Research note · method

Utilization is not one number.

AI capacity can be fully booked while the equipment is underused. Booked space, allocated chips, active workloads and financial returns measure different things.

Published 5 September 2026 Draft for review
The measurement problem

Five measurements, one word.

North American data-centre markets are commercially tight. CBRE reports primary-market supply up 33.7% year over year in H1 2026 while vacancy fell to 1.4%, with 80.4% of capacity under construction preleased. JLL reports 1% vacancy and 95% pre-commitment across a broader universe.

At the same time, Cast AI reports average GPU utilization of 5% across the tens of thousands of Kubernetes clusters it sampled through April 2026. And US grid planners are confronting more than 700 GW of data-centre power requests, some of which regulators now treat as speculative.

These numbers do not describe the same thing. They measure different stages between an intention to build and economically productive compute.

The conclusion

The current utilization threshold should stay unresolved. Future assessments should separate six measures: requests, bookings, delivered capacity, allocated equipment, active workloads and financial returns.

Each stage would need its own supporting evidence.

From plans to returns

Six stages between a request and a return.

Demand does not become useful capacity in one step. Infrastructure investments usually fail when plans, bookings or delivered equipment do not make it to the next stage.

01
Planning demandHow much capacity is being requested?
Utility interconnection requests, announced campuses, prospective megawatts. Some requests reserve options for projects that may never be financed or built.
Medium, noisy
02
Contracted capacityHow much future capacity has a financially identified customer committed to take?
Preleased megawatts, signed cloud capacity agreements, take-or-pay contracts. Stronger evidence of demand, but not evidence that infrastructure exists.
Medium-high
03
Delivered capacityHow much contracted capacity has been commissioned and can carry load?
Energized megawatts, delivered halls, functioning cooling and network, live grid connection. This separates construction ambition from physical delivery.
Low-medium
04
Compute occupancyHow much available accelerator capacity is provisioned to customers or workloads?
Allocated accelerator-hours over available accelerator-hours. A commissioned facility can still contain unallocated compute.
Low
05
Active workloadHow intensively is provisioned compute actually being used?
Accelerator busy-time, memory and computation occupancy, serving and training duty cycle. This is the layer the Kubernetes telemetry measures for its own sample.
Low outside selected fleets
06
Financial returnIs the active infrastructure earning enough to justify its installed cost and cost of capital?
Revenue or gross profit per accelerator-hour or megawatt, contract yield against annualized installed cost, return on invested capital against cost of capital. Equipment can be busy without earning enough to cover its cost.
Low-medium, company-specific

Public-data quality is assessed as of September 2026 and is expected to improve at stages 3 and 4 before stages 5 and 6. For each layer FP1 publishes the numerator, denominator, geography or fleet scope, observation date, primary source, source independence, known exclusions and confidence.

Measurement rule

Keep the six measures separate.

The six layers are not summed into a single utilization score.
  • Wait until at least three successive assessments show stable, comparable public data.
  • Combining them would suggest more certainty than the evidence supports.
  • Using CBRE vacancy would redefine utilization as leased physical capacity. Using grid requests would redefine it as planning demand. Using a Kubernetes sample would redefine it as workload activity in one fleet. None is equivalent to the R-012 concept.
  • We cannot track change by switching to whichever measure happens to be available in September.

When this framework would no longer be needed. The Utilization Stack should be abandoned or collapsed if a reliable public series emerges that directly and consistently measures delivered accelerator capacity, active workload and economic yield across the relevant hyperscaler and neocloud universe with a stable denominator.

Consequence for the Radar: no new crossing is recorded for the existing utilization dimension until we have a clear, comparable measure. The old reading remains part of the record; the revised method applies to future assessments. See the September evidence update.