Three Banks Now Hold the AI Trade

The best business in the AI boom right now isn’t building models. It’s lending to the people betting on them.

Goldman Sachs, JPMorgan and Morgan Stanley have made a combined $52 billion in equity trading revenue this year, according to CNBC. That’s nearly 80% of what the top five banks earned from equity trading. The engine isn’t a clever new product. CNBC says the growth came from “insatiable” hedge fund borrowing for leveraged bets on AI hardware and Asian semiconductor supply chains. Looser leverage rules freed up the balance sheets to meet that demand. The International Monetary Fund, among others, has warned about rising leverage.

The headline treats this as a story about winners: three fortress banks taking share from weaker rivals. That’s accurate, and it’s the less important part. What matters more is where the risk ends up.

Start with the seat these banks are in. A prime broker gets paid when the fund borrows, not when the thesis turns out right. Goldman doesn’t need Nvidia’s customers to earn a return on their GPUs. It needs hedge funds to keep wanting more exposure than they can pay for in cash. In a gold rush you sell shovels. In a leveraged gold rush you sell the loan for the shovel, and you collect interest whether or not anyone finds gold. It’s the best seat in the house. That’s why Dimon, Solomon and Pick are fighting over it.

The strongest counterargument is in the CNBC piece itself. The Big Three’s huge balance sheets let them absorb possible losses and take on risks that push smaller rivals out. On that reading, concentration is good. If the leverage has to sit somewhere, it’s better with the firms holding the biggest cushions and the most experienced risk teams than with a mid-tier bank stretching to keep up. I take that seriously. Size really is protection.

But size protects against one kind of loss. A big balance sheet absorbs losses that hit one client at a time: a blown-up fund, a bad single-stock bet, a client that disappears. It does much less against losses that hit every client at once. When most of the book is levered to one theme, AI hardware plus the Asian chip supply chain behind it, the book isn’t diversified. It’s one trade spread across many tickers. Margin lending makes that worse. Prices fall, margin calls go out, funds sell to meet them, and prices fall further. The more of this business three firms hold, the more any unwind depends on three risk committees making the same call on the same morning. Concentration is the cushion and the single point of failure at once.

Then look at the gap between stated policy and outcome, because that’s where power actually shows up. Since 2008 the stated goal has been fewer banks that are too big to fail. In practice, leverage rules were loosened, and the banks with the most capital to deploy took the share smaller rivals gave up. Nobody planned that. It’s just what happens when capital is the binding constraint: the firm with the most capital gains most when the constraint relaxes. Every rule change shifts optionality to someone. This one shifted it to three firms.

The financing layer elsewhere in the AI build-out is starting to get picky, too. HSBC CEO Georges Elhedery told CNBC this week that AI is transformational and worth investing in, but that data center deals depend on the quality of the tenants and how the contracts are priced. McKinsey estimates global data center spending could reach $7 trillion by 2030. Mohamed El-Erian has pointed out that growing hyperscaler bond issuance is adding to the competition for financing. Put plainly, lenders are starting to separate good AI credit from bad. That separation usually shows up in credit before it shows up in equity.

I’m not calling a top. Markets are usually right, and I think the technology is real and compounding. But inflection points are where the alpha is, and I’m betting the next one in the AI trade shows up first at the margin desk, not on an earnings call. Watch who gets financed and on what terms. That will tell you more than another capex guidance raise.

Here are the stakes. If AI compounds the way I think it will, the technology gets through any unwind fine. The leveraged positions sitting on top of it don’t have to. Technology compounds. Leverage just speeds up whichever direction the price is already going.

The AI trade now has three shock absorbers, and shock absorbers only work until every wheel hits the same pothole.