Circular Deals & Supply Chain Dynamics

The big story of 2025 has been the direct transfer of data center demand risk from hyperscalers to NeoClouds and chip companies. Circular deals crystallize this dynamic.

Last year, in “The AI Supply Chain Tug of War”, I wrote about pairwise game theoretic dynamics up and down the supply chain:

In the supply chain, risks are transferred from suppliers, who need to build CapEx to manufacture products, upstream up to their customers, who pay a margin that compensates for this capital expenditure over time.

Each player wants to maximize profit while minimizing risk. This creates supply chain conflict, which lurks behind the scenes, and exposes itself in pairwise game theoretic interactions between suppliers and their customers.

If you think about a supply chain (say for clothes), the end consumer pays some price for the goods, and then those dollars get split by all the vendors up and down the supply chain. In a mature supply chain, everyone makes money, and hence, the supply chain continues.

In AI, by contrast, the end customer is not funding the supply chain. Rather, the supply chain is funding its own buildout, in the hope that more paying customers will eventually arrive. Said differently, the amount of capacity now being built vastly outstrips current demand.

Early in technology cycles, this type of risk is normal. For example, during the mobile transition, investors financed the build-out of massive transportation and food delivery fleets, in the hopes that eventually the math would tie out. It ultimately did. In other cycles, such as railroads or the telecom bubble, investor “animal spirits” get ahead of reality, and there is so much overbuilding that it can take a long time for the math to work, if it ever does.

What’s unique about the AI cycle is a) The sheer scale at which this is happening, and b) That this “demand risk” (or “demand hot potato”) has been financed by mega-corporations of world-historic scale. From the Supply Chain Tug of War piece:

Today, Big Tech companies have stepped up to alleviate some of this tension. They are acting as risk-absorbers within the system, taking on as much demand risk as they possibly can, and driving the supply chain toward greater and greater CapEx escalation.

In a separate piece on the Game Theory of AI CapEx I explained further:

The key to understanding the pace of today’s infrastructure buildout is to recognize that while AI optimism is certainly a driver of AI CapEx, it is not the only one. The cloud players exist in a ruthless oligopoly with intense competition. This is no small prize to defend—the cloud business today is a $250B market, roughly the same size as the entire SaaS sector, combined. The cloud giants see AI as both a threat and an opportunity and do not have the luxury to wait and see how the technology evolves. They must act now.

If the story of 2024 was the hyperscalers stepping in to subsidize land and power, natural gas generator construction, fab buildouts, and much more, then the story of 2025 has been the shift in who is holding this hot potato. This story has been quietly playing out since January, but has now become a centerpiece of public attention with the announcement of recent “circular financing” deals led by chip companies.

It all started with Microsoft walking away from two data center leases at the very beginning of the year. This sent a message to the market: Microsoft was making it known that they would no longer be paying the bill for the entire AI ecosystem’s data center buildout. It’s not just Microsoft. Investors think that Amazon is underweight on GPU purchases, relative to the scale of AWS, because of its more limited risk appetite vs. peers. Both companies have suffered this year as investor sentiment reflected that these companies were “underinvesting in AI.”

At the same time, Oracle and Coreweave stepped up this year, absorbing the latent demand left on the table by Microsoft and Amazon. Both companies have been heralded as winners, with surging stock prices and investor interest. Nebius and a longer list of NeoClouds have been ramping up as well. But these companies, even combined, are much smaller than Microsoft and Amazon, and so the recent moves by Nvidia, AMD and now Broadcom were ultimately inevitable: These companies are stepping up to absorb some demand risk, too. Previously, all of the risk was sitting with their customers, the hyperscalers.

There are now four deals on the table: First, Nvidia announced that it will invest $10B in OpenAI for every 1 GW of new data centers, up to 10 GW. Then last week, AMD announced that it will build 6 GW with OpenAI, in return for up to 10% equity in warrants. And then today, Broadcom announced that it too will build 10 GW of capacity. Finally, there is a fourth less well-publicized deal from earlier this year, where Nvidia is backstopping demand for Coreweave, should future revenue not materialize.

