W.161Goldman raised its 2027 AI project finance forecast to $545B from $300B, accounts say
Goldman Sachs’s credit strategists raised their forecast for AI project finance in 2027 to $545B from $300B, according to accounts of a note dated November 17; the note is not public and we could not read it. The accounts split the forecast into about $225B for data center shells and $320B for chips. Chip financing so far in 2026 is put at $48B, so the chip forecast is 6.7 times this year’s total to date. Two accounts give $150B for all AI project finance so far this year; one adds that Goldman’s initial full-year forecast was $76B.
What was reported
Goldman’s credit strategists raised the 2027 forecast to $545 billion from $300 billion. Dataconomy, two KuCoin News reposts, Sina Finance and Caiwen carry the figure. One repost gives the note’s title, “The Credit Line: From Shell to Compute: An Update on AI Project Finance”, and its date, November 17. We could not read the note, and we found no report of it by a major news agency. The rise is $245 billion, or 81.7%. Goldman’s head of credit strategy research gave the $300 billion figure in August and described it as financing beyond the hyperscalers’ own bond issuance.
Chips are the larger part. The accounts put data center shells, meaning buildings and power connections, at about $225 billion and chips at $320 billion. Chips are 58.7% of the forecast and shells 41.3%. No account gives a split for the earlier $300 billion.
Chip financing so far this year is $48 billion. Dataconomy, both KuCoin reposts and Caiwen give that figure, and none gives the date it runs to. The 2027 chip forecast is 6.7 times it. The figure is close to the $48.4 billion of GPU-backed debt a separate tracker counted to October 10, which we covered in W.046; we do not know whether the two counts cover the same deals.
One account gives the rest of 2026. The KuCoin repost of 528btc says AI project finance so far this year is about $150 billion: about $80 billion of data center bonds, $22 billion of non-syndicated construction financing and $48 billion of chip financing. Caiwen gives the same total. The repost says Goldman’s initial full-year forecast was $76 billion, so the year to date is 2.0 times it. The same repost’s summary misstates the chip figure as $3.2 trillion, so we treat figures that only it carries with caution.
The 2027 forecast is 3.6 times this year’s total so far
Billions of dollars of AI project finance, from accounts of Goldman Sachs’s note of November 17. Faint grey: Goldman’s initial forecast for 2026 and its previous forecast for 2027, for which no split is reported. Grey: data center shells; for 2026 so far, data center bonds plus non-syndicated construction financing. Pink: chips. The 2026 figures are for the year so far, to a date the accounts do not give; the initial forecast and the 2026 split come from one account.
The figures
| $B | Shells | Chips | Total |
|---|---|---|---|
| 2026, initial forecast | n/a | n/a | 76 |
| 2026, so far | 102 | 48 | 150 |
| 2027, previous forecast | n/a | n/a | 300 |
| 2027, new forecast | ~225 | 320 | 545 |
Chips would grow 6.7 times and shells 2.2 times
Billions of dollars. Faint grey: 2026 so far. Grey: the 2027 forecast for shells. Pink: the 2027 forecast for chips. The 2026 shell figure is our sum of two lines one account reports, $80 billion of data center bonds and $22 billion of non-syndicated construction financing; the accounts do not say Goldman defines shells that way. The comparison is a partial year against a full-year forecast.
What the forecast assumes
The cost assumptions come from one account. The 528btc repost says Goldman assumes about $15 million a megawatt for a shell and $25 million a megawatt for the chips in it. It says 45% to 50% of shell cost is financed with debt, 30% of hyperscalers’ compute and 45% of others’.
On those assumptions the two halves describe a similar amount of capacity. This is our arithmetic, not Goldman’s. About $225 billion of shell debt at $15 million a megawatt and 45% to 50% debt finances 30.0 to 33.3 GW. $320 billion of chip debt at $25 million a megawatt finances 28.4 GW if 45% is borrowed and 42.7 GW if 30% is.
The note is reported to flag a capacity gap. The repost says Goldman’s equity analysts expect 113 GW of compute to ship by 2028, against 80 GW of data center capacity identified, a gap of 33 GW. Sina reports the same gap.
Chip debt is priced wider and runs shorter. The repost says new investment-grade chip financings priced at spreads of 150 to 240 basis points, against about 100 for five-to-seven-year investment-grade bonds. Dataconomy says chip contracts run five years or less. Sina quotes the note, without naming a speaker, as saying investors will ask for more equity, lower loan-to-value ratios and tighter terms.
On the relayed assumptions, each half finances about 28 to 43 GW
Gigawatts; our arithmetic on assumptions one account attributes to Goldman. Dashed pink: the forecast debt divided by cost per megawatt times the share financed with debt. Shells: $225 billion at $15 million a megawatt, with 45% to 50% borrowed. Chips: $320 billion at $25 million a megawatt, with 30% (hyperscalers) to 45% (others) borrowed. The solid part of each bar is the low end and the outlined part runs to the high end. Goldman is not reported to have published these capacity figures.
