HindeSight letters: August · A Credit Cycle Wearing AI’s Costume | September · The Credit Line | October · The AI Catch-22
HindeSight
Oct 2026Hinde Capital

The AI Catch-22: how AI can be right, yet the capital behind it can still lose

The AI Catch-22: how AI can be right, yet the capital behind it can still lose
What the series is attempting

The October letter is one thesis in three parts, and it fits in a sentence: the AI build cannot be completed on the timetable its financing assumes, at the cost its financing assumes.

In Joseph Heller's Catch-22, a pilot could be excused from flying if he was mad, but asking to be excused proved he was sane. There was no way out. The AI build has a catch of its own: each escape from one problem leads into the next. The power, equipment and materials arrive too slowly for the loans already taken out. The debt falls due too soon, before the revenue can start. And when the revenue does arrive, competition may make intelligence too cheap for the owners to earn back what they spent. None of it requires the technology to fail.

The technology can be right. The financing can still be early.

Part I · 23 min read Too Slow How much is planned, how much is needed, and why it cannot arrive on time Read Part I → Part II · 29 min read Too Soon The bill, three real projects, and who carries the timing Read Part II → Part III · 25 min read Too Cheap Who gets paid when intelligence gets cheaper Read Part III →
The three units

The series measures everything in three units. Fix these and the rest follows.

One megawatt (MW)
Electricity, running all year
About 800 American homes, or more than 3,000 British ones
What it takes
Fewer than ten of the newest AI racks
Cost, built and equipped
About $17.6m for the facility (Cushman's average, excluding chips) plus about $20m of chips
Land alone
About $584,000 for powered land in primary US markets
In this series
Costs are quoted per MW
One hall (100 MW)
A hall is one data-centre building: the computer rooms, power and cooling that house the racks of AI chips.
Electricity, running all year
About 80,000 American homes
What it takes
A single building of several hundred racks
Cost, built and equipped
About $3.8bn: $1.8bn of facility and $2bn of chips
Land alone
About $58m
What it must earn
About $850m of revenue a year before any delay or overrun
In this series
Part II's model
One gigawatt (GW = 1,000 MW)
Electricity, running all year
About 800,000 American homes, or more than three million British ones; roughly one large nuclear reactor
What it takes
Several thousand racks, a power station's worth of transformers and switchgear, and about 35,000 tonnes of copper
Cost, built and equipped
About $30bn to $40bn, more than half of it chips
Land alone
Over half a billion dollars before anything is built
In this series
Jupiter is 2.5 GW; the US plans 509 GW, more than eight Britains at peak
Catch-22 in twelve numbers

The series in the figures worth remembering. Each is explained and sourced in the part named.

Too slowPart I →
509 GW planned in the US; about 183 GW likely built by 2032
Most of the plan is paper gigawatts
About 5 years to connect to the grid; about 18 years to open a copper mine
The physical clock cannot be hurried
35.7 million sq ft approved but unbuilt in Loudoun County alone
Approval is not power
Too soonPart II →
Switchgear prices +60% since 2021, against +7% for building materials
The ratchet: costs only go one way
$2.9trn → $5.3trn → about $7trn
The spending estimates rise with the plans
About $800bn of private credit, roughly 70% of the borrowing
Who funds the build, and who ends up holding it
About $1.1trn of debt in place; about $3trn more in commitments
The stack today
CoreWeave: bonds below 80; about 64% implied chance of default
The second station has lit up
Too cheapPart III →
Price actually paid for AI: down about half since May
The product deflates; the debt does not
OpenAI: $856bn on computing against $840bn of revenue, 2026 to 2030
The Catch-22 on the company's own numbers
Gartner: $8.5trn of revenue needed against $3.6trn forecast
About half the investment at risk of write-down
The clocksPart II →
Credit at about July 2007; equities at about October 2007
Where we are in the sequence
LADDERLADDER · Today to 2030 · Morgan Stanley; Goldman Sachs; Gartner; HindeSight analysis
The Ladder of Trillions
What exists, what is promised, what is planned, and what it must earn. $ trillion.
Reported and estimated
024684 · WHAT IT MUST EARNRevenue neededto 2030~8.5Revenue expectedto 2030~3.6The gap: ~$5trn3 · WHAT IS PLANNEDGartner2026–30~7.0Goldman Sachs2025–305.3Morgan Stanley2025–282.9Oct 2026 estimate: ~$4.4trn (2026–28, incl. leases)2 · WHAT IS PROMISEDCommitments signed, not yet debttoday~3.01 · WHAT EXISTSDebt in placetoday~1.1
4 · WHAT IT MUST EARNRevenue needed · to 2030~8.5Revenue expected · to 2030~3.6The gap: ~$5trn3 · WHAT IS PLANNEDGartner · 2026–30~7.0Goldman Sachs · 2025–305.3Morgan Stanley · 2025–282.9Oct 2026 estimate: ~$4.4trn (2026–28, incl. leases)2 · WHAT IS PROMISEDCommitments signed, not yet debt · today~3.01 · WHAT EXISTSDebt in place · today~1.1
Rungs measure different things over different horizons and must not be added. Debt in place combines hyperscaler balance-sheet debt and leases with named project debt; commitments are Morgan Stanley's off-balance-sheet count; revenue needed and expected are Gartner's estimates to 2030.

HS Chart Note About $1.1 trillion of debt exists and about $3 trillion more is promised. The plans call for $3 trillion to $7 trillion of spending. To earn a 12% return, that spending would need about $8.5 trillion of revenue by 2030; Gartner expects about $3.6 trillion. The financing is built on the top rung of the ladder; the revenue is on the one below it.

Source: Morgan Stanley Research (Jul 2025) and Morgan Stanley via Bloomberg Law (Aug 2026); Morgan Stanley (Oct 2026); Goldman Sachs via Cleary Gottlieb (Sep 2026); Gartner (Oct 2026); HindeSight analysis.

How to read the numbersPlanned is not built, committed is not drawn, a year's issuance is not the stock of debt, gross is not net, and a run-rate is not revenue. Part I sets these out in a short panel.