Skip to content
Advertisement
Goldman Sachs forecasts $545 billion in AI project finance for 2027

Goldman Sachs forecasts $545 billion in AI project finance for 2027

The bank's credit team expects chip financing to grow almost sevenfold, and much of that debt may come without a traditional guarantor

Goldman Sachs’ credit team has raised its AI project finance forecast for 2027 to $545 billion. Its previous estimate was $300 billion.

Goldman splits the $545 billion into two buckets. About $225 billion would fund data center shells, meaning the buildings, power hookups and physical structures that house the hardware.

The other $320 billion would fund the chips inside those buildings. Put simply, Wall Street now expects to lend more against what sits in the server racks than against the walls around them.

Compute is where the growth really shows up. Chip financing has reached $48 billion so far this year, according to Goldman, and the bank expects it to grow almost sevenfold to $320 billion by 2027.

The revision also fits a wider shift in Goldman’s thinking. The bank has lifted its estimate for hyperscaler capital expenditure in 2027 to between $1.2 trillion and $1.4 trillion.

At the top of that range, annualized spending on AI hardware and IP would come to nearly 4% of US GDP.

Advertisement

How a GPU deal actually works

A special-purpose vehicle, or SPV, owns the chips. An SPV is a standalone legal entity created for one job, which keeps the assets and debt separate from the parent company’s balance sheet.

A customer then signs a lease or service contract for the computing capacity. Under the structure Goldman describes, those contracts run five years or less.

Lenders are paid back from the cash that contract produces. If the customer pays, the debt gets serviced. If the customer stops paying, lenders are mostly left holding the hardware and the paperwork.

That leads to the key caveat in Goldman’s research. A large share of the $320 billion in chip financing may come without traditional corporate guarantees.

In a typical corporate loan, a parent company with deep pockets promises to cover the debt if things go sideways. Without that backstop, lenders are betting on two things: the contract itself and the creditworthiness of the customer who signed it.

The deals already on the table

A $40 billion GPU financing package involving SpaceX was recently reported, with Goldman Sachs and Apollo backing the arrangement. Separately, a $22 billion chip financing has been tied to Crux AI, a venture connected to Blackstone and Alphabet.

These deals sit alongside a separate wave of borrowing. Earlier reports pointed to hundreds of billions of dollars in direct investment-grade bond issuance from the major hyperscalers in 2026.

What this means for lenders, borrowers and everyone else

The research flags several pressure points that could test that assumption. They include power supply constraints, regulatory opposition to data center projects, and whether end users can earn enough revenue to meet their contract obligations.

Growth from $48 billion to $320 billion means a market that barely existed at scale is expected to become one of the largest corners of project finance within roughly two years.

Banks like Goldman and alternative asset managers like Apollo and Blackstone are clearly positioned to arrange and fund these deals, which gives them a direct stake in the market’s growth.

The figures worth tracking from here are concrete. Watch how much of the $320 billion chip forecast actually closes, how many of those deals carry a corporate guarantee, and whether power and permitting constraints start delaying the facilities those chips are meant to fill.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.