Financing the Machine: Bank Profits and the AI Build-Out

Where the money came from depends on where you look. Net interest income, the rate-sensitive spread business, grew only in the mid single digits and is now a mild drag as the Fed eases. The quarter was made elsewhere: in trading and in underwriting, and a large share of the underwriting was financing one thing, the AI build-out. The banks' profits have quietly become tied to data-centre capex. This piece works out how that dependence is built, and whether the credit markets that fund the build-out give an early read on the cycle turning.

The beat was capital markets, not lending

Rank the five banks that reported by profit growth and the order is almost exactly their exposure to capital markets.

* JPMorgan's headline was $21.2bn and +41%. It included $5.6bn of pretax one-off gains, mostly on Visa shares. Excluding significant items, net income was $16.9bn, up roughly 13%; we use that figure.

Goldman sits at the top with almost no ordinary lending book. Wells Fargo sits at the bottom as the most deposit-and-loan bank of the group. The spread between them, +78% against +17%, is the whole argument in one line. The bank most tied to plain lending grew the slowest. The bank that is pure capital markets grew the fastest.

Inside the capital-markets line, two things ran hot. Trading was the first. Equity trading revenue rose 86% at JPMorgan, 72% at Goldman, 70% at Bank of America. Prime-brokerage balances at Citi climbed around 60%, which is hedge-fund leverage flowing into the same names. The second was underwriting and advice. Goldman's equity underwriting rose 130%, its debt underwriting 75%, and announced large-cap M&A volumes ran about 90% ahead of last year. These are the fee lines that only move when companies are issuing securities and doing deals. In the June quarter they moved a lot, and the question is what they were issuing for.

What the banks are underwriting for

The AI build-out is spending roughly $725bn this year across the big hyperscalers, about three quarters of it on AI infrastructure, up about 77% on last year. For most of the past decade that spending was funded from internal cash flow, and internal cash pays no fees to a bank. In “The AI Capex Machine” and its sequel we tracked the point where that stops being true. On a trailing-twelve-month basis Amazon now runs capex at 102% of operating cash flow and Oracle at 174%, both with free cash flow negative. Once capex passes cash flow, the marginal dollar of spending has to be financed outside the company. That is the dollar the banks get paid on.

The financing arrived in size. The five most active hyperscalers sold $121bn of bonds in 2025, more than four times their 2020–24 average. Meta printed a single $30bn deal, the largest investment-grade bond ever done outside an acquisition. Amazon followed with about $54bn. Bank of America now expects $175bn of hyperscaler issuance for the full year, and the Street models roughly $300bn of AI and data-centre bonds over the next twelve months.

Below the investment-grade names, a second financing layer has emerged that barely existed a few years ago. Data-centre debt issuance roughly doubled to $182bn. Private-credit loans to AI companies went from near zero to more than $200bn in a few years. Meta financed a two-gigawatt campus in Louisiana through a $27bn special-purpose vehicle arranged with Blue Owl and Pimco. The neocloud operators that rent out GPUs, CoreWeave and its peers, borrow in the high-yield and private markets at rates around 11% against the chips themselves as collateral.

Every one of those flows pays a bank. Debt underwriting on the investment-grade bonds. Structuring and warehouse fees on the private credit and the securitisations. Leveraged-finance fees and bridge loans on the neoclouds. Equity underwriting when an AI-adjacent name comes to market. Advisory when two of them combine. That financing layer runs through the banks, which is why Goldman's 75% jump in debt underwriting and 130% jump in equity underwriting are not a general capital-markets recovery. They are, in large part, the AI build-out passing through the income statement of a bank.

The seductive idea

If bank capital-markets revenue rides the capex cycle, it should be possible to run the logic backwards. Financing conditions set the cost of the marginal capex dollar. When credit is cheap and the issuance window is open, the spending continues. When spreads widen and the window narrows, the spending gets trimmed, and the companies that sell into the build-out see it in their order books a quarter or two later. The chain looks like this:

financing conditions → capex plans → supplier revenue → analyst estimates → share prices

Credit sits at the front of that chain. Analyst estimate cuts sit near the back. In the sequel piece we noted that not a single AI name has had next-year earnings cut yet, while the share prices have already repriced most of them. If credit leads the whole chain, then watching spreads and the issuance window should give an early read on the capex top, ahead of the estimates and perhaps the share prices themselves. It is a clean idea. We built the series to test it, and it did not survive.

What the data said

We proxied the financing conditions with liquid instruments that have a long history: investment-grade and high-yield credit against Treasuries, the relative strength of recently-listed companies for the issuance window, and a basket of the listed neoclouds against the semiconductor complex. The target was the semiconductor index, the cleanest daily proxy for the AI-capex theme. The test was simple. Does the state of financing conditions today tell you anything useful about where the AI complex goes next. We ran it back to 2015, which includes the late-2018 credit scare, the 2020 shock and the 2022 tightening.

Two results, both against the thesis.

First, changes in credit move at the same time as the AI complex, not ahead of it. The day-to-day correlation between credit risk appetite and the semiconductor index is about +0.36 at zero lag and essentially zero at any positive lead. Credit does not front-run the complex. The two reprice together, and if anything credit lags by a day or two.

Second, the level of credit stress points the opposite way from a warning. Wide spreads have been followed by stronger returns in the AI complex, and the effect grows with horizon. Sixty trading days after credit sits in its most-stressed third, the semiconductor index has returned about 13% on average, against under 5% from its easiest third. This is the oldest fact in credit. Spreads are tightest when everyone is comfortable, which is near the top, and widest when everyone is frightened, which is historically near the bottom. Stress has read as a contrarian marker rather than a lead.

