The Numbers That Look Too Good

The neocloud sector generated more than $25 billion in revenue in 2025. Fourth-quarter revenues surged 223 percent year-over-year to $9 billion alone. By 2031, analysts project the market will grow to nearly $400 billion at a compound annual growth rate of 58 percent. These numbers move through investor decks with the ease of settled facts. However, beneath the headline figures sits a financing architecture that deserves considerably more scrutiny than the market currently applies. Neoclouds operate on a deceptively simple playbook. They sign multi-year offtake agreements with AI labs and startups, borrow heavily against those contracts, then use the proceeds to finance the next stage of GPU cluster expansion. The contracts become the asset. The buildings, the power, and the liquid cooling infrastructure exist primarily to service what those contracts represent on paper. Consequently, the valuation of the facility depends almost entirely on the creditworthiness of whoever signed the lease — and in the neocloud world, that counterparty is frequently an early-stage AI startup.

Paper Commitments, Real Concrete

This is where the distortion begins. Private equity valuations and debt underwriting models treat backlog figures as confirmation of real demand. CoreWeave disclosed a remaining performance obligation of $99.4 billion as of March 2026. Nebius reported a backlog of $46 billion, including a major Microsoft commitment. Those numbers are extraordinary by any infrastructure standard. Moreover, they have attracted the capital needed to build facilities at extraordinary speed. However, the financing structure contains a hidden asymmetry. Lenders underwrite neocloud debt on what practitioners call look-through credit — they evaluate the end buyer’s financial strength, not the neocloud operator’s own balance sheet.

A large commitment from a Fortune 500 counterparty clears credit committee easily. In contrast, a commitment from a Series B AI startup creates an entirely different risk profile. As of mid-2026, most Series B-plus AI startups are seeking compute allocations of two megawatts or more. Many of those commitments back debt that finances custom liquid-cooled facilities designed specifically for high-density AI workloads. Therefore, the moment the AI startup’s own funding position shifts, the anchor holding the facility’s valuation in place shifts with it.

The Meta Signal Nobody Should Ignore

The fragility of customer concentration crystallized in a single day. On July 1, 2026, Bloomberg reported that Meta was exploring its own cloud business under the name Meta Compute, potentially selling off excess AI capacity to external customers. Shares of Nebius and CoreWeave temporarily plunged around 15 percent on the report. One company exploring a tangential product line moved two neocloud valuations by double digits in a single trading session.

That reaction is instructive. It reveals how thinly distributed the real demand base currently is beneath the headline backlog numbers. Meanwhile, Microsoft already represented 62 percent of CoreWeave’s total revenue in 2024. Meta signed deals worth $35.2 billion with CoreWeave and up to $27 billion with Nebius. Furthermore, the hyperscalers themselves increasingly sign neocloud commitments as operating expense structures rather than capex — meaning the same companies building neocloud competitors are simultaneously their largest customers. The circular exposure this creates is not a theoretical risk. It showed up in real-time valuations on July 1.

When the Tenant Becomes the Competitor

The neocloud market has developed a structural contradiction that the valuation models have not fully priced. The largest and most creditworthy customers are also the most likely to build competing capacity. SpaceX already demonstrated this pattern clearly. Following its acquisition of xAI in February 2026, SpaceX began marketing its own Colossus data center capacity externally, competing directly for the same enterprise AI customers its infrastructure had previously served. The biggest customer became the competitor almost overnight.

This dynamic is not limited to SpaceX. Hyperscalers are simultaneously signing neocloud contracts as customers, building their own competing capacity, and developing proprietary silicon specifically designed to reduce GPU dependency. AWS Trainium, Google TPU, and Microsoft Maia each represent a long-term substitution threat to the very infrastructure neoclouds are financing today. Therefore, the backlog figures that underpin neocloud valuations today assume continued customer commitment from counterparties who are actively reducing their dependency on the product those backlogs represent.

The Economics Don’t Forgive Utilization Drops

The financial structure underneath these valuations offers almost no margin of safety. Gross margins for GPU rental businesses sit between 55 and 65 percent before depreciation. After labor, power costs, and depreciation, that margin compresses to roughly 14 to 16 percent — lower than many non-technology retail businesses. If utilization slips below 80 percent, returns flatline entirely. Moreover, depreciation on GPU fleets consumed more than half of revenue at both CoreWeave and Nebius in recent reporting periods.

The chip release cycle compounds this pressure further still. Over a typical five-year depreciation horizon, the price of a GPU hour could decline by half or more. That dynamic requires every neocloud to recover capital within the first four to five years after a GPU becomes active, then reinvest in the next generation before the existing fleet loses commercial relevance. Custom liquid-cooled facilities built for specific GPU architectures introduce an additional layer of risk here. Unlike air-cooled generic shells, which can be repurposed for different tenant types, a facility designed around 120-kilowatt rack densities for H100 clusters is not easily redeployed for enterprise workloads at 8 to 15 kilowatts. The infrastructure becomes asset-specific at precisely the moment asset-specificity becomes a liability.

What Stranded Assets Look Like in Practice

Two brand-new Silicon Valley data centers sat idle in 2025 waiting for electrical equipment on back-order. They had signed tenants, committed capital, and a liquid cooling infrastructure specification that assumed a specific workload profile. However, without power, the facility produced no revenue while debt service ran continuously against it. This is the physical expression of what happens when infrastructure build timelines and utility synchronization timelines diverge.

The stranded asset risk in the neocloud context is structurally similar but operationally more severe. A generic data center shell that loses its primary tenant can be remarketed to alternative users. A facility specifically engineered for maximum GPU density, built with direct-to-chip liquid cooling manifolds, custom power distribution, and a thermal architecture optimized for 100-kilowatt-plus racks, has a dramatically narrower re-leasing market. Consequently, when the underwriting thesis depended on a single AI startup that subsequently pivots, runs out of funding, or consolidates onto hyperscaler infrastructure, the developer is left holding a highly specialized asset in a market that values generalist flexibility.

The Discipline the Supercycle Is Missing

None of this argues that neocloud infrastructure is without genuine value. Compute demand brokers have described demand as outpacing supply at a 50-to-1 ratio in 2026. The structural shortage of AI compute is real and the operators building into it are responding to actual market signals. However, the discipline those signals demand is not being applied uniformly across the capital stack. Utility-synchronized capital allocation — sizing facilities to confirmed grid capacity, structuring debt against verified rather than projected utilization, and applying the same credit scrutiny to AI startup counterparties that would apply to any sub-investment-grade industrial tenant — would not slow the market. Instead, it would make the valuations the market assigns to these assets defensible over the asset’s actual operational life.

Sub-investment-grade neocloud tenants introduce a category of risk beyond traditional hyperscalers that the current underwriting environment is only beginning to acknowledge through parent guarantees and hyperscaler credit wrappers. Building that discipline into the capital allocation process from the design stage is fundamentally different from retrofitting accountability onto a stranded liquid-cooled hall after the tenant has moved on. The supercycle story is real. The phantom demand layer growing inside it deserves equal attention.