Nvidia's $500bn compute financing platform — the moment GPUs became collateral

The question hanging over AI capital expenditure has been the same one for two years. Where does the money come from?

There is a limit to how much datacenter capacity can be built out of hyperscaler cash flow, and the 2026 buildout plans had already passed it. Nvidia's answer, announced on August 10, went in an unexpected direction. Not we will spend more of our own, but we will build the road other people's money travels on.

1. What was announced

The identity of the six tells you what kind of announcement this is. These are not venture funds. They are the largest managers of infrastructure equity and private credit in the world, and their usual assets are power plants, toll roads, ports and fibre networks.

The proposal is to add datacenters to that list — not as real estate, which lenders already understand, but with the compute inside them treated as part of the security.

2. What "investable asset class" actually claims

Jensen Huang said compute chips are becoming an investable asset class for the first time. The supporting argument has three parts, and each maps onto a standard test in infrastructure lending.

Aircraft finance is the closest analogy. A plane earns revenue, lasts decades, and can be repossessed and leased to a different airline. That is why banks accept aircraft as security, and why airlines can fund fleets with debt rather than equity.

Huang's argument is that GPUs now pass the same screen. If lenders accept it, the cost of capital for the entire AI buildout falls — a large share of what is currently funded with equity can be refinanced with cheaper debt.

If they do not accept it, the $500bn stays a target.

3. The circular financing objection, and the number 25

The criticism arrived with the announcement. If Nvidia helps its customers raise money, and that money buys Nvidia chips, then revenue reflects a loop Nvidia constructed rather than demand the market generated.

Huang's rebuttal ran three ways.

Demand is genuine. The constraint is capital, not customers — capacity is being rationed, not searching for users. Underwriting is independent. Each capital partner performs its own due diligence on every project. Nvidia does not approve deals. Nvidia's exposure is bounded. What Nvidia provides is a residual value support mechanism of up to 25% in certain cases.

The third point carries the weight. A residual value guarantee means the guarantor covers the shortfall if the collateral is worth less than a set floor at the end of the term. A 25% cap means Nvidia absorbs at most a quarter of that risk and lenders carry the other three-quarters.

As long as that ratio holds, the structure is difficult to characterise as circular. But a disclosed cap is not a disclosed contract, and the individual platform terms have not been published.

4. Why Seoul rallied two days later

On August 12 the KOSPI closed at 6,579.04, up 3.68%. Samsung Electronics rose 6.68% and SK Hynix around 7%, with foreign buying concentrated in electronics. The first reason Korean analysts cited was this announcement, converted in local coverage to roughly 710 trillion won.

The chain is short.

1. Capital is arranged for datacenter construction
2. Datacenters need HBM and server DRAM alongside the accelerators
3. Those are made in Korea

Layered on top was reporting that the memory shortage was spreading from HBM into advanced packaging and wafers. The phrase repeated in Korean analyst commentary that day was that the AI bubble concern had eased — and the evidence cited for the easing was this $500bn. The session itself is covered in "KOSPI closes at 6,579.04."

5. What is unresolved

These are MOUs, not contracts. When each of the six platforms is established, under what structure, at what spread, is undisclosed. In infrastructure finance the gap between a headline target and deployed capital is routinely wide and slow.

No public data on used-GPU recovery values was located. The entire collateral argument rests on repossess-and-redeploy, and unlike aircraft or property, that market has no long price history. Accelerator generations turn over roughly every two years. How you underwrite the residual value of an asset whose successor arrives in 24 months is the weakest joint in the structure.

The 710 trillion won figure is a conversion and differs by outlet. The announcement is denominated in dollars.

One note on scale. $500bn is comparable to a full year of global semiconductor revenue. If capital of that size actually moves, the effects will not stop at chips — they run through power generation, cooling, construction and land. The conditional is still doing a lot of work in that sentence.