoined Eric on the Newcomer Podcast just in time to debate Meta’a aggressive hiring strategy and Shaun Maguire’s latest Twitter firestorm.
Plus, we make the case for why Grok 4 can stand out despite the chatbot’s sudden tendencies to go full “MechaHitler” in the replies.
Talent Wars
See the AI Researchers and Executives Lured by Zuckerberg With Extra-Fat Packages
The race to build superintelligence is nothing if not expensive.
Over the past month Meta has been using the promise of life-changing compensation to woo star AI researchers and executives to its Superintelligence Labs. After spending a whopping $14.7 billion to get a stake in Scale AI and steal its CEO, Alexandr Wang, Zuckerberg followed up by buying out a chunk of Nat Friedman and Daniel Gross’ venture fund and poaching close to a dozen OpenAI researchers, as well as a top AI exec at Apple.
As The Verge notes, not all engineers are willing to trade millions of dollars for the grueling work schedule Meta expects from its new hires. And OpenAI is making moves too, nabbing four researchers from Tesla, xAI, and Meta, Wired reported.
The cost of talent is so steep that OpenAI’s stock compensation commitments for the year have outpaced its revenue estimates, per The Information.
Here’s a list of all of Meta’s big AI hires over the last month whose moves have been made public, including any details we could find on their compensation.