GPU Rents Are Rising. AI Margins Need a Better Test.

AI infrastructure can be scarce while its owners earn disappointing returns. The investment question is who can turn a constrained resource into durable cash after operating costs, financing and replacement investment.

1. The spot market is a signal, not the entire business

The supply chain has several distinct owners of scarce resources: chip designers and manufacturers, owners of energized buildings, compute operators, and businesses selling model access. They do not share a single revenue model. One sells equipment, another rents time, and another monetizes an outcome. A constraint can support one layer while raising costs for another.

Figure 1 records five GPU benchmarks from the same settlement date. The dispersion matters: a newer accelerator can command a higher hourly rate while delivering a different quantity of useful work. Comparing those dollar rates as if they were identical services confuses the price of an input with the cost of production. [ornn]

Ornn’s methodology uses active on-demand transactions and a volume-weighted, winsorized calculation. Long-term and reserved contracts are outside that spot benchmark; contributors are not publicly identified. A change in the index should therefore prompt investigation of availability and contract renewals. It cannot establish the repricing of every installed machine. [ornn-method]

The timing of exposure is decisive. A business that owns usable capacity and reprices frequently may benefit promptly. An operator that already sold capacity under a long contract may see little immediate upside. A model provider with committed supply could be insulated from a temporary spot spike, while a competitor relying on short-term capacity could face a higher bill. Contract structure determines whose earnings respond and when.

2. Translate the GPU-hour into a comparable task

A GPU-hour is a time-based input. A token is an output unit whose economic meaning depends on the model, workload and serving configuration. Even “one million tokens” is an incomplete product definition: input and output tokens can consume different resources, and cached requests are not equivalent to uncached generation.

Compute cost per million billable tokens = rental price per GPU-hour × GPU-hours required per million billable tokens

The second term must be measured at the required service level. Increasing concurrency can improve aggregate throughput while degrading the experience of an individual user. A lower-quality model or looser response-time limit can make a benchmark cheaper without delivering an equivalent product. MLCommons explicitly defines inference benchmarks through models, datasets, quality targets and scenario-specific metrics. Its Closed and Open divisions also permit different degrees of model flexibility. [mlcommons]

For commercial research, the comparable unit may be a successfully completed coding task, a generated video meeting a quality threshold, or an accepted business workflow. Token economics remain useful, but they should sit underneath a task-level view whenever the customer buys a result. More tokens generated can represent more productive work, more reasoning effort, or simply more verbosity.

Figure 3 below makes this distinction explicit with a scenario rather than a forecast. With hourly rent up 20%, unchanged throughput increases unit compute cost by 20%. A 40% throughput improvement instead lowers that cost by 14.3%. Neither outcome can be inferred from the rent increase alone.

3. EBITDA does not settle the return on the asset

The operator’s problem extends beyond renting or owning an accelerator. If hardware is owned, cash was committed before revenue arrived. Financing, economic obsolescence and replacement capacity still matter when a quarterly operating metric excludes them. If hardware is rented, the contractual bill may be visible in operating expense instead. Two businesses can report different margin structures while bearing related economic risks.

CoreWeave reported Q2 2026 revenue of $2.575bn and adjusted EBITDA of $1.510bn. Its GAAP operating loss was $49m and net loss was $626m. The reconciliation below shows where the differences arise, without treating accounting depreciation as current cash spending. [coreweave]

This is not an argument that EBITDA is useless. It helps isolate an operating layer of the business. But the owner of a rapidly evolving asset base must eventually earn enough to service its obligations and sustain competitive capacity. A metric that removes depreciation cannot, on its own, answer whether the assets earn their replacement cost.

Depreciation is an estimate of expense allocation; replacement investment is an actual financing requirement that may occur on a different schedule. Maintenance capex and growth capex are also economically different, yet clean separation can be difficult when a new generation improves both capacity and efficiency. Analysts should test several useful-life and residual-value assumptions rather than substitute the accounting schedule for the investment thesis.

The hurdle becomes higher when selling prices fall. Under the chart’s assumptions, preserving the original compute-only revenue-to-cost ratio after a 10% selling-price decline requires throughput to rise 33.3%: 1.20 ÷ 0.90 − 1. Improving throughput by 20% merely holds compute unit cost flat; it does not offset the lower selling price. These are algebraic conditions, not predictions about any provider.

4. Read the contract before assigning the winner

The most attractive bottleneck owner is not necessarily the company with the highest headline rental price. Reliable availability, software integration and customer switching costs can support realized value, but their benefit has to exceed the cost of delivering them. Conversely, a cheaper service may have sound economics if it operates efficiently and fills otherwise idle capacity.

The counterargument to permanent scarcity deserves equal attention. New supply can arrive, workloads can migrate, and better software can reduce the resources needed for a comparable task. At the same time, lower task cost can stimulate additional demand. The balance cannot be settled by assuming either unlimited efficiency or unlimited consumption.

5. The evidence that would change our view

Monitor the chain in order: available capacity, paid utilization, realized pricing, task productivity, operating cash and sustaining investment. Each link should carry its own period, definition and source. A strong observation at one link should not fill an empty field elsewhere.

ConstraintAlpha view: Higher GPU rents are a reason to investigate. Sustainable value capture requires a fuller test of useful output, contractual risk allocation and lifetime asset returns. The scarce input matters; so does the cost of keeping it productive.