Parceling the Risk Stack

Why equity cannot carry the buildout

Marc Rowan frames the scale of AI-related capital spending, data centers, chips, robotics, manufacturing, and defense combined, as roughly every dollar spent since the invention of fire, and argues it cannot be financed with equity for two separate reasons. Equity is the most expensive capital in any structure, so using it to fund a building or a rack of hardware with real resale value destroys returns unnecessarily. And there is simply not enough venture and growth equity in existence to fund a buildout at this scale at any price.¹ His answer is to decompose a single project by risk type rather than fund it as one instrument: the fundamental business risk, whether the company itself will work, stays with equity, while the reusable parts with real hard-asset value get offloaded into credit markets at whatever rate and rating that collateral earns on its own.

The 2026 call

Rowan treats 2025 as proof of concept, the year the market confirmed data centers, chips, and energy were all genuinely needed, and identifies 2026 as the year the market starts recognizing a structural limit: with roughly eight hundred billion dollars of capital expenditure from just four public companies in a single year, before counting private spending, investors are approaching concentration limits in a narrow set of names, and he expects credit spreads on AI-adjacent debt to widen as a result.

Why robotics is financed like equipment, not like venture

The same logic extends to autonomous physical equipment. Rowan's reasoning is that once the hardest version of the self-driving problem, a constantly changing environment where safety is paramount, gets solved, applying the same autonomy to construction equipment should not be harder. The financing insight is the actually novel part: there is already a large, mature equipment rental market with a far lower cost of capital and far greater scale than venture capital, and there is no reason autonomous construction equipment should be financed any differently from ordinary construction equipment. Autonomy, on this view, is simply a software layer sitting on top of a depreciating hard asset that lenders already know how to price.

Whose credit is actually behind the paper

A separate 2026 interview supplies the underwriting test that determines which side of the split a given asset lands on. Rowan distinguishes strong credits, where a diversified investment-grade company such as Amazon, Google, or Microsoft is willing to lend its own credit to a project specifically so the debt does not consolidate onto its own balance sheet, from weaker ones that are more equity story than sensible underwrite. Good underwriters pick good credits, he says, and bad underwriters pick everyone.² Much of the tranching this framework describes turns out to be a hyperscaler's off-balance-sheet structure in practice: the equity story and operating risk stay with the AI company itself, while the creditworthiness is supplied by a diversified counterparty that wants the capacity built without owning the debt, which means the real credit question is frequently about the strength of a large tech company's guarantee rather than about the data center it is attached to.