The Provider Shakeout: What Q1-Q2 Exits Taught the Market

◆ NOT BAD LUCK, NOT BAD OPERATORS

A model with almost no margin of safety

The Q1 and Q2 exits were not bad luck or bad operators. They were a business model with almost no margin of safety meeting three pressures at once.

The model itself is simple to state. Sign multi-year take-or-pay contracts for GPU capacity, borrow against those contracts, and race depreciation to the bank. It works when utilization holds and the debt clears before the hardware loses value. The whole art of the raise is arranging capital so the debt is repaid before the collateral decays and before the customer contract runs out. Get that right and a forty-person operator raises hundreds of millions on terms that used to be reserved for utilities. Get it wrong and the facility that funded the growth becomes the thing that ends the company.

◆ A 1,024-GPU H100 CLUSTER, BY UTILIZATION

Sources: American Compute 2026 via ModulEdge for the 55% and 85% figures, Spheron cost model dated 4 July 2026 for the 70% and 90% figures. The two models use different rate and colocation assumptions, which is why the bands differ

◆ UTILIZATION BREAKS FIRST

A fifteen-point swing decides solvency

A debt-financed cluster breaks even somewhere around 70 to 75% utilization. On a 1,024-GPU H100 cluster, 55% utilization produces a loss of roughly $330,000 a month while 85% produces a gain of roughly $340,000. Sit at 70%, right at the threshold, and the same cluster still loses $80,000 to $130,000 a month. Push to 90% and it generates $170,000 to $220,000. Bare-metal gross margins run 55 to 65% before depreciation is counted at all.

That is a business where a fifteen-point utilization swing decides solvency. When spot pricing falls below cost basis, the revenue side of that equation stops working regardless of how well the cluster is run.

◆ THREE FORCES CONVERGED, AND A QUIET FOURTH

All in the same two quarters

Capital raised in a different interest rate environment came due. Debt service on hardware depreciating faster than model generations is brutal arithmetic, and it does not improve with time. Over a hundred providers were competing for investment precisely when investor scrutiny of unit economics intensified, which meant the capital needed to refresh aging fleets became hardest to raise at the moment it was most needed.

Chip availability shifted the advantage. Blackwell production ramping through 2026 expanded supply, which sounds helpful and was not, for smaller players. Expanded availability makes volume purchasing power and supplier relationships more decisive, not less. A provider without scale loses the pricing advantage that scarcity had been providing.

Hyperscaler price cuts closed the gap that justified the category. Neoclouds existed partly because they were cheaper than AWS and Azure. When the hyperscalers cut, the differentiation narrowed for providers who had nothing else to offer.

First-generation infrastructure also hit depreciation limits at the same time, which is the quiet fourth factor. Fleets bought in the 2023 and 2024 build-out reached the point where the accounting and the competitive reality both said replace, and replacing requires capital that had become expensive.

◆ A DISCLOSURE BEFORE THE NEXT PART

Capital-heavy against asset-light

There is a structural distinction worth naming, and I should be upfront that GPUaaS sits on one side of it.

Capital-heavy providers raised billions, took on debt, and built or leased data center capacity to compete on infrastructure scale. CoreWeave, Lambda, and Nebius each made the same bet, that owning the hardware is the path to owning the market. Asset-light platforms aggregate existing capacity without carrying the balance sheet. Analysis from Axe Compute argues the asset-light model wins structurally, because it does not carry the debt, the hardware depreciation risk, or the customer concentration pressure. We are asset-light, so read that argument knowing we benefit from it being true.

What is not in dispute is that the two models fail differently. A capital-heavy provider under pressure has to keep utilization above a hard threshold or lose money every month. That constraint is real regardless of how the market eventually shakes out.

◆ THE SYMPTOMS ARRIVE BEFORE THE ANNOUNCEMENT

What provider stress looks like from the customer side

For buyers, provider financial stress shows up before any announcement does, and the symptoms are specific. Capacity rationing. Lower responsiveness for non-enterprise customers. Sudden pricing changes. Those are what a provider does when utilization is below break-even and cash is tight, and they are visible to a customer well before a shutdown notice.

There is a pricing tell worth understanding too. Providers financed through vendor arrangements carry locked-in cost structures tied to revenue-share terms, so their on-demand pricing has to cover both debt service and the vendor's cut. Providers without vendor-financed capital expenditure can price closer to market because they are not servicing those obligations. A price that seems inexplicably high may be structural rather than opportunistic.

The consequences of a mid-contract exit are worse than an inconvenience. Workload migration under time pressure, service disruption, renegotiated pricing from a weak position, and in the worst case stranded workloads with no clear recovery path.

◆ THE NEXT COHORT UNDER PRESSURE

The evaluation standard moved

The timing matters for anyone signing now. Commitments made in 2024 and 2025 come up for renewal through late 2026, and the evaluation standard has changed underneath them. The question in 2024 was whether GPUs could be accessed at all. The question now is data sovereignty, geographic resilience, and compliance architecture. Providers who could clear the first bar and cannot clear the second are the next cohort under pressure.

None of this argues against using a neocloud. It argues for asking about the balance sheet with the same seriousness as the fabric specification, because a provider's utilization threshold is a risk the customer carries whether or not anyone mentions it. For what to ask specifically, see the vendor evaluation checklist and the three questions most providers cannot answer.

◆ FAQ

Frequently asked questions