Goldman Sees $545B of AI Project Finance; Chips Take 58.7%

Goldman Sachs credit analysts led by Shamshad Ali raised the bank’s 2027 global AI project finance forecast to $545 billion from $300 billion, the Financial Times reported on 17 November. The bank splits it into $225 billion for data-center shells and $320 billion for compute, the chips, so chips make up 58.7% of the forecast, by our arithmetic.

This year’s total so far is $150 billion, including $48 billion of compute financing, according to ZeroHedge’s summary of the note. Chips are 32.0% of this year’s total, and the 2027 chip figure is 6.7 times this year’s, by our arithmetic.

Business Pill · Vendor Guarantees

An evergreen Business Engineer explainer on vendor guarantees: when a supplier backs its customers’ financing, which can leave hidden leverage after the sale. It teaches the general idea only and does not discuss Goldman Sachs or this forecast.

The key insight: By our arithmetic, chips are 58.7% of the $545 billion Goldman’s credit strategists expect for AI project finance in 2027, against 32.0% of this year’s $150 billion so far, according to ZeroHedge’s summary of the note, and $272 billion, or 68.9%, of the $395 billion increase.

What the Goldman Note Says, as Reported

ZeroHedge names the note as “The Credit Line: From Shell to Compute: An Update on AI Project Finance”. The FT says a growing share of AI capex will be funded via bankruptcy-remote special purpose vehicles, even though issuers pay larger risk premiums, and quotes the note: “the private market’s ability to provide certainty of financing and execution may be especially valuable — even if at a higher cost of capital”.

Compute financing would grow “almost sevenfold” from the $48 billion done year to date, while shell financing roughly doubles, the summary says. Per megawatt, the chips cost about 1.7 times as much as the building, the summary says.

By our arithmetic, the new forecast is $245 billion, or 82%, above the earlier $300 billion, and 3.6 times this year’s $150 billion so far. Shell financing this year is about $102 billion ($150 billion less $48 billion), so the $225 billion forecast is 2.2 times that.

Share of AI project finance by what it pays for: chips (compute) vs data-center shells, 2026 so far ($150 bill
Share of AI project finance by what it pays for: chips (compute) vs data-center shells, 2026 so far ($150 billion) and Goldman’s 2027 forecast ($545 billion). Data: Goldman Sachs credit strategists’ note as reported by ZeroHedge on 17 November 2026, not independently verified; shares and the 2026 shell figure are our arithmetic.

How a Chip Deal Is Built

The summary describes the structure. An SPV owns the GPUs, a customer signs a lease or service contract of five years or less, and lenders are repaid from the contract.

Lenders have little interest in owning the collateral, the summary says, so the weaker deals lean on a residual value guarantee from the chipmaker. Even then the paper comes at 150 to 240bps in investment grade, against about 100bps for ordinary credit, according to the summary.

Chip debt lands in the 3 to 7-year part of the curve, where 60% of investment grade issuers borrow, the summary says. It adds that Goldman expects the damage to stay contained for now, since the deals are off-index and hard to analyze, which also means the buyer base is narrow. On buyers, the FT quotes the team: “material overweights, in aggregate, are unlikely”.

The FT also quotes the team on who borrows this way. Loans securitised on GPUs “have generally come from neoclouds, lower-rated hyperscalers, or chip suppliers themselves, often with some ambiguity around the ultimate end user”, it writes, adding: “we think even higher-rated firms will still lean on the project finance market for a part of their funding plans”.

Grouped bar chart: AI project finance in billions of dollars; shells about 102 in 2026 so far and 225 in the 2027 forecast, 2.2 times; chips 48 in 2026 so far and 320 in the 2027 forecast, 6.7 times
AI project finance, 2026 so far vs Goldman’s 2027 forecast: chips (compute) from $48 billion to $320 billion, 6.7 times; shells from about $102 billion to $225 billion, 2.2 times. Data: Goldman Sachs credit strategists’ note as reported by ZeroHedge on 17 November 2026, not independently verified; the 2026 shell figure and multiples are our arithmetic.

Where the Guarantor Question Comes From

ZeroHedge’s headline says $320 billion of the total “Needs A Guarantor”. That framing is ZeroHedge’s: the $320 billion is the chip financing in the forecast, and ZeroHedge writes that Goldman’s note mentions Nvidia once, in a chart footnote.

ZeroHedge’s own math, labelled as such in its summary, puts the guarantees needed at $80 billion at 25% cover, and at $230 billion to $275 billion on the terms it attributes to Broadcom. It says the three big chipmakers have $94 billion of bonds outstanding between them.

For scale, Nvidia’s latest 10-Q puts the maximum gross exposure under its guarantees at $108.5 billion, mostly lease and power guarantees for SB Energy’s Ohio campus, as our Nvidia guarantees piece reported. Bank of America analysts estimate Nvidia has provided $230 billion in financial guarantees, Reuters reported on 16 November. The FT notes Bloomberg’s report last week that a $500 billion AI finance platform launched in August had delivered no deals, as lenders wanted bigger backstops.