How should we look at these circular deals? The simplest way to look at it is as vendor financing. Chip companies are subsidizing their customers — effectively lowering prices to stimulate more demand. Additionally, they are priming the pump: By showing that further development (outside of the traditional hyperscalers) is inevitable, they hope to attract more capital from new parties.

A lot of people have been talking about how the AI CapEx buildout is debt financed, but in reality, what’s interesting about the circular deals is precisely that they are not debt financed. At least for now, debt investors are unwilling to bet on data center demand in 15-years (this is what they would be doing if they agreed to unsecured loans against new data center construction). Instead, they have been requiring the credit-worthy companies to back-up their loans. Hence, Microsoft underwriting the early Coreweave deals, Oracle underwriting Stargate, and now Nvidia and AMD underwriting whatever comes next. In fact, when I talk to debt investors, they balk at taking on the demand risk directly — which is precisely why first the hyperscalers, and now the NeoClouds and chip companies, are having to step up.

Another aspect of the deals is that they seem more like statements of intent rather than firm commitments. Morgan Stanley’s Tom Wig reported in his newsletter that “the 10 GW OpenAI deal is an ‘aspirational’ number (could be higher or lower).” On the AMD side, the 6 GW deal seems to hinge on some technical milestones being cleared, which Jensen himself has called into question. Details are still emerging around the Broadcom deal.

The most important aspect of the deals is not what is said, but what is left unsaid. And that is the dollar cost of these deals, the revenue that will be required to pay them back, and where the rest of the money is going to come from. Increasingly, all of the deals are being quoted in GW (gigawatts of power) instead of dollars. To put the concept of a GW in context, a very big data center pre-AI used 50 MW of power, so even a single 1 GW data center is as big as 20 cloud data centers. The concept of GW is being used to reflect that power is the new constraint on data centers, rather than money. But this has an effect of obfuscating just how much spending we’re talking about.

Here’s a chart that does the conversion from GW back to dollars and shows the implied revenue needed to pay back these dollars. For this analysis, I’m conservatively using $40B per GW for the cost of data center construction, inclusive of chips and interconnect. Going forward, Nvidia has said that the total cost will be up to $50-60B per GW for upcoming Vera Rubin based data centers, so this is a lower bound.

The columns in the chart above reflect 1 GW (unit cost), 6 GW (size of the AMD deal), 10 GW (size of the Nvidia and Broadcom deals), and then 100 GW and 250 GW, which are estimates now being widely used for base case and bull case forecasts on what CapEx through 2030 could end up looking like. These numbers are now also being used by AI leaders to describe their goals. Like my analysis in AI’s $600B Question, I assume a 50% gross margin for the ultimate user of the compute (and conservatively, don’t factor in a hyperscaler margin), to arrive at the lifetime revenue required to payback these investments.

AI is always full of surprises, but it does feel like we’ve reached a new steady-state in terms of who is financing data center buildouts, and it’s logical that the semiconductor companies are now incentivized (game theoretically) to take on some of the risk for the buildout directly. It’s possible that debt investors get over their fears and join in next, and it’s also possible that sovereigns increase their spending on data centers, as these are now being marketed as national assets. Equity investors may also increasingly finance this growth, to the extent that they fund NeoClouds directly.

One year later, the conclusion of AI’s Supply Chain Tug of War is only more true:

Supply chain players understand AI’s $600B question, and they are working to navigate it—maximizing their profit margins and minimizing their demand risk. The result is a dynamic tug of war between some of the most sophisticated companies in the world.

Today, the tug of war has resulted in a temporary equilibrium. Supply chain players are offloading their demand risk to Big Tech, to the maximum degree possible. Big Tech companies—either due to AI optimism or oligopolistic competition—are stepping in to absorb this risk and keep CapEx cranking.

This equilibrium is fragile: If at any point the tech giants blink, demand all along the supply chain will decline precipitously. Further, the longer the Big Tech companies continue to double down on CapEx, the more they are at risk of finding themselves deeply in the hole should AI progress encounter any stumbles.