The Neocloud Problem

Where the AI Datacenter Value Chain Actually Breaks

Earlier this summer, our Artificial Intelligence research series examined the AI investment cycle from the perspective of public markets: capital expenditure, semiconductor leadership, monetization, portfolio concentration and the exposures sitting on the other side of the trade.

With this new research stream, we move one layer deeper into the infrastructure financing that makes the AI buildout possible.

The starting question was straightforward: what electricity price does an AI datacenter need in order to remunerate both its equity and debt investors?

The answer led somewhere else.

Our unit-level model suggests that electricity is not the binding constraint in AI datacenter economics. For a specialist operator, power represents less than 10% of the annual cost base. The more important variables are capital structure, hardware depreciation and the ability of revenue to keep pace with an asset whose economic value declines over time.

The distinction becomes clear when the same physical asset is viewed through three different balance sheets.

One gigawatt of AI computing capacity costs approximately $38 billion to build. A hyperscaler self-build requires roughly $7.94 billion of annual revenue to remunerate its capital. A specialist or neocloud requires approximately $10.36 billion. Observed specialist revenue is currently around $10.30 billion per gigawatt-year — effectively break-even under a four-year hardware-life assumption.

That leaves very little room for error.

A hyperscaler can absorb a material revenue decline before the economics of the asset break. A leveraged specialist cannot. The same demand disappointment can therefore remain an earnings event for one balance sheet and become a debt-service event for another.

The research also examines a less visible vulnerability: collateral.

More than $20 billion of lending is secured against computing hardware whose secondary-market value can decline rapidly. At observed depreciation rates, increasingly long financing structures can reach the point where collateral loses value as quickly as the loan amortises. The analysis therefore shifts the focus from electricity prices to a more fundamental question: how long does the hardware remain economically productive, and how safely has it been financed?

The study maps this risk across specialist operators, hyperscalers and regulated utilities, showing how the same underlying shock changes form as it travels through the value chain: credit risk at the specialist level, earnings and valuation risk for hyperscalers, and regulatory or allowed-return pressure for utilities.

The video above presents the principal findings.

It is based on a broader 36-page vertical research note, including the unit economics model, stress testing, collateral analysis, company-level exposures, assumptions and source work. For readers who wish to explore the underlying analysis in greater detail, the complete research note is available upon request.

This material is provided for informational and research purposes only and does not constitute investment advice, a recommendation, an offer or a solicitation.