From Cost Centre to Revenue Line: How Enterprises Are Rethinking Owned GPU Infrastructure

◆ THE 30-POINT SWING

Where the loss becomes a gain

A debt-financed GPU cluster breaks even at roughly 70% utilization. Below that line, every idle GPU-hour is depreciating and accruing interest without generating anything against either cost. Above it, the same hardware, financed the same way, running the same power draw, starts producing real margin. Gross margins on bare-metal GPU capacity typically run 55 to 65% before depreciation is even factored in, which leaves almost no room to absorb idle time once the debt service and operating costs are counted.

That's the part standard cost-centre accounting doesn't show. A GPU cluster sitting at 5% utilization isn't just "underused." It's actively bleeding against the breakeven point every single hour, in a way a simple capex-and-depreciation line item never surfaces to whoever's actually reviewing the budget.

◆ NEO-REAL ESTATE, NOT NEO-CLOUD

5% average utilization across 23,000 clusters

Cast AI's 2026 measurement across 23,000 production clusters found average enterprise GPU utilization at 5%. Cast AI's co-founder named what many fleets have become. Not cloud infrastructure. Neo-real estate. A separate industry survey found two-thirds of enterprises report peak GPU utilization below 70%, meaning even at the single busiest moment of the day, most fleets never even touch the point where a debt-financed cluster breaks even.

◆ THE FIRST STEP: MAKING USAGE VISIBLE INTERNALLY

Chargeback and showback, before any external monetization

Before an enterprise even considers offering idle capacity externally, the more common first step is making internal usage visible. Assigning AI infrastructure costs to the specific business units consuming it creates accountability that a shared infrastructure pool rarely has on its own. Even showback, giving teams visibility into what they're consuming without actually billing them for it internally, changes behavior simply by making consumption visible to the people using it.

Hard allocation limits per team, combined with queue-based scheduling for non-urgent workloads, prevent any single team or model from monopolizing shared infrastructure indefinitely. This isn't about slowing anyone down. It removes the incentive teams currently have to over-provision as a hedge against not getting capacity later, which is exactly the dynamic that produces a fleet running at 5% utilization in the first place.

◆ A LEASABLE ASSET, NOT A SUNK COST

What external monetization actually looks like

Once internal allocation is under control, the enterprises going further aren't buying fewer GPUs. They're treating owned hardware as a leasable asset rather than a fixed cost. A rack not fully booked for internal work during a specific window gets offered to a vetted external buyer for that window specifically, not indefinitely. Capacity returns to internal use the moment it's actually needed again.

This is closer to how a landlord thinks about an office building's unused floors than how a traditional IT department thinks about a server room. The asset stays owned. The utilization schedule around it becomes something actively managed rather than something that just happens.

◆ THE UTILIZATION SWING ON A REAL CLUSTER

Source: American Compute 2026, cited in ModulEdge's Neocloud unit economics analysis

GPU rental rates on the exact hardware most enterprises already own run a few dollars an hour to well over ten, depending on chip generation and demand. Multiply that by the hours a rack sits idle in a typical month. The uncollected revenue on a rack at 5% utilization is not a rounding error, and once a cluster's actual breakeven point is known, the size of the gap becomes a specific number rather than a vague sense that something is being wasted.

The same scarcity pressure that drove overbuying is why nobody releases capacity back, even capacity sitting unused. Procurement and runtime get managed as two separate problems, usually by two different teams reporting up two different chains. They're one problem from two ends. A procurement decision made under scarcity pressure creates the idle capacity that shows up on a utilization dashboard months later, and the team looking at that dashboard rarely has any input into the original purchase decision.

A rack stays available for internal use the moment it's needed. Still owned. Still on the balance sheet. What changes is whether the idle hours sit unused or get offered against real external demand, and whether the utilization number itself becomes something finance and infrastructure teams review together instead of something only engineering ever looks at.

◆ FAQ

Frequently asked questions