Latest full assessment · 29 June 2026

The AI buildout is now a balance-sheet bet.

Four of five conditions crossed their historical thresholds. The unresolved question is whether the new capacity gets used enough to justify its cost.

The five signalsWhat would change it?How to read it
4 / 5
Beyond the band
Four measures have crossed their historical thresholds.
1
Still uncertain
Utilization remains inside the uncertainty band.
0
Goalposts moved
The thresholds were not reset after the fact.

5 September update ↗ Utilization needs a clear, comparable measure. The June assessment is unchanged.

Hyperscaling · R-01229 June 2026
Threshold
4 crossed1 unresolved
Utilization · 40–60%

Inside the band in R-012. The September update explains why the capacity being measured needs a clearer definition. The estimate is unchanged.

The beam tours the signals. Select one to hold its explanation. Positions show status, not probability or comparable magnitudes.

What moved

The five signals behind the reading.

Cost

$725B

Announced 2026 capex, up sharply year over year.

Crossed
Credit

$121B

AI-infrastructure debt issued in 2025, far above the recent norm.

Crossed
Profitability

~10:1

Capex relative to AI-services revenue in the highest cited estimate.

Crossed
Utilization

40–60%

Inside the R-012 band; capacity definition under review.

Approaching
Competition

61%

NVIDIA revenue concentrated in four customers in the cited reading.

Crossed
The correspondents at work

Booked. Busy. Productive?

Five perspectives on the utilization question.

Explore the example →

Explore a hypothetical case

What would change this assessment?

Choose a possible development to see what it would tell us—and what we would still need to check.

Published assessment stays unchangedR-012 · 29 June 2026, with the 5 September clarification. These cases do not change the scores, thresholds or Radar graphic.

Hypothetical · Utilization · quality of the evidence

Productive use improves

Weakens the concern about unused capacity

If comparable measurements show more capacity doing useful, paying work, the underuse concern becomes weaker. It would not, by itself, settle whether the returns cover the investment.

Evidence needed

  • Active work measured against a clearly defined amount of available capacity.
  • The same population and observation period, so the new reading can be compared with the old one.
  • Evidence of useful output and revenue, with costs shown separately.

What this would not establish

More bookings or fewer vacant facilities would not establish this case.

Read the published update ↗

Hypothetical · Profitability · returns on investment

Revenue falls short

Strengthens the concern about recovering the cost

If comparable AI-services revenue weakens while investment and operating costs remain high, the case that the buildout can pay for itself becomes harder to support.

Evidence needed

  • Revenue for a defined set of services and companies, over comparable periods.
  • The investment and operating costs associated with those services.
  • An explanation of changes in reporting, prices or the mix of services.

What this would not establish

One disappointing company result would not establish a sector-wide change.

Read the published update ↗

Hypothetical · Credit · ability to fund commitments

Financing gets tighter

Strengthens the concern about financing the buildout

If borrowing becomes more expensive or harder to renew, operators may have less room to fund expansion or wait for returns. That does not tell us how much capacity is doing useful work.

Evidence needed

  • Comparable borrowing terms: spreads, maturities, collateral and lender conditions.
  • Evidence of completed financing or refinancing, rather than announcements alone.
  • The companies and commitments affected, with the dates of the change.

What this would not establish

A wider credit spread alone would not prove that demand or productive use has fallen.

Read the published update ↗

Hypothetical · Demand · what reservations can establish

More capacity is booked

Leaves the utilization question unresolved

More reservations can show commercial demand. They do not establish that the capacity has been delivered, put to work or earned a return.

Evidence needed

  • A clear distinction between requests, signed bookings and delivered capacity.
  • A link from those bookings to available equipment, active workloads and returns.
  • Comparable measures over time, including cancellations or unused reservations.

What this would not establish

Demand can strengthen while the published utilization assessment stays unchanged.

Read the published update ↗

“Strengthens” and “weakens” refer to the concern that the buildout may not justify its cost. They describe a possible argument, not a new measurement or a probability. See how the capacity measures differ →

Why it matters

Can the capacity earn its keep?

01

Separate demand from use.

Contracted capacity is not the same as active or economically productive capacity.

02

Duration matters.

Long-lived infrastructure is being financed against shorter hardware and market cycles.

03

Utilization is the open question.

The outcome depends on whether expensive capacity does useful work or sits underused.

Why trust it

The rules are written before the answer is known.

The thresholds stay fixed. The result gets published either way.

1

Choose the lines.

Track the measures that distinguished past buildouts that paid off from those that failed.

2

Set the thresholds.

Fix what counts as meaningful movement before the window closes.

3

Publish either way.

If nothing changes, say so. If the method fails, leave that visible too.

See the record ↗Read the latest update ↗Explore the method →