One of this week's biggest stories was Nvidia agreeing to invest up to $100bn in OpenAI, much of which OpenAI will use to buy GPUs from Nvidia. This is, depending on your view, either exciting or distressing. The exciting part is that Nvidia is so bullish on long-term demand for their products that they're willing to front the money for them in exchange for some of the upside from using them. The negative case is: what if OpenAI only raised that money from Nvidia because the best terms they could get were from someone trying to ramp up their own revenue?

Vendor financing reliably produces horror stories. One of the popular ones people talk about right now is all the lending that cash-flush telecom equipment companies did in the late 90s. The telecom buildout was capital-intensive, but telcos were betting on a massive increase in demand for data that would make all that spending worth it. They tapped equity and credit markets, but every bit of cash helped, so Cisco lent them the money they'd use to buy Cisco's equipment. In a way, this made perfect sense: Cisco's revenue had compounded 36% annualized over the prior half-decade, and they liked to do stock-based acquisitions, so they were in a solid cash position. The financial system already intermediates between cash-generators who want a return and cash consumers who can produce one, and Cisco could have argued that it had a competitive advantage in assessing which borrowers would be good for the money.

Of course, if they made that argument, they were wrong, and some of that equipment turned out to be, through many layers of indirection, a charitable contribution to YouTube and Netflix: the bandwidth demand projections were basically right, but given enough overbuilding, the price of using a fixed asset with near-zero marginal cost and significant overcapacity is close to zero.

There are older examples; a bowling boom in the early 60s was built on vendor financing from Brunswick, but that just meant that every city where bowling was trendy had lots of mostly-empty bowling alleys. That case seems like simple irresponsibility: if there's some trendy new consumer product, and you're a trusted brand selling it, you want to make sure your customers are patient and well-capitalized, because it'll be easy for them to overshoot and ruin the market.

In Nvidia's case, there are many moving parts—so many that it's hard to make direct comparisons with any prior episode of vendor financing:

1. They generated $70bn+ in free cash flow over the last four quarters, so they can afford a lot of lending before there's a serious risk to their business. Even if OpenAI somehow wipes out entirely, and Nvidia recovers zero cents on the dollar of that $100bn, Nvidia's not going anywhere. (At least in a world where that happens to OpenAI and nothing else is affected.)

2. Nvidia has a strategic interest in ensuring that there are multiple viable players who operate AI infrastructure, and probably in ensuring that some of them are cloud businesses that lease their infrastructure to third parties while others are buying GPUs in order to use them directly.

3. One of the reasons for that is that a big uncertainty in Nvidia's long-term model is where value and durable competitive advantage will accrue. One of the ways they hedge this is by ensuring that companies like CoreWeave, Nebius and Lambda Labs get access to compute, and making sure there's always a lab with an intimidatingly huge capital footprint, like xAI's. But, in the end the value accrual is hard to resist. Just as telcos don't have a good way to identify the cheap or expensive bits and mostly have to focus on volume, Nvidia's safest hedge against a world where one lab wins and can dictate terms to hardware providers (or just make its own hardware in-house) is to pick that lab, give it the resources to win, and own a piece of it in exchange.

4. Nvidia would probably prefer that scaling laws hold, and models keep improving linearly given superlinear increases in compute. That's a world where, every year, Nvidia's slice of global capex is bigger. But on the downside, a world where models stop improving so much is one where the useful life of GPUs is potentially longer, since more of them will be used for inference rather than training, and more AI companies will get clever about routing simpler queries to dumber models on older hardware.

There aren't many direct historical analogs to Nvidia's situation right now. OPEC in the 1970s comes close, and the global commodities supply chain during the Second World War is also a reasonable case, at least in the sense that there were some inputs (oil, rubber, bots, winter coats, etc.) that were make-or-break for various participants in the war, such that the ones who offered them had enormous negotiating leverage.¹ It's rare that there's just one swing producer, at a time when changes in the economy are dominated by one main factor, so the set of relevant examples is small.

It's still very likely that, even if AI produces superhuman intelligence and we either achieve global post-scarcity or get wiped out entirely, there will be a messy quarter or two in the meantime. If that happens, this deal and the recent Oracle one will be trotted out as examples of how crazy things got at the peak. But from the perspective of people building AI companies, there's still more downside risk than upside risk—as long as their competitors are scaling fast, the only choices are to match them or to be irrelevant. So in that sense, while it would be impossible to predict specifics in advance, deals like this are inevitable.

In The Diff, we’ve written quite a bit about vendor financing, cross-shareholdings, and other such complications, including:

• We’ve examined the general phenomenon of companies taking equity in their customers ($).

• The Nvidia debate ($) is similar to what it was in 2023, just with bigger numbers.

• We also covered the recent Oracle/OpenAI tie-up ($), and how recent tech history shows that investors tend to underreact to events like this.

• If you’re on the skeptical side of things, consider the general theory of grift.

• Here’s a writeup of a company on the receiving end of an equity investment from a big customer.

Join the discussion!

1  The US took advantage of that leverage by placing an oil embargo on Japan, but didn't use it that aggressively when negotiating with the Soviet Union.