How it compares
It is larger than Goldman’s figure for the hyperscalers’ own bonds. An account of Goldman’s research on November 6 put direct hyperscaler issuance expected in 2027 at about $440 billion. That and the new project finance forecast sum to $985 billion; the two figures were made eleven days apart.
JPMorgan’s comparable line is smaller and narrower. JPMorgan’s forecast of November 13, covered in W.124, has $225 billion of net new data center and chip financing in US investment-grade bonds in 2027 and $225 billion of net new supply from hyperscalers, $450 billion together. Goldman’s count for 2026, per one account, includes high-yield bonds and non-syndicated construction financing, and some accounts call the forecast global, so the two are not like for like. Fitch said on November 18 that five hyperscalers could add about $280 billion in 2027 if 25% of their capital spending is debt-funded.
Two of the largest chip financings reported this year had not closed when last reported. KBRA gave preliminary A- ratings on November 16 to about $29.6 billion of notes funding Nvidia systems leased to SpaceX (W.153). Bloomberg reported on September 16 that ten banks were lining up a $22 billion loan for Crux AI, a Blackstone-backed venture, to buy Google chips; we found no report that it has closed. Together they exceed the $48 billion Goldman counts so far, so that count cannot include both in full.
Several accounts tie the note to Nvidia’s guarantees. Nvidia’s CFO commentary of November 17 puts its maximum gross exposure under guarantees at $118.0 billion at October 25. That is $9.5 billion more than the $108.5 billion in its 10-Q for July 26; the subtraction is ours. Of that, $105.0 billion supports the building of an Ohio campus by SB Energy, not chip purchases. The commentary says no payments have been made under any guarantee. We cannot tell whether the framing that a guarantor has stepped back is Goldman’s or the accounts’.
Goldman’s project finance forecast exceeds its figure for hyperscalers’ own bonds
Billions of dollars forecast for 2027. Grey and pink: Goldman’s project finance forecast, shells and chips. Black: Goldman’s figure for direct hyperscaler issuance, as relayed on November 6. Strong grey: JPMorgan’s forecasts of November 13 and Fitch’s estimate of November 18, both for US investment-grade bonds only. The scopes differ: Goldman’s count for 2026, per one account, includes high-yield bonds and non-syndicated construction financing.
What this changes on inthesingularity.ai
- W.124 used JPMorgan’s $225 billion for data center and chip financing in 2027. Goldman’s figure for a wider market is 2.4 times that.
- W.046 counted $48.4 billion of GPU-backed debt announced by October 10. Goldman’s $48 billion so far this year is close to it; the definitions have not been compared.
- Lead for Hyperscaler capex funding and Neocloud capex funding: a project finance line for 2027 split into shells and chips, with Goldman’s $225 billion and $320 billion as one estimate, once the note’s definitions can be read.
- Lead: a count of chip financings closed in 2026, to test the $48 billion against named deals.
Sources and method
| Item | Figure | Source |
|---|---|---|
| The note | 2027 AI project finance forecast raised to $545 billion from $300 billion; about $225 billion for data center shells and $320 billion for chips; chip financing of $48 billion so far in 2026; chip contracts of five years or less | Dataconomy, Nov 18 |
| The note | Title, date (Nov 17) and length of the note; about $150 billion of AI project finance so far in 2026, of which about $80 billion data center bonds, $22 billion non-syndicated construction financing and $48 billion chip financing; initial 2026 forecast of $76 billion; about $15 million a megawatt for shells and $25 million for chips; debt shares of 45–50% for shells, 30% for hyperscalers’ compute and 45% for others; 113 GW of compute shipped by 2028 against 80 GW of data center capacity identified; chip-financing spreads of 150–240 basis points | KuCoin News, reposting 528btc, which credits CoinDesk, Nov 18 |
| The note | The same headline forecast and split; the hyperscaler capital spending and GDP figures we did not use | KuCoin News, reposting Crypto Briefing, Nov 17 |
| The note | Shamshad Ali named as leading the analysts; the forecast is still nearly 30% below the spending implied by Goldman’s chip sales estimates; investors will seek more equity and tighter terms; a 33 GW gap | Sina Finance, Nov 17 |
