The cost of idle AI capex is a lag problem, not a rate problem

Our read: the cost of AI capex sitting idle is a $10–20B-a-year problem, not a rate problem (our scenario).

Start with the run-rate. $AMZN, $GOOGL, $META, $MSFT and $ORCL spent about $65.8B a month on capex in their latest quarters, $META's being Q3 from its release (our calc). That is total capex; none of them splits out AI.

Now add a lag. Gear that is bought but not yet earning, for 3 to 6 months, means $197–395B sitting idle at any time (our scenario).

$NVDA's Jensen Huang put it at "a few hundred billion dollars sitting idle." Our range brackets his.

The carrying cost: at a 5.1–5.6% weighted average coupon on their existing debt (our calc; $MSFT uses the five-company average as a proxy), that idle pile costs about $10.4B a year at a 3-month lag and $20.7B at 6 months (our scenario, pre-tax). Cutting the lag from 6 months to 1 saves ~$17.3B a year, about 3 points of interest cost on ~$545B of debt (our calc; no offset for interest income on cash).

So the number to watch is how fast new capacity gets switched on, more than where rates go.

The live meter: construction in progress (spending not yet in service) as a share of gross property and equipment (our calc):
$META 15.3% (mid-2024) → 27.4% (Sep 2026), flat since June
$GOOGL 20.2% (end-2024) → 28.8% (mid-2026)
$ORCL 16.2% (May 2024) → 32.6% (May 2026), 31.6% (Aug 2026)

Which do you watch more closely: the lag or the rate?

What would prove us wrong

Set before publication. We report back in public when one trips.

• That share turns down at $META and $GOOGL and keeps falling at $ORCL (it eased to 31.6% in Aug from 32.6% in May), in the next 10-Qs. Not tripped: $GOOGL is still climbing and $META held flat at 27.4% in Q3. Open · next check: $GOOGL Q3 10-Q (late Oct), $ORCL 10-Q (Dec), $META 10-K (early 2027)

Method and sources

Figures marked "our calc" are our own arithmetic on the cited primary sources (company filings, earnings releases, public datasets). Every figure was checked against its source before publication.

Original post on X → Download data note

General research, the same for every reader. Not investment advice.