← Back to feed

The global AI data center buildout is routinely quoted in the tens of trillions of dollars over a decade. What gets less attention is the financial architecture holding it together — specifically, who is actually writing the checks and what happens if the revenue projections miss.

The short answer: a growing share of the capital is coming not from equity raises but from structured debt instruments that Wall Street and sovereign wealth funds are absorbing at scale.

Nebius: $4.5B in Convertible Notes

The clearest recent example is Nebius. The AI cloud infrastructure company — spun out of the former Yandex international business — announced a $4.5 billion convertible debt offering in August 2026, with proceeds earmarked partly for GPU purchases and data center expansion.

A convertible note is a bond that can flip into equity at a set price. For the issuer, it’s cheaper than pure equity dilution at current valuations. For the buyer, it’s a bet on continued AI infrastructure demand with downside protection: if the equity story collapses, they hold debt ahead of shareholders.

The Nebius deal is large but not exceptional. Similar structures have been used by CoreWeave, Lambda Labs, and several other hyperscale-adjacent players trying to accumulate GPU fleets without giving up control at depressed valuations.

Nvidia’s Role in Packaging GPU Assets

Nvidia is not a passive vendor in this arrangement. The company has been working with major financial institutions on frameworks to treat AI infrastructure as a structured investable asset class — closer to real estate investment trusts or aircraft leasing than conventional corporate debt.

The proposed structures involve private credit backed by GPU assets and future customer revenue commitments. The logic is straightforward: a Blackwell Ultra cluster has a reasonably predictable utilization curve given the demand backlog, a secondary resale market, and a calculable depreciation schedule. Those properties allow it to serve as collateral in a way that software assets cannot.

For Nvidia, this matters because it expands the buyer universe. If a midsize AI cloud company can raise GPU acquisition capital from private credit markets rather than equity, it can scale faster without waiting for a public offering or a strategic investor.

The Unit Economics Argument

Gavin Baker, founder of Atreides Management, has made the bull case for this financing structure explicit. His firm has seen a 100-fold increase in token consumption over the tracking period. Using that trajectory, Baker estimates that a $40 to $50 billion, one-gigawatt GPU data center could recover its upfront capital cost in nine to ten months, given customer prepayments and access to low-cost debt financing.

If the nine-to-ten-month payback figure holds, it restructures the risk profile of AI infrastructure investment substantially. Infrastructure with sub-year payback on the right demand assumption looks more like a toll road than a speculative venture bet. That framing is exactly what structured credit investors need to get comfortable with GPU-backed paper.

The caveat buried in every version of this argument: the payback depends on sustained demand, sustained pricing power, and access to power and chips that Baker himself flags as persistent supply constraints. If any one of those legs gives, the recovery timeline stretches.

Compute Inequality as Systemic Risk

Baker’s concern is not overbuilding — it’s the opposite. Persistent supply constraints in power, semiconductor capacity, and construction could produce what he calls compute inequality: a world where frontier AI access concentrates among a small number of well-capitalized incumbents who locked in capacity early, while everyone else queues.

That’s a structurally different risk than the data center glut narrative that circulated in 2023 and 2024. The argument then was that hyperscalers would overbuild and write down assets. The argument now is that even at current build rates, the demand trajectory makes underbuilding the more plausible failure mode.

The financing structures being assembled — convertible notes, GPU-backed private credit, customer prepayments as collateral — are optimized for the underbuilding scenario. They move capital into GPU capacity fast, before the supply constraints bite.

Who Holds the Tail Risk

The risk concentration question is real. Convertible notes convert to equity if things go well; they remain as debt claims if things go badly. Private credit backed by GPU assets is only as solid as the GPU secondary market, which has not been tested through a demand downturn.

If a significant portion of the $3.6 trillion projection ends up financed through structured instruments rather than equity, the distribution of losses in a downturn shifts toward credit holders — pension funds, insurance companies, and private credit vehicles — rather than equity investors who price in that volatility explicitly.

That is not necessarily a problem. Infrastructure financing through debt is normal. But the novelty of GPU-backed credit, the recency of the asset class, and the optimistic demand assumptions baked into the structures warrant scrutiny that the headlines about trillion-dollar buildouts tend not to provide.