Goldman collects the AI boom's fees and leaves the debt on other balance sheets

Goldman Sachs▲+0.38% — Goldman Sachs, up 0.38 percent today has turned the AI build-out into a fee line, and Tuesday's third-quarter report↗ put the number on it. Profit rose 29 percent on the strength of the financing desks and the best advisory quarter in the bank's history ¹. Revenue from equities financing, lending to investors against their positions, nearly doubled, though total revenue cooled from the second quarter's record ¹. The other half of the story is what the bank did with the debt behind those fees: it arranged the borrowing, then stepped out of the way of the risk.

The build-out arrives at the bank as fees

Follow the money's path. A new AI cloud company such as Volta Infrastructure, founded this year by two former Brookfield infrastructure executives, needs tens of thousands of Nvidia▲+0.74% — Nvidia, up 0.74 percent today GPUs and a data center to run them, and the AI lab AnthropicPVT$350B — Anthropic, private, latest valuation $350B has signed a six-year agreement for capacity at the Norwegian site ². That spending outruns a startup's cash, so Volta is borrowing $5 billion in the leveraged-loan market ².

The arranging banks take the fees. Chief executive David Solomon told investors this week↗ that financing activity has become a larger and more durable component of Goldman's markets businesses, and that demand for compute still significantly exceeds supply ³.

The arranger underwrote the safe slice and passed on the chips

Underwriting is the step where a bank commits to keep the paper if buyers do not appear, and Volta's structure shows the choice. JPMorgan Chase and Goldman underwrote only the $1.3 billion loan, which sits as cash collateral behind a data-center lease ². The $3.7 billion chip loan, the part that buys the GPUs, has no underwriter at all↗ ⁵.

The package sells as a strip, so a lender wanting the collateral-backed piece must also take the chip loan ². That debt is pitched at a yield of about 11 percent↗, among the highest of the AI boom, and it is unrated, which keeps out the collateralized loan obligations that usually buy leveraged loans ². Goldman's own strategists, having tallied the issuance, are underweight hyperscaler credit, the debt of the big cloud operators ⁶.

Higher rates press on the project's returns

Rates press on the far end of the chain. The Federal Reserve raised rates in September, its first hike in three years, and Goldman's economists expect another by December↗ ⁷. On the earnings call, Solomon said pressure on rates could affect the returns and viability of AI projects, and that Goldman remains highly selective about the AI risk it underwrites ³.

The bank's own book is the analysis

The arranging is genuine work and deserves its fee: turning a capital-spending need into a fundable structure is what makes the compute exist at all. It is also work the bank has chosen to do without holding the resulting debt. Where Goldman put its own capital instead is the tell. It spent $3 billion buying back its stock last quarter↗, at an average price about 12 percent above the market's ⁸.

Two tests land on Wednesday

The question for anyone sizing this boom is not whether the fees are real; it is whether the credit is worth holding when the bank that built the market in it will not. Two markers arrive this week. Morgan Stanley▼−1.52% — Morgan Stanley, down 1.52 percent today reports on Wednesday, October 14, and its equities desk will show whether the surge is industry-wide or Goldman's alone ⁸. The same day, commitments close on Volta's $3.7 billion chip loan, which still needs buyers at that yield ².