The shortest distance between two balance sheets

NVIDIA is putting $2 billion of equity into CoreWeave at $87.20 a share, deepening a relationship that already runs through chip allocation, a multi-year compute backstop, and a software distribution agreement that now extends CoreWeave's stack to other cloud service providers and enterprises. CoreWeave, for its part, is targeting 5 gigawatts of AI factory capacity by 2030 and has just expanded its capacity deal with Meta. The press release calls this an alignment. I'd call it something more specific: NVIDIA financing the demand for NVIDIA's own product, through a counterparty whose largest customer is Microsoft, whose second-largest is now Meta, and whose single most important supplier is NVIDIA.

This is not a scandal. It is, however, worth being precise about what it is, because the shape of this deal tells you more about where the AI infrastructure cycle has got to than any capex number on a hyperscaler's earnings call.

Start with the mechanics. NVIDIA sells GPUs to CoreWeave. CoreWeave rents those GPUs, packaged as AI factories, to Microsoft, Meta, OpenAI, and a growing book of enterprises. NVIDIA also holds equity in CoreWeave, now meaningfully more of it, and has, in prior arrangements, committed to taking compute capacity from CoreWeave itself if needed. The cash NVIDIA puts in at $87.20 a share funds, among other things, CoreWeave's ability to keep buying NVIDIA chips at the rate its customers demand. The equity stake appreciates if CoreWeave's revenues compound. CoreWeave's revenues compound if it can secure more chips and more power. The chips come from NVIDIA. The power is the actual constraint.

NVIDIA is financing the demand for NVIDIA's own product, through a counterparty whose single most important supplier is NVIDIA.

If you've read the financial press on this for the last eighteen months, you'll have seen the word "circular" used to describe arrangements like this, the OpenAI–Microsoft–NVIDIA triangle, the Oracle–OpenAI compute deal, the various Anthropic–Amazon–Google permutations. The word is doing a lot of work, and not always honestly. So let me try to separate what's genuinely circular from what isn't.

What isn't circular: end demand. Microsoft is paying real money for CoreWeave capacity because Microsoft's customers, and Microsoft's own internal AI workloads, are paying real money for inference and training. Meta's expanded deal is funded by Meta's advertising business, which throws off cash at a rate that makes the AI capex line look almost reasonable. OpenAI's revenues, whatever you think of the unit economics, are demonstrably climbing. The dollars at the customer end of the chain are not NVIDIA's dollars coming back around. They are Fortune 500 IT budgets and consumer subscription revenue.

What is circular: the financing of the middle layer. CoreWeave is a balance-sheet-intensive business that buys depreciating hardware on debt and equity, parks it in leased data centres, and rents it out on multi-year contracts. Its ability to scale faster than its operating cash flow allows depends on capital markets believing the contracts are good and the hardware will hold its value. NVIDIA's $2bn equity cheque is, among other things, a signal to those capital markets, a vote of confidence priced at $87.20 that makes it cheaper for CoreWeave to raise the next tranche of debt. That is a real economic transfer, and it has the effect of underwriting GPU demand at the margin.

The question isn't whether this is happening. It plainly is. The question is whether it matters.

The bull case runs roughly: NVIDIA is doing what every dominant platform vendor does at the start of a category, seeding the ecosystem that consumes its product. Microsoft did it with software developers in the 1990s. Apple did it with the App Store. AWS did it with startup credits. Vendor-financed adoption is how new categories are built, and the test of whether it's healthy is whether end demand eventually shows up to retire the financing. On current evidence, end demand is showing up. Hyperscaler capex guidance for 2026 is up again. Inference revenues are climbing. The contracts CoreWeave is signing are with counterparties, Microsoft, Meta, who can pay them.

The bear case runs: vendor financing of demand is exactly what late-cycle behaviour looks like, and the precedents are not encouraging. Lucent and Nortel both extended supplier credit to telecoms customers in the late 1990s; the equipment got bought, the customers went bust, the equipment came back at cents on the dollar, and the supplier balance sheets ate it. The structure here is different, CoreWeave isn't buying chips on NVIDIA paper, it's taking equity from NVIDIA, but the economic substance has a family resemblance. If AI inference demand at the enterprise tier disappoints, the contracts unwind, the GPUs get repriced in the secondary market, and the equity stake is worth what the equity stake is worth.

The structure is different from Lucent–Nortel. The economic substance has a family resemblance.

I find myself unable to land cleanly on either case, and I think that's the honest answer rather than a hedge. The end-demand question is genuinely open. Enterprise AI deployment is happening, but unevenly, and at a pace that's a function of organisational absorption rather than capability. The hyperscalers are real customers with real revenues. The second-tier enterprise customers CoreWeave is now signing through this expanded distribution agreement are a different proposition, many of them are still in the proof-of-concept phase, and proof-of-concepts don't pay for 5 gigawatts of capacity.

What I am confident about is that the deal tells you NVIDIA's strategic posture has shifted. A year ago NVIDIA was rationing supply and choosing winners through allocation. Today it is putting equity in to make sure the winners it chose can keep buying. That is not a company that thinks demand will take care of itself. It is a company that has decided the binding constraint on its own growth is the capacity of the middle layer to scale, and is willing to use its balance sheet to relieve that constraint.

A year ago NVIDIA rationed supply. Today it is putting equity in to make sure its chosen counterparties can keep buying. That is a different posture.

The 5-gigawatt target is the number to watch, because it is, in the end, a power number rather than a chip number. CoreWeave can buy all the GB300s NVIDIA can ship; it cannot conjure substations, transmission interconnects, or the cooling water that a gigawatt of inference draws. The constraint has moved. NVIDIA's $2bn doesn't solve it. It just buys CoreWeave more time and more credibility while the physical world catches up, or doesn't.

That's the deal. Not a scandal, not a stitch-up, not the end of the cycle. A specific, legible move by the dominant supplier to underwrite the demand for its own product through a counterparty whose growth is now constrained by electricity. Whether that's clever or late depends entirely on what the enterprise inference revenue line looks like in 2027.

I wouldn't bet against NVIDIA. I also wouldn't price CoreWeave as if the financing structure were incidental to the story. It isn't. It's most of it.

Footnotes

¶