So the financing layer confirms the cycle. It does not lead it. A bank's capital-markets revenue is a real-time gauge of how the build-out is being funded, and a good one. It is not an early-warning system for the funding drying up. The early warning still has to come from the place it always comes from, which is the fundamentals: the first hyperscaler to soften its capex language, the first supplier whose next-year numbers get marked down. Oracle, the most leveraged spender and the one the sequel piece already flagged, is the name to watch for both.

Where the stress actually shows

The composite reading today is benign. Broad high-yield is calm and the issuance window is open. That is consistent with a build-out still accelerating and no capex cut expected before late this year at the earliest. But two narrower gauges are not calm, and they are the two the thesis would have pointed to.

The first is Oracle against Microsoft. One is rated BBB−, one notch from junk after S&P's July downgrade, and is funding its build-out with debt. The other is rated AAA and is funding its own from cash. Their shares have pulled sharply apart, which is the market pricing the leverage rather than the business. The second is the neocloud basket against the chips it runs on. Those operators sit on the most fragile funding in the whole structure, and they are the first to wobble when lenders get cautious. Blue Owl's attempt to raise $4bn for a CoreWeave project stalled this year when lenders balked at the credit. Broad credit says nothing is wrong. The AI-specific credit says the weakest points are being repriced. That is exactly the order in which these things start.

It matters because the banks are not only earning the fees. They are holding the risk. The build-out has a circular quality that rhymes with the last one. Nvidia holds a stake in CoreWeave and has committed up to $100bn to OpenAI, money that returns to Nvidia as GPU orders. This is vendor financing, the same mechanism that flattered Nortel and Lucent into the 2001 telecom bust, when the equipment makers lent their customers the money to buy the equipment. The banks are the other vendor-financing channel this time. They lend to a build-out whose revenue depends on the build-out continuing. In an upswing that is a fee engine. In a downswing it reverses twice over. The fee income falls as issuance stops, and the loans already made start to sour. A capex slowdown would hit the capital-markets banks on both sides of the ledger at once. That is why a downturn would bite harder than the placid quarterly numbers suggest.

What it means

The banks' capital-markets profits are a coincident gauge of the AI financing cycle. They will most likely stay strong through the capex peak rather than roll over ahead of it. The reason is mechanical. Financing need is largest late in the cycle, when the spenders have exhausted their own cash and must borrow for every additional dollar. Bank AI revenue peaks when the build-out is most stretched, making it a late-cycle signal.

So the banks read the cycle rather than call it. Their earnings confirm how the build-out is being funded and how much risk that funding carries. The turn, when it comes, will surface first in the spenders themselves, in the same capex language and supplier estimates the two prior pieces watch. The banks will register it afterwards, in the same lines that are booking records now.

The banks made themselves part of the AI build-out, deliberately, for the fees it pays. The June quarter is what that choice looks like on the way up. The way down runs through the same lines. The fees fade as issuance stops, and the credit already extended to the build-out stays on the books.

Method notes

• Bank figures: Q2 2026 releases (14–15 Jul 2026) for JPM, GS, C, BAC, WFC. Operating adjustments noted in-text (the JPMorgan one-off gains). Segment growth rates as reported by each bank. Morgan Stanley reported on the same cycle and is broadly consistent; it is omitted from the table for space.
• Capex, cash-flow ratios and the depreciation context are carried from the two prior pieces; source model ai_earnings_peak/model.py, yfinance quarterly financials on a trailing-twelve-month basis.
• Financing-conditions proxies (yfinance, daily): high-yield HYG versus Treasuries IEF and versus investment-grade LQD; IG LQD versus IEF; the Renaissance IPO ETF versus SPY for the issuance window; an equal-weight basket of listed neoclouds (CRWV, NBIS, APLD, IREN, WULF, CIFR, CORZ, HUT) against the semiconductor index SMH; Oracle versus Microsoft for credit-quality discrimination. The composite is a weighted z-score, signed so positive means tightening.
• Event study, two lenses. (1) Cross-correlation of daily proxy returns against forward SMH returns at lags out to 20 days. (2) The stress level against forward SMH returns at 5, 10, 20, 40 and 60 days, with a top/bottom-tertile split. The credit-only composite is run on the full 2015–2026 sample (2,764 days); the neocloud proxies only exist for about the last two years and are excluded from the long-horizon test.
• Key results: contemporaneous credit-versus-SMH correlation +0.36, best positive-lag correlation −0.11 at one day. Level test: correlation rises from +0.09 at 5 days to +0.28 at 60 days, all positive; top stress tertile forward-60-day return ~13% versus ~5% for the bottom tertile. The lead the thesis assumed is not present, and the level effect is contrarian.
• Financing-flow figures (hyperscaler and private-credit issuance, neocloud terms, securitisation volumes) are from market reporting and sell-side estimates, not primary filings, and are cited as orders of magnitude. The AI-linked share of each bank's underwriting is an estimate from deal commentary; banks do not disclose it.
• Limits: the long test spans several credit cycles but only one AI-capex cycle, which has not yet turned. The contrarian level result is robust across cycles. The absence of a lead is what the current data can support; it is not proof that credit could never lead a genuinely AI-driven funding stop. Company names illustrate the framework; nothing here is a recommendation.

Not investment advice.