Goldman Sachs put two curves on one chart: chip sales, and new data center capacity. By the end of 2028, the gap between them is 33 gigawatts.
33 GW is roughly the UK's baseline electricity demand.
Cards can be reordered. The grid cannot.
A gigawatt is 1,000 megawatts. A Shanghai computing center designed for 18 MW draws 0.018 GW. Run 33 GW flat out for a year: 289 billion kilowatt-hours.
The number comes from Goldman's credit strategy team, in a November 17 note led by Shamshad Ali. The note is not about power but money: it lifts the forecast for global AI project financing in 2027 from $300 billion to $545 billion, an upgrade of more than 80%.
The money lands before the power
Split the $545 billion: $225 billion into data center buildings, $320 billion into GPUs and other compute.
The note flags both: buildings may not be energized, chips likewise. Those are the coldest lines in the report.
A hall goes up, a card goes in, the power arrives on another schedule. Building a data center and clearing an interconnection queue run on two clocks, with permits, land and construction between them. 33 GW measures not annual consumption but the slope difference between two curves, and that difference compounds.
This year to date: $150 billion of AI financing, only $48 billion of it for chips.
Next year, chips alone: $320 billion. 320 over 48 is 6.7 times, close to 7 times, as the note puts it.
Spread $545 billion across 365 days: about $1.5 billion a day, 3.6 times this year's total. That is why the gap lands in 2028: money first, power second.
Money multiplies nearly 7 times in a year. The grid does not.
Sellers are more optimistic than lenders
In the same week, SemiAnalysis cut its forecast for AI accelerator capital spending and shipments, by nearly 30% for 2028. Goldman raised its financing forecast by more than 80%.
One trims equipment, one marks up money. Different measurements, not a contradiction.
Goldman's chip sales desk implies capex nearly 30% above that. One firm, two desks, two numbers, 30% apart. Not an error: a position. Credit asks whether the borrower repays. Sales asks how much product moves.
Careful leaves traces. Cloud providers have issued $295 billion of debt globally this year, breaking Goldman's full-year expectation two months early. The same note concedes that tighter terms have slowed issuance, with deals piling up, and advises selectivity in data center credit.
A forecast broken early, a backlog forming. Together they describe the market: money still moving, but pickier. Picky about what? Power. A data center with no interconnection date and one with a 20-year power purchase agreement do not clear at the same price or covenants.
Three trillion off the books
Federal Reserve governor Michael Barr gave two numbers. On-balance-sheet debt at the seven largest AI companies should pass $800 billion by the end of 2026. Off-balance-sheet items, including long-term contracts, leases and residual value guarantees, carry a nominal ceiling above $3 trillion.
$800 billion on the books, $3 trillion off. Barr compared the cycle to telecom investment in the late 1990s. That round left fiber in the ground. This one leaves machine halls and cards. The point is structure, not scale: borrow first, lay the infrastructure, wait for demand. Demand is slow. Interest is not.
Jim Zelter, president of Apollo Global Management, is more direct: loans fully amortized in five years, backstopped by investment-grade counterparties, no residual value risk.
In plain English: lend, but do not leave me the last hand.
Maturities stretch. Backstops move further away. At the end of the chain, the price gets hard. Credit markets do not fear expense. They fear losing the anchor, and when leases and residual values sit off the statements, the anchor is not on them.
Who pays the $35 billion
US Democrats are drafting a data center power reform plan. By one estimate, operators may pay $35 billion more over 10 years: $3.5 billion a year, $290 million a month. Not chips, not halls. The grid. The plan is not law; the number is out, and it says one thing: interconnection is becoming a separately priced service.
Operators will not absorb it. It travels down to the cloud provider renting the racks, then to the customer buying tokens. How power is priced in a rack lease is the least glamorous and most decisive line in this business. Any inversion shows up first in cloud gross margin. So $35 billion is not an environmental figure. It is a cost figure, and it will reach the model price sheet.
The cost table needs a new line
In Pudong, Shanghai, the Liming smart computing center is in trial operation: 18 MW, mainly domestic compute, biomass power and waste-heat cooling. Reports name Zhipu as a user; Zhipu declined to comment. Filling 33 GW takes 1,833 of them. The point is structure, not volume: more output per watt, not more watts.
Back to 33 GW. It is not a forecast. It is the distance between two curves.
Curves bend. Physics does not. Financing buys cards, not an energized line.
For two years, AI's cost curve has been set by card price, utilization and model efficiency. A fourth is arriving: the price of a kilowatt-hour. Once it enters the table, handsome algorithmic numbers get reseeded. A model that pushes inference to the floor, running where power is dearest, may not beat a rival on the ledger.
33 GW does not say power is scarce. It says power arrives slower than money.
The divide is not capacity. It is the interconnection queue.
One question: if an interconnection agreement is harder to sign than a procurement contract, what should an AI company book first, the chips or the power?
The table will be rewritten. The top line will not be cards.
It will be the price of a kilowatt-hour.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
