Scarcity is visible. Productive utilization is not.
New evidence shows demand for AI infrastructure remains strong and financing is getting harder. But we still lack comparable data on how much capacity is productively used. The June assessment stands; this update explains what has changed.
Why the June assessment still stands.
R-012, completed 29 June 2026, remains the latest full Hyperscaling assessment. This update adds evidence without changing its results or the number of thresholds crossed.
To track change, we need to measure the same thing each time. We could update four of the five measures today. The fifth, utilization, still lacks a clear definition of the capacity being measured. It was already the central uncertainty in R-012.
Updating the other four measures would leave that gap unresolved. Changing what we count as capacity could move the number because we changed the math, not because the world changed.
What changed
Available capacity is scarcer. Lenders are less forgiving. And the figures called “utilization” measure different things.
Six findings and what they mean.
Each finding links to a dated source. Figures covering different markets or fleets are kept separate.
Commercial capacity remains scarce
CBRE's H1 2026 figures show North American primary-market vacancy at 1.4% even after supply grew 33.7% year over year to 10,903 MW. Preleasing rose to 80.4% of capacity under construction, leaving under 1,500 MW available across eight primary markets.
CBRE, North America Data Center Trends H1 2026, 27 August 2026Available space is being taken up, even as more is built.
Forward commitments remain large
JLL reports North American vacancy at 1% for a third consecutive year and 95% of 66 GW under construction pre-committed. Its market universe is wider than CBRE's, so the construction totals should not be added together.
JLL, Data center demand exceeds expectations in H1 2026Customers are reserving future capacity. A booking does not tell us whether the equipment is installed or doing useful work.
Hardware demand remains strong
NVIDIA reported Q2 FY2027 revenue of $96.2B on 26 August, with Data Center revenue of $89.0B, up 117% year over year, and guided Q3 to $108B excluding any China Data Center compute.
NVIDIA Q2 FY2027 results, 26 August 2026These results do not suggest a broad collapse in accelerator demand.
Credit is available, but tolerance has narrowed
Apollo reports hyperscaler bond cover ratios falling from nearly 5x in February to below 2x in July. Reuters reported roughly $194B of hyperscaler issuance through 7 July alongside wider concessions and weaker secondary performance.
Apollo, Cover Ratios for Hyperscaler Bonds Declining, 15 July 2026; Reuters, 29 July 2026Borrowing is still possible, but lenders are demanding better terms. This is the clearest change since R-012.
Grid requests contain a speculative layer
Reuters reports more than 700 GW of US data-centre power requests, over ten times estimated current US data-centre power use. Texas has halted new grid connections while auditing project legitimacy; other utilities now require deposits or proof of financial capability.
Reuters, Texas' halt on powering data centers, 1 September 2026A request for power does not show that capacity has been delivered or put to use.
GPU workload is not physical occupancy
Cast AI reports average GPU utilization of 5% across tens of thousands of AWS, Azure and GCP Kubernetes clusters analysed through April 2026. The sample is cloud Kubernetes infrastructure observed before optimization was enabled, not a census of frontier training clusters or physical data centres.
Cast AI, 2026 State of Kubernetes Optimization ReportBooked capacity can be scarce while the equipment inside it is underused. These figures measure different things; this sample cannot replace R-012's utilization estimate.
Keep different measurements separate.
- Facility occupancy is not substituted for compute utilization.
- Preleased megawatts are not substituted for active workload.
- Grid interconnection requests are not substituted for delivered capacity.
- One Kubernetes fleet's GPU utilization is not substituted for the hyperscaler and neocloud universe.
The original threshold stays in the record. The revised method will apply to future assessments. Utilization Is Not One Number sets out the proposed six-stage replacement.
What this evidence changes.
Capacity remains scarce and accelerator demand is strong. Credit is harder to obtain on favourable terms. Public utilization figures still do not give us a comparable measure.
Capacity can be booked long before the equipment does useful, profitable work. In the historical cycles used for comparison, that gap determined whether the investment paid off.
Will booked infrastructure receive power, do useful work and earn enough to cover its cost before interest and depreciation erode returns?
High on physical and commercial scarcity. High on continued accelerator demand. Medium-high on credit-market deterioration. Low-to-medium on fleet-wide productive utilization, because public telemetry is fragmented.
No aggregate crossing count is issued, so no confidence is claimed for one.
- OpenedTrack booked, delivered, allocated, active and profitable capacity separately in the next version.
- NarrowedUsing grid queues, vacancy or preleasing as direct evidence of productive compute utilization.
- UnchangedContinue monitoring the buildout rather than calling either an overbuild or a clean all-clear.
- ClosedNone this update.
Comparable data connecting booked capacity to power delivery, equipment allocation, active workloads and returns. Or a sustained fall in operating demand, alongside rising vacancy and customers renegotiating capacity bookings.
Set the v0.2 utilization definitions and clarify the other R-012 measures. Publish the next full Hyperscaling assessment when the data support a like-for-like comparison.
Evidence update, published 5 September 2026. The latest full assessment remains R-012, assessed 29 June 2026. Method: NCB-004 v0.3.