The Money Is Coming From Inside the House

Nvidia reported $96.22 billion in revenue this week for its second quarter, against analyst consensus of $92.11 billion, and guided to $108 billion for the third. A day or so later, the Wall Street Journal reported that the company had paused the revenue-sharing financing deals it announced back in July under the name AI Compute Partnership. The program was aimed at smaller cloud providers who need to spend billions on Nvidia chips and the data centers around them before they can rent out an hour of anything. What WAS in place, but is no longer, was that Nvidia and the provider would agree on a base hourly rental rate that covers the provider's costs, Nvidia would provide credit support against the chip purchases, and Nvidia would take 50% of whatever revenue comes in above the base rate. The reasons for the pause, about eight weeks after launch, were internal worries about antitrust exposure, plus prospective partners getting irritated at how much control Nvidia wanted over the operations it was financing.

So the chip vendor extends credit so the customer can buy the chips, then takes half the upside on the rental revenue those chips produce. I would call that a loan with profit participation before I called it a sale, and I think Nvidia's own lawyers apparently agreed.

BUT, the paused program was also the small version of something much larger that is still running. OpenAI has committed to spending roughly $1.4 trillion over eight years on data centers and compute, against annual revenue somewhere near $13 billion. In contrast, simultaneously, NVIDIA has committed as much as $100 billion to OpenAI directly. OpenAI has committed hundreds of billions to Oracle and Microsoft for capacity, and Oracle and Microsoft buy NVIDIA GPUs to build that capacity out. Bloomberg and others have now traced more than $800 billion of these interlocking arrangements through the AI supply chain. Morgan Stanley reportedly expects Microsoft's entire Azure AI growth for the fiscal year that ended in June, north of $20 billion, to come from OpenAI, a counterparty in which Microsoft also holds a large stake. If you try to name the party in that chain who put outside money at risk on the proposition that end demand exists at these prices, somebody who is not also a supplier, customer, or shareholder of one of the others, I think you will end up struggling. While this is good in many ways (people are putting their capital where their words are), it also creates a lot of interdependence.

As with so many things nowadays, this feels like a replay; in this case, Telecom ran this experiment 25 years ago, and the filings are still on EDGAR.

Specifically, in the late 1990s the hot buyers of network equipment were the CLECs, the competitive local exchange carriers, startups laying fiber to compete with the Baby Bells under the 1996 Telecom Act. They had orders and no cash, so the equipment makers lent them the purchase price. Lucent committed $8.1 billion in vendor financing over the period, including a $2 billion credit line to WinStar Communications that WinStar drew on to buy Lucent switches, and the switch sales went into Lucent's reported revenue, where analysts read them as demand. Between 1996 and 2001, the sector overbuilt by an estimated $60 billion, and once outside funding dried u,p, the upriers started failing: Covad, Focal, McLeod, NorthPoint, and then, in the spring of 20011, WinStar itself, days after Lucent declined to advance the final $90 million on the line. That single relationship cost Lucent a $700 million write-off. The whole book deteriorated even faster than the WinStar piece did, with bad loans going from 2.6% of Lucent's financing portfolio at the end of 2000 to 60% a year later (Nortel's went from 25.5% to 80% over roughly the same stretch). Lucent took billions in provisions against customer loans, shed most of its workforce, and was eventually absorbed by Alcatel, having discovered that a meaningful slice of its late-90s revenue had been its own treasury making a round trip.

And before anyone emails/texts/DMs me to complain that "this time it's different", YES, vendor financing on its own is not a scandal. IBM Global Financing has been lending customers the price of IBM equipment since 1981, and it worked for four decades because the collateral was mainframes running payroll at companies with thirty years of verifiable cash flow. In THIS case, Lucent's collateral was the business plan of a five-year-old CLEC whose model assumed the 1999 growth curve was permanent. And, unfortunately, disclosure doesn't rescue it either, since Lucent's loans were disclosed too, in the same filings everyone now cites as the warning nobody read. The thing that went missing in the CLEC years, and is missing now, is an outside party who validated the demand with their own capital before the revenue got booked.

Which brings me back to the pause. Nvidia had just posted the best quarter in its corporate history and could have coasted on the program for another year. Instead, eight weeks in, somebody internal looked at the antitrust exposure, or at the $11 billion-plus already lent against GPUs across the neocloud sector before this program even existed, or at what taking half your customers' upside implies about eating half their downside, and shut the thing off. The company with the best view of real rental demand on earth got offered a machine for manufacturing more of it, and declined.

None of which means AI demand is fake. A lot of it is real, VERY real, and expensively so. But some fraction of the headline numbers is the same dollar passing through three income statements; nobody knows the size of that fraction, including the participants, and we find out when a payment gets missed somewhere in the circle. The 1996 to 2001 fiber did get built, the builders mostly died, and the rest of us spent a decade lighting up dark strands bought for cents on the dollar. The capacity outlived the counterparties. WinStar felt like growth too, right up until the $90 million did not arrive.

Want to learn how intelligent data pipelines can reduce your AI costs? Check out Expanso. Or don't. Who am I to tell you what to do?

NOTE: I'm currently writing a book based on what I have seen about the real-world challenges of data preparation for machine learning, focusing on operational, compliance, and cost. I'd love to hear your thoughts!