Opinion | Finance & Investment
A model built for speed
Vendor funding, special-purpose vehicles and GPU-backed debt keep AI capacity moving – but they also bundle assets with very different lifespans and risk profiles into structures that few outsiders can read, making the financing behind AI infrastructure harder to follow than the infrastructure itself: A supplier backs a customer, the customer buys the supplier’s compute, an operator builds around that hardware, and a tenant services the debt.
The scale is substantial
An estimated $466bn of AI-linked debt has been raised so far in 2026, against $101bn in the same period last year and much of this sits in special-purpose vehicles that hold the chips and the customer contracts separately from the operator. Earlier this year an $8.5bn facility became the first investment grade rated GPU-backed financing and more than $20bn of such facilities have now been announced.
None of this is improper; Vendor financing is a long-established commercial practice and ring-fenced vehicles can allocate risk sensibly – the concern is structural: what happens when several layers of risk are wrapped into one arrangement?
When suppliers, customers and lenders are intertwined, it becomes difficult to separate genuine end-user demand from demand supported by financing. The IMF flagged this in April when it noted increased circular financing across the value chain, though it judged the financial stability impact minor at this stage, “minor for now” is not the same as “understood.” If investors cannot see how much revenue depends on a counterparty who is also a funder, how can they price the risk properly?
Assets with mismatched lifespans
The physical data centre is long life infrastructure – the chips inside are not. Useful life is typically assumed at four to six years for GPU equipment whilst lifespans for physical infrastructure are much, much longer.
Debt sized against hardware should be repaid while the collateral loses value, which makes depreciation assumptions the load-bearing wall of the model – and they are contested. The optimistic case has evidence: a three year old, high end GPU resells for around 60% of its original price – but residual values do not follow a simple depreciation curve and a price cut on new chips could see used values pushed far below projections – every new generation of hardware ages the one that came before it.
The real collateral is the counterparty
Lenders have learned this; One analysis argues the market is overwhelmingly pricing the credit quality of the tenant and the length of the contract, not the residual value of the chips – which is rational, but it does shift the risk. If the tenant is a well-capitalised, long-term customer, the structure is robust. If the tenant is an early stage business whose own funding depends on the same ecosystem, the loop closes – a missed payment, lost contract or a funding squeeze could potentially cascade through the vehicle, the operator and the lender all at once.
Delivery risk sits outside the model
Financing models assume a facility will be energised on schedule – grid constraints, permitting and equipment lead times make this an “assumption”, not a certainty. Some recent reporting showed tenants invoking contractual protections over power delays where the tenant could potentially secure a three-year delay on rent once payments begin. Whilst a rent deferral on one project is a cash-flow problem for its equity and a covenant question for its lenders – multiply this across a pipeline and it becomes a market-wide issue.
Specialised assets are also hard to redeploy: An AI facility built for high density, liquid cooled racks is a specialised environment; If a lender has to enforce or convert it for general use, this is a very complex and costly exercise – collateral that works well for one use case but works poorly for others is not only harder to value – it’s harder to sell.
What does better look like?
The AI data centre industry does’nt need less capital – it needs clearer separation as assets are often financed as one, when they are fundamentally different:
- A chip is a fast-evolving, high-depreciation technology asset.
- A data centre is a long-term real estate asset.
- Future AI demand is an unproven market assumption.
Should financing should reflect this ?
Should short life equipment be paid down on a schedule that matches it’s realistic life? Long life infrastructure carry long term capital with demand assumptions stress tested openly, and not embedded in the structure of the debt.
Does the market need disclosure?
Who are the ultimate counterparties, how much revenue depends on them, what are the depreciation assumptions, and who bears the loss if these assumptions prove to be wrong? Perhaps the real danger is not that AI infrastructure is over financed – maybe it is the shared assumption that someone else is carrying the downside.
When the funder, supplier, operator and the tenant are all closely linked together like this – does the question silently move from who is building the next generation of AI infrastructure, to who ultimately owns the risk when assumptions change.