The FT also quotes a speech by Federal Reserve governor Michael Barr on the obligations that sit outside balance sheets: “The FSR estimates that the potential maximum nominal value of these types of obligations, or off-balance-sheet liabilities, now exceeds $3 trillion.” Barr lists residual value guarantees among them.

Against the Deals in Our Coverage

The largest chip financing in our coverage is SpaceX’s. Bloomberg reported on 4 November that an Apollo-backed entity would buy approximately $47.3 billion of Nvidia GPUs and related equipment, funded by roughly $40 billion of debt and a $7 billion lease prepayment from SpaceX (our SpaceX piece). Goldman’s $320 billion chip forecast equals about 6.8 deals of that size, by our arithmetic.

Three Bloomberg reports in the past two days involve AI data-center or chip loans. Bloomberg reported that HSBC sold a roughly $300 million slice of an $18 billion Oracle-tied data center loan at prices in the mid-80s (our HSBC piece), and that two loans tied to ByteDance’s computing, $480 million and $488 million, were frozen or delayed, $968 million together (our ByteDance loan piece).

On the shell side, the summary says funds are underweight 13 of 16 junk data center bonds, and that at a 10% coupon a tenant has to pay 2.7 percentage points more in rent yield to keep the same debt coverage.

The Structural Read

As we read the summary, the risk moves with the mix: the chip deals it describes rest on contracts of five years or less and collateral lenders have little interest in owning, which is why the weaker ones lean on a chipmaker’s residual value guarantee.

As we read it, the pricing the summary gives, 150 to 240bps in investment grade against about 100bps for ordinary credit, is the cost of that structure, and the off-index, hard-to-analyze deals it describes leave a narrower set of buyers for the paper.

ZeroHedge, opening its 17 November 2026 article on the Goldman note

“The corporate bond market is full, so Wall Street has a new plan: finance the chips.”

Three Implications

FOR THE CHIPMAKER GUARANTEES Nvidia’s filed guarantee exposure is in our Nvidia guarantees piece.

FOR A GPU FINANCING IN PRACTICE The SpaceX-linked $47.3 billion package is in our SpaceX piece.

FOR LOANS MEETING RESISTANCE The HSBC sale of an Oracle-tied loan slice is in our HSBC piece.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Capex Map & The State of AI Hyperscalers.

The analysis sets out its subject as: “How the money flows, where it cracks, what it buys, who can absorb it, and what happens when revenue can’t catch up.”

As we read it, Goldman’s note, as summarised by ZeroHedge, puts a number on the “who can absorb it” question for the chip layer: $320 billion of compute financing in 2027, 58.7% of the total by our arithmetic. The analysis predates the note and does not discuss it.

What Is Not Established

We have not seen the Goldman note. The forecast and its split are in both the FT report and ZeroHedge’s summary; the year-to-date figures, deal structure, pricing and curve figures come from ZeroHedge’s summary only, which sits in front of a premium article we could not read in full and mixes the note’s figures with ZeroHedge’s own math. The FT calls the new forecast more than double the previous estimate; $545 billion is 1.82 times $300 billion, by our arithmetic. The 2026 shell figure is our subtraction.

We do not use the summary’s claims about what Nvidia’s leadership will do or about the take-up of Nvidia’s financing platform, which we could not check. Goldman Sachs was not contacted for this piece. Not investment advice; for information only.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Capex Map & The State of AI Hyperscalers

The Business Engineer analysis of how the AI build-out is financed and who can absorb it.

Read the premium analysis →Get 50% off →

The Bottom Line

Goldman’s credit analysts expect $545 billion of AI project finance in 2027, $320 billion of it for compute, the FT reported. That puts chips at 58.7% of the forecast against 32.0% of this year’s $150 billion so far, by our arithmetic.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The forecast, its split and the quotes from the Goldman team are from the Financial Times (FT Alphaville) of 17 November 2026; year-to-date figures, deal structure, pricing and curve figures are from ZeroHedge’s summary of the same Goldman Sachs credit note, published the same day; we have not seen the note itself. Fed governor Barr’s words are as quoted by the FT. ZeroHedge’s own math is labelled as such. SpaceX, Oracle-tied, ByteDance-linked and Nvidia guarantee figures come from our earlier pieces, linked, and the outlets they cite. Shares, multiples and the 2026 shell figure are our arithmetic. Goldman Sachs was not contacted for this piece. Not investment advice; for information only.

Sources: Financial Times (FT Alphaville), 17 Nov 2026 · ZeroHedge, summary of Goldman Sachs’ note, 17 Nov 2026 · @zerohedge on X, 17 Nov 2026

Scroll to Top
FourWeekMBA