Datacentre leases run ten to fifteen years. The customer contracts that fill them run four to five. Someone has to hold the difference. That someone is quietly becoming the most important institution in AI infrastructure — and almost nobody is underwriting it as such.

Australian Gulf Capital · Investment Office · Abu Dhabi

There is a structural flaw running through the centre of the AI infrastructure boom, and it is not a technology problem. It is a balance-sheet problem. We have spent the better part of two years underwriting the compute supply chain — at the equity layer through our venture book, at the asset layer through Aurecore, and at the credit layer through our structured-credit programme — and the same fracture appears at every level. We call it the tenor gap.

The mechanics are simple enough to state in a sentence. A hyperscale or neocloud datacentre is financed and leased on a horizon of ten to fifteen years, because that is how long the physical asset and its power contracts last. But the customers who rent that capacity — the model labs, the inference platforms, the enterprises — sign for two to five years, because no AI company will commit a decade of demand to hardware that may be two generations obsolete by then. The asset is long. The revenue is short. The difference does not disappear. It sits on someone's balance sheet as duration risk.

This is the oldest structure in finance. An institution that funds long-dated assets with short-dated liabilities — or the reverse — is performing maturity transformation. That is the definition of banking. The neoclouds have, without quite intending to, become unregulated banks for compute — and the market is still valuing most of them as if they were software companies or, at best, real-estate trusts.

An institution that holds a long asset against a short contract is not a landlord. It is a bank. The AI infrastructure cycle has produced a new generation of them, and they do not yet know it.

Why the gap is widening, not closing

The intuitive response is that the gap should close over time as the market matures — that customers will eventually sign longer contracts, or that hardware cycles will lengthen. We think the opposite is true, for three reasons.

First, hardware cadence is accelerating, not slowing. The move from one accelerator generation to the next is now an annual rhythm rather than a multi-year one. Every compression of the hardware cycle shortens the rational customer commitment and lengthens the relative duration of the asset. The gap is a function of the ratio between asset life and contract life, and the denominator is shrinking.

Second, the capital required is migrating from equity to debt. Early datacentre build-out was equity-funded because it had to be — there was no collateral history to lend against. As the asset class matures, the financing stack is rotating toward debt, which is cheaper but far more sensitive to duration mismatch. A mispriced tenor gap that an equity investor simply rides out becomes, for a lender, a covenant breach.

Third, power is the binding constraint, and power contracts are long. The scarce input is no longer chips — it is contracted megawatts. Securing power means signing decade-plus offtake. That deepens the asset side of the mismatch precisely as the revenue side gets shorter. The firms that control contracted power are taking on the longest duration in the entire stack.

What the market is mispricing

If neoclouds are banks, then the correct lens is not revenue multiples — it is the quality of the maturity transformation. Two neoclouds with identical revenue can have wildly different risk profiles depending on how they finance the gap, who holds it, and how the contracts are structured. The market is not yet pricing this distinction. We think that is the single largest source of mispriced risk — and mispriced opportunity — in AI infrastructure today.

The Three Questions We Underwrite

1. Who holds the duration? If the operator holds it on-balance-sheet, it is a bank and should be underwritten as one. If it has been transferred to a financing counterparty, the question is whether that counterparty has priced it correctly.

2. What backs the short contract? A two-year contract from a frontier lab with billions in funding is a different instrument from a two-year contract from a startup burning capital. The credit quality of the tenant is the credit quality of the asset.

3. Is the power contracted? Uncontracted power is the most dangerous position in the stack — long asset, short revenue, and no certainty on the most expensive input. Contracted power converts the gap from speculation into an underwritable spread.

The institution that holds the gap

Here is the conclusion that organises everything we are building. In every mature financial system, the institution that holds the maturity transformation — that funds the long asset and manages the short liability — captures the largest and most durable share of the economics. Not the borrower. Not the depositor. The bank in the middle.

The AI infrastructure cycle is producing the same structure, and the role is open. It will be filled either by the operators themselves — taking on bank-like risk they are not yet measuring — or by purpose-built financing institutions that underwrite the gap deliberately, price it correctly, and hold it against the right collateral and the right power contracts.

This is the thesis behind AGC's structured-credit programme. We provide senior-secured financing against GPU fleets, contracted datacentre capacity, and long-duration offtake — the precise instruments that sit inside the tenor gap. And it is the thesis behind Aurecore: by controlling the asset and the power at the IRIS Campus, we hold the long side of the gap on our own terms rather than financing someone else's mismatch blind.

We do not finance the tenor gap because it is safe. We finance it because it is mispriced — and because we underwrite it at the equity, asset, and credit layers at once, which almost no one else can.

The firms that understand this will not describe themselves as banks. The most valuable institution of the AI infrastructure era will look, from the outside, like a datacentre company or a credit fund. Underneath, it will be doing what banks have always done: holding the duration that no one else wants, and being paid for it.