AI accelerators are quietly changing accounting categories. A recent analysis from Global Data Center Hub examines three deals and argues that the GPU has stopped being a component and started being infrastructure — something financed, depreciated and securitized like a turbine or a transformer.

The shift is more consequential than it sounds. When a chip is a component, it sits on a server bill of materials and follows the capex cycle of whoever buys the server. When it is an infrastructure asset, it enters the world of project finance, leasing structures and long-term offtake agreements. The three transactions the outlet reviews show that transition in progress.

What changes when a chip becomes an asset?

Finance changes first. Assets that can be valued, collateralized and depreciated on a schedule attract a different class of capital than components do. Infrastructure investors — the funds that own fiber routes, power plants and tower portfolios — have spent two decades learning to underwrite long-lived hardware with predictable cash flows. AI silicon, packaged at data center scale, now resembles those familiar instruments.

The comparison is instructive: a rack of accelerators under a multi-year service contract starts to look less like IT equipment and more like a leased power turbine. The revenue is contracted, the utilization is measurable, and the residual value is a debate rather than a mystery. That is precisely the shape infrastructure capital likes.

The second change concerns who bears the risk. Component purchases concentrate demand risk on the operator. Asset structures spread it across lenders, lessors and offtakers, each with its own view of how long the hardware stays productive. Global Data Center Hub's framing suggests the market is now building that distribution machinery around AI compute.

Why do three deals matter?

Three is a small sample. It is also how asset classes begin. Early project-finance structures in renewables and fiber looked equally provisional until the paperwork standardized and the capital followed. The deals in the analysis function as templates: each one demonstrates a workable structure for turning compute capacity into a financeable object.

The commercial stakes are considerable. If AI silicon can be financed like infrastructure, deployment no longer depends on the balance-sheet appetite of a handful of hyperscalers. A wider pool of capital can fund capacity, which in turn shapes ordering patterns across the supply chain — from wafer allocation at the foundry to advanced packaging slots to power procurement at the site.

It also reshapes the pricing conversation. Infrastructure assets are judged on yield rather than on performance per watt. A finance-driven buyer cares about revenue per dollar deployed over the asset's service life, not quarterly benchmark positioning. That is a subtle but real change in what the market rewards.

What could still go wrong?

Asset treatment cuts both ways. Infrastructure underwriting assumes a useful life, and accelerators remain subject to an unusually fast product cadence. Lenders that treat a five-year depreciation schedule as settled fact may discover that silicon does not age like concrete. The residual-value question — what a three-year-old accelerator earns in year six — remains the least tested assumption in the stack.

Geopolitics adds a second complication. Export controls and supply concentration already constrain where advanced AI silicon can be deployed, and an asset class needs predictable rules to attract patient capital. Regulatory volatility is, in effect, a pricing input that infrastructure models have not historically had to carry for hardware.

Where this leads

The direction of travel is clear enough. Global Data Center Hub's account indicates that the financial plumbing for AI compute is being assembled now, deal by deal, and that the winners will be those who standardize the structures first. If the template holds, expect AI silicon to be priced, financed and argued over the way power capacity is — a commodity with a queue.