Compute Is Becoming an Asset Class — and Nobody Has Priced the Downside Yet

A year ago, an AI data centre was a line of capital expenditure on a hyperscaler's balance sheet. It is turning into something else: a financeable, leasable, insurable asset with its own capital structure. The deals we tracked over the past three weeks are not, individually, dramatic. Read together, they describe the machinery of an asset class assembling itself in public.

Start with the lease. SpaceX has signed a multi-year compute lease with Reflection AI, giving it priority access to Nvidia GB300 GPUs in the Colossus 2 data centre — worth up to $6.3 billion through 2029. Note the structure, not just the number. This is not a cloud contract billed by the hour; it is a long-dated claim on physical capacity, the kind of instrument that looks much more like commercial real estate than like software.

Then the debt. A Blackstone-led consortium has provided a $10 billion debt facility to Firmus Technologies to build up to 1.6 gigawatts of Nvidia-backed AI data centres across Australia. Debt is the tell. Equity investors buy a story; lenders require a cash flow they can model, collateral they can seize, and a term over which both survive. When a consortium of that size writes a facility of that size, someone has convinced a credit committee that AI compute produces predictable, secured, bankable revenue.

Then the revenue share. In a parallel arrangement, Nvidia will supply 170,000 GPUs to the same company under a revenue-sharing partnership, as Firmus builds a 360MW facility in Batam, Indonesia, targeting up to $30 billion in cloud revenue over six years. The supplier is no longer merely selling hardware; it is taking a position in the operating economics of the buyer. That is vendor financing with the serial numbers filed off, and it is what chip vendors do when they want demand to exist.

And finally, the one that made me sit up: insurance. A Chinese partnership is co-developing insurance-backed security solutions for AI data centres and what the release calls "token factories". Insurance is not a technology milestone; it is a maturity signal, and a strangely honest one. Underwriters only price what they can model, and they only model what has enough operating history to produce a loss curve. Nobody insures a science project. The moment compute becomes insurable, it has been reclassified — from experiment to infrastructure.

Even the corporate structures are re-forming around the idea. Jet.AI's shareholders approved a combination that spins out its aviation business and leaves behind a pure-play AI infrastructure and cloud services company. Companies do not restructure themselves into a category unless they think the category has a multiple attached to it.

## What financialisation actually buys

The bull case is straightforward, and I think largely right in the near term. Infrastructure that can be leased, levered and insured can be built far faster than infrastructure that must be funded out of retained earnings. If you have to pay for a gigawatt of capacity from cash flow, only about five companies on earth can play. If you can finance it against contracted revenue, the field widens to anyone who can sign a creditworthy tenant. That is precisely how railways, telecoms and shipping were built, and it is why the AI build-out is now spilling into Indonesia and Australia rather than concentrating in five US metros.

It also explains a pattern we have been tracking for months: capital moving down-stack, into the plumbing. The interesting money is no longer chasing the model layer. It is chasing power, land, interconnect and silicon — right down to optical I/O chiplets that move data between accelerators as light rather than electrons. Investors are behaving as though the durable returns sit in the pipes.

## The thing nobody has priced

Here is my discomfort, and I want to be careful to state it as a question rather than a prediction.

Every one of these structures assumes the asset holds its value over the term. A $10 billion debt facility, a lease running to 2029, a six-year revenue-share target — these are long-dated instruments written against hardware with a brutally short technological half-life. GPUs do not depreciate like rail track. Colossus 2 is being leased on GB300s; the useful economic life of a specific accelerator generation, in a market where the frontier vendor ships a new architecture roughly annually, is not obviously six years. It may be. But the financing is being priced as though the question is settled, and it is not.

There is a second, subtler mismatch. The bankability of these deals rests on contracted revenue from AI companies — and a striking share of the AI companies signing those contracts are themselves financed by the same capital cycle. Nvidia supplies GPUs to Firmus and shares Firmus's revenue. Nvidia also invests in the labs that rent capacity. Circular flows are not inherently fraudulent — vendor financing built the telecom boom — but they do mean that the "predictable cash flow" underwriting the debt is less independent of the equity story than a credit model would like.

And then there is the demand-side irony sitting in the same week's deal flow. AWS and Nvidia are making open-weight Nemotron 3 models fine-tunable through a managed SageMaker API, removing the operational tax that made open weights hard to use. Every step like that pushes inference cost per token down. Cheaper, more efficient inference is unambiguously good for the world and ambiguous for anyone who has borrowed ten billion dollars against the assumption that compute stays scarce and dear. Efficiency is the one risk an infrastructure lender cannot hedge, because it arrives disguised as progress.

## What I would watch

I am not calling a bust. The demand is real: enterprises are signing multi-year, billion-dollar AI transformation contracts — HCLTech's roughly $1.14 billion, five-to-six-year deal with a Fortune Global 50 client in Europe is a serious commitment from a serious buyer. Something is being bought here, and it is not vapour.

But the near-future question is not whether AI compute earns a return. It is what happens the first time an operator's contracted tenant fails to renew, and a lender discovers what a three-year-old GPU cluster actually fetches in a liquid market. We do not know that number. No one does, because that market barely exists yet. The insurance deals are, in a sense, the first attempt to invent it.

Asset classes are made, not born, and they are usually made twice: once on the way up, when the structures are invented, and once on the way down, when the recoveries are discovered. We are watching the first half with unusual clarity. The deal record is public and the structures are legible. Anyone underwriting this ought to be asking not what compute is worth while it is scarce, but what it is worth the morning after it stops being.