The GPU is the asset.
Everything else is the enclosure.

Every AI compute business, however it dresses itself up, sells the same thing: a GPU for an hour. That is the meter. Everything else — the campus, the substation, the cooling loop, the software layer — exists to make that hour available and to make it sellable again next hour.

Published on-demand rates across the neocloud market run roughly $2.20 to $8.60 per GPU-hour depending on generation, provider and configuration, with hyperscaler on-demand pricing higher again. Newer silicon commands more: current-generation parts list meaningfully above the previous generation, because the work they do per hour is worth more.

For everything that follows, take a round $6.00 per GPU-hour — a mid-market figure for a current-generation part, chosen because it is easy to multiply, not because it is anyone's actual quote. Substitute your own number and every figure below scales linearly.

Land, shell, power and cooling are indispensable, and we spend most of our time on them. But they do not bill anybody. The GPU is the only object in the building with a meter attached, and that single fact reorganises how the whole stack is financed.

An asset that meters can be underwritten. It has a measurable output, a market rate for that output, a depreciation schedule, and — crucially — customer contracts that can be pledged. That is the recipe institutional capital recognises, and it is why so much AI infrastructure finance is structured around hardware and its contracts rather than the real estate, a pattern we traced in turning compute into an asset class.

There is a timing problem sitting inside this, and it is the most important thing on the page. Six years straight-line has become the common accounting life for these parts, extended in recent years from four or five. Customer contracts, meanwhile, typically run two to five. The schedule outlives the contract, which means the final stretch of a GPU's book life is usually uncontracted.

Whether those late years earn is genuinely open, and the evidence so far is encouraging. NVIDIA points to the A100, introduced in 2020 and still in active commercial use six years later for training, fine-tuning, inference and HPC — with customers still committing multi-year capacity against it, which the company argues extends its economic life toward a decade. Older silicon has kept re-contracting rather than going idle. The honest counter is that this has not yet been tested through a full demand cycle. Anyone underwriting a GPU is underwriting a view on that tail, whether they have stated it or not.

There is only one product, but there are four ways to sell it, and they differ in exactly one respect: who carries the risk that the hour goes unsold. Every pricing decision in this industry is a trade along that axis.

The commercially decisive one is the bottom row. Multi-year take-or-pay capacity — paid for whether consumed or not — is the only version of this revenue that a lender will treat as bankable, because it converts an uncertain stream of hours into a contracted cash flow. Operators borrow against those contracts to buy the next tranche of hardware. Contracted backlog, not spot demand, is what funds growth.

Here is the cleanest thing in this whole business, and it falls straight out of the multiplication. When an operator is offered a committed contract at a discount to list, the question “should I take it?” has an exact form: a reserved discount is break-even against the utilization it implies.

Accept 45% off and you are earning 55% of the list rate — guaranteed, on every hour. Selling on-demand instead beats that only if you can keep the rack above 55% utilization. Take 35% off and the bar rises to 65%. Take 60% off and it falls to 40%. The discount and the required utilization are the same number, viewed from opposite sides.

That is why committed contracts at what look like punishing discounts get signed cheerfully. Set the implied bar against a market break-even in the region of 70% and a large discount stops looking like a concession and starts looking like insurance bought at a sensible price — all the more so because the contracted version is the one that can be financed.

Everything above describes a machine that produces measurable revenue against contracts. That is also, precisely, the description of something a lender can underwrite — and in August 2026 that possibility stopped being theoretical.

On August 10, NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish compute financing platforms intended to mobilize more than $500 billion of third-party capital for AI infrastructure over time. Jensen Huang described AI compute as an emerging investable asset class. We looked at the architecture of that initiative in detail in turning compute into an asset class; what matters here is what it does to the ladder.

Worth being exact about the number, because it is widely misread. $500 billion has not been raised. It is not a single fund, and it is not NVIDIA revenue. It is the capital the proposed platforms hope to mobilize over time, under memorandums of understanding still subject to final agreements. The accurate reading is a stated ambition backed by six of the most credible allocators in the world — which is genuinely significant, and different from a cheque.

The reason it matters is access. An advanced AI facility costs billions, and until now the ability to build one was effectively limited to companies that could fund it from their own balance sheet. A working financing market changes that: smaller clouds, labs and enterprises can build against the revenue the equipment will earn rather than the cash they already hold — the same shift that let airlines fly planes they could never have bought outright.

The mechanics resemble how aircraft, ships and power plants have been financed for decades. The asset earns income and simultaneously secures the loan, backed by the equipment itself, the customer contracts, the rental revenue and the expected residual value. NVIDIA adds one thing the aircraft market does not have: limited residual-value support for up to 25% of an opportunity, assessed project by project, and expressly designed to complement rather than replace independent underwriting.

Multiply that ceiling by the headline and you get $125 billion — a figure Huang has himself described as an option to backstop up to. Read it as a ceiling on optional, negotiated support, not a booked liability. The capital would be deployed gradually across many projects, the support is discretionary and per-deal, and final structures have not been disclosed.

Securitisation has an unfortunate reputation, earned in one market and unfairly generalised to every other. Aircraft leases, auto loans and commercial mortgages are financed and packaged routinely and unremarkably. The label is not the risk. The assumptions are. And for compute, the assumptions are unusually inspectable — which is a point in this market's favour, not against it.

Every question that matters here is answerable with evidence rather than faith, because the asset publishes its own rate card and its utilization is measured continuously:

The last one deserves particular attention, and the structure anticipates it: independent underwriting is the stated control, and vendor support is explicitly designed to complement it rather than substitute for it. That is the right design. A financing market is healthy when the money asks harder questions than the vendor does — and six allocators of this calibre are not short of reasons to ask them.

The honest summary is that this is a young market being built with mature instruments. Hardware depreciation remains the genuine open variable, and analysts are right to keep pressing on it. But the direction of travel is toward more disclosure, more standardisation and more independent scrutiny, which is how an asset class matures rather than how a bubble inflates.

Everything above is revenue. It is deliberately the top line and nothing else, because the top line is the part that is publicly knowable — published rates, published rack configurations, arithmetic. Below it sits the part that is specific to every operator and every site: power price, financing cost and structure, staffing, software margin, contract mix, and the depreciation view discussed above.

We are not going to publish those here, and readers should be wary of anyone who does with confidence. Market analysis puts gross margins in the region of 55-65% before depreciation, which tells you the shape of the thing — enough to know that idle capacity is expensive, and that a rack is a serious piece of revenue equipment — without pretending to a precision the inputs do not support.

What the ladder does give you is the right instinct. When someone quotes you a number in this market, find out which rung it is on — chip, node or rack, list or contracted, gross or net of utilization. Most of the confusion in AI infrastructure discussion is people comparing rungs. Most of the clarity comes from putting them back in order.