| The note | Published Tuesday by the credit strategy team; about $150 billion so far in 2026; the raise is over 80% | Caiwen, via 163.com |
| The note | The forecast is global; chips are mainly GPUs and custom semiconductors | The Dark Side of the Boom, Nov 17 (teaser) |
| Earlier Goldman figures | About $440 billion of direct hyperscaler issuance expected in 2027 | The Dark Side of the Boom, Nov 6, citing Goldman |
| Earlier Goldman figures | Data center and project finance of about $300 billion in 2027, beyond direct hyperscaler issuance | Goldman Sachs Exchanges, August |
| Other forecasts | JPMorgan, 2027, US investment grade: hyperscalers $225 billion; data center and chip financing $225 billion | Briefs, citing Bloomberg, Nov 13 |
| Other forecasts | Fitch: about $280 billion of 2027 issuance from five hyperscalers into US investment-grade non-financials, assuming 25% of capital spending is debt-funded | TNGlobal, Nov 18 |
| Nvidia | Maximum gross exposure under guarantees of $118.0 billion at October 25: SB Energy $105.0 billion, power and systems $9.5 billion, neocloud land, power and shell $3.5 billion; no payments made to date | Nvidia CFO commentary, Nov 17 |
| Nvidia | Maximum gross exposure under guarantees of $108.5 billion at July 26 | Nvidia 10-Q, quarter ended Jul 26, 2026 |
| Deals | Preliminary A- ratings on about $27.2 billion of senior and about $2.4 billion of subordinated notes of Saturn Funding Trust, which leases Nvidia systems to SpaceX | KBRA, Nov 16 |
| Deals | Ten banks lining up a $22 billion secured loan for Crux AI to buy Google chips; Blackstone committed $5 billion of equity | Pulse 2.0, citing Bloomberg, Sep 16 |
| Not used | Goldman’s strategists: five US hyperscalers’ capital spending of $1.2 trillion in 2027 and about $1.4 trillion in 2028 | PYMNTS, citing Bloomberg, Sep 25 |
| Not used | Goldman’s economists: annualized spending on AI hardware and intellectual property reaching $1.4 trillion by the end of 2027, nearly 4% of the US economy | Reuters, via TimesLIVE, Nov 12 |
What we read. Accounts of the note by Dataconomy, KuCoin News (two reposts), Sina Finance, Caiwen and a newsletter teaser; Goldman’s August podcast page; accounts of JPMorgan’s and Fitch’s forecasts; Nvidia’s CFO commentary; KBRA’s release. We could not read Goldman’s note or the CoinDesk report that one repost credits, and ZeroHedge’s article is listed as a trigger from its headline only. No account is independent of the note, and several appear to derive from one another. Arithmetic. Raise: 545 − 300 = 245; 545 ÷ 300 = 1.817, so 81.7%. Shares: 320 ÷ 545 = 58.7%; 225 ÷ 545 = 41.3%; 225 + 320 = 545. 2026 so far: 80 + 22 + 48 = 150; shells so far, 80 + 22 = 102. Multiples: 320 ÷ 48 = 6.67; 225 ÷ 102 = 2.21; 545 ÷ 150 = 3.63; 150 ÷ 76 = 1.97. Capacity: 225,000 ÷ (15 × 0.50) = 30000 MW; ÷ (15 × 0.45) = 33333 MW; 320,000 ÷ (25 × 0.45) = 28444 MW; ÷ (25 × 0.30) = 42667 MW. Gap: 113 − 80 = 33 GW. Goldman’s two 2027 figures: 440 + 545 = 985. JPMorgan: 225 + 225 = 450; 545 ÷ 225 = 2.42. Nvidia: 105.0 + 9.5 + 3.5 = 118.0; 118.0 − 108.5 = 9.5. SpaceX and Crux: 29.6 + 22 = 51.6, more than 48. Conflicts between sources. Sina calls the raise more than double; $545 billion is 81.7% above $300 billion, and Caiwen’s “over 80%” matches. Some accounts call the forecast global and others give no geography. Dataconomy and one KuCoin repost say Goldman lifted its estimate of hyperscalers’ 2027 capital spending to $1.2 trillion to $1.4 trillion, nearly 4% of US GDP at the top of the range. Those figures match two other Goldman publications, a strategy forecast of September 25 and an economics note reported on November 12, and the detailed accounts of the credit note do not carry them; we did not use them. The 528btc repost’s summary gives chip financing as $3.2 trillion, and its body gives $320 billion. Limits. Every figure from the note is second-hand. The 2026 total, its bond and construction parts, the initial forecast, the cost and debt-share assumptions, the capacity gap and the spreads rest on one repost, with the total and the gap each confirmed by one other account. The date to which “so far” runs is not given. The capacity figures are our arithmetic and assume the forecast is all debt on new capacity priced at the relayed costs. The 2027 comparisons mix forecasts of different scope, made on different dates.
Wire reports are written within hours of the news, on a six-hour schedule. Each lists its sources and shows the arithmetic behind its figures. Not investment advice.