$545 billion.
On November 17, Goldman Sachs' credit strategy team raised its forecast for global AI project financing in 2027 from $300 billion to $545 billion, an increase of more than 80%.
That same week, Nvidia's $500 billion chip financing platform, launched in August with institutions including Goldman, had gone three months without completing a deal.
One says the money is coming. One says the money has not moved.
Where the $545 billion goes
Split the number. $225 billion goes into data center buildings, civil works not necessarily energized. $320 billion goes into GPUs and other chips, cards and racks also not necessarily energized. Together, exactly $545 billion.
The chip piece is 6.7 times the year to date. AI project financing totals about $150 billion this year, of which chips are $48 billion. 320 divided by 48 is 6.67.
Chip financing must reach seven times the year-to-date figure in one year. Where does the money come from?
Its name is project finance, not corporate finance: repaid from the project's cash flow, not the parent's credit. Data centers fit: long leases, power contracts, stable inflows. GPUs do not. They depreciate fast, residual value after three years is unclear, and cash flow must land in the first two. One asset is underwritten over 20 years, the other over three.
Goldman's answer is debt. Large cloud providers issued $295 billion globally this year, passing its full-year forecast two months early. 545 divided by 295 is 1.85. JPMorgan adds the rest: 2027 TMT investment-grade issuance of $675 billion, two thirds AI-related, about $450 billion. 545 minus 450 leaves $95 billion, to come from equity, private capital or vendors.
Goldman's chip sales team carries an implied capex estimate nearly 30% above $545 billion: $708.5 billion.
25% against 86%
Now the platform: $500 billion, three months, no deals. Nvidia guarantees up to 25% of each transaction. Lenders want more.
On a $10 billion financing, Nvidia covers 2.5 and the other 7.5 sits with lenders and borrowers. In May, Broadcom guaranteed $30 billion of Anthropic's $35 billion in chip financing. 30 divided by 35 is 86%. 25% against 86%, a 61-point gap. That is what "cannot agree" looks like in numbers.
For lenders, residual risk is too large. A GPU is not a house: depreciation in three years, resale and buyer have no standard answer. Without enough guarantee the deal does not enter the risk model. Risk control is not bearish on AI; it must price the worst case, and nobody can quote it. With the residual column blank, the sheet fails.
For Nvidia, 25% is already the offer. Its total guarantees are about $118 billion; 25% of a $500 billion platform implies a $125 billion cap, added on top if the platform fills. It lends the balance sheet to customers a second time, off the balance sheet.
Signing is harder than lending. A guarantee trades vendor credit for orders: a contingent liability, often kept out of the main liability line, which vendors like and banks do not.
Who carries the residual
Apollo Global Management President Jim Zelter set out the terms: full amortization within five years, an investment-grade counterparty backstop, and no residual value risk.
Three conditions, one meaning: the risk should not land on us.
Five-year amortization shortens the tenor; a stronger counterparty narrows default risk; no residual pushes the final disposal price to the vendor or borrower. That is why the platform stalls: not for lack of money, but for lack of anyone willing to sign the residual column.
Federal Reserve Governor Michael Barr supplies the scale. On-balance-sheet debt for the seven largest AI companies should exceed $800 billion by the end of 2026. Off-balance-sheet obligations, including long-term contracts, leases and residual guarantees, carry a potential notional ceiling above $3 trillion. That is 3.75 times. The risk sits in the notes, not the liability line. Barr compares it to the telecom buildout of the late 1990s: demand did not fail first, balance sheets did.
Who underwrites compute
Put the two stories on one sheet. A chip company sells cards. Customers lack money, so it helps them finance. Banks balk, so it guarantees. The guarantee is too thin, so the deal stops.
One step further and the chip company is the industry's credit intermediary: it earns the spread and pays for the cycle. Nvidia still earns chip money, but may pay a cycle bill.
$118 billion is not debt. It is a promise. A promise is a moat in a boom and a rope in a turn.
The money is waiting for a signature
One line in the report matters more than the number. It concedes that tighter terms slowed issuance this quarter, with unfinished deals piling up. It recommends being selective on data center credit and expects spreads to widen. The same report raised the forecast while saying issuance is slowing. It forecasts demand; terms are stuck.
On demand, the chip sales curve and new data center capacity diverge by 33 gigawatts by the end of 2028, roughly UK baseload demand. On supply, SemiAnalysis cut its 2028 AI accelerator forecast by nearly 30%.
A 33-gigawatt gap and a 30% cut say the bottleneck is not the will to buy. It is whether power can be connected. An unpowered data center is an empty building, and an empty building still pays interest. Rent waits; interest does not.
Whether $545 billion lands depends on three things: a guarantee above 25%, someone taking the residual, and power prices held down. None is decided by a model company.
Compute has been financialized to the point where the vendor guarantees the customer. The risk has moved off the balance sheet.
One question left. If the guarantee moves from 25% to 50%, who pays the extra exposure: shareholders, or the customers renting the cards?
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
