Who gets paid when AI crashes?

Happy Saturday!

Brian McGleenon here, Global Head of News at BeInCrypto.

July was wild. Major chip stocks dropped nearly 28% from June highs, while major international exchanges hit daily volatility limits. Days later, big tech bounced hundreds of billions of dollars in market cap practically overnight.

None of this was fundamentally about whether AI works as a technology. It was pure market leverage. When debt calls come, forced liquidations can wipe out portfolios, leaving buyers to rush back in the second the mechanical selling clears.

No one has a crystal ball for how things will play out this August. But instead of guessing which single semiconductor stock wins the next round, there is an alternative lens: looking at the financial institutions that service the capital flow either way.

Here is the structural case often made for institutional financiers like Goldman Sachs ($GS).

Goldman doesn’t make GPUs, build hyper-scale data centers, or train foundation models. It arranges the capital, and earns fees when money moves in either direction.

Think of an auction house. They take a fee when a rare collection is assembled at record valuations, and they take a fee again when assets are restructured or liquidated. Same gavel, same percentage, either direction.

• During the Boom Phase: Wall Street banks get paid to underwrite and raise capital. Hyperscalers and tech firms continue to deploy massive capital expenditure on AI infrastructure, often requiring capital markets to organize bond and equity issuances. Wall Street earns advisory and underwriting fees whether Semiconductor Brand A stays on top, Brand B takes market share, or cloud providers manufacture custom silicon internally.

• During a Market Drawdown: Financial institutions like Goldman Sachs get paid to facilitate restructuring, mergers, acquisitions, and private debt financing as volatile valuations spark consolidation.

Goldman reported on 14 July, two weeks before the worst of the selloff:

• Record earnings of $20.98 per share, against $10.91 a year earlier

• Return on equity of 23.5%

• $3.40bn in fees from arranging deals, up 55% on last year

• $1.03bn from bond sales, its best quarter ever

• $985m from share sales, up 130%

• $1.2 trillion of merger deals advised on in six months, roughly $425bn more than its closest rival

• A pipeline of unbilled work still at a five-year high, even after all of the above

So the boom half of my argument is settled; the money has already been banked.

The bust half is still just a claim, and I’ll show you where it’s weakest before the end of this letter.

First, the month that put it to the test.

Memory went first. SK Hynix fell a record 15% in Seoul on 13 July, pulling the Kospi down 9%.

Then, on 27 July, Nvidia dropped 4.99% to $196.51 and handed the world’s-most-valuable-company title back to Apple.

The real signal was in credit, not equity. Nvidia’s five-year CDS spread jumped to a record 82 basis points intraday, the biggest single-day move since those contracts began trading actively.

That is nowhere near distress. The velocity was the point. Credit desks had treated Nvidia as boring, and they stopped.

Korea then broke properly, with circuit breakers on two consecutive sessions, a first, after SK Hynix posted its most profitable quarter ever and still missed what the market demanded.

ChangXin Memory’s $8.6bn DRAM-expansion IPO and China’s move into domestic lithography landed in the same 48 hours. The Philadelphia Semiconductor Index closed the month 28.6% below its 22 June peak.

Market sentiment leading into the Fed meeting was heavily weighted by hawkish positioning anyway, with CME’s FedWatch reflecting a one-in-three chance of a surprise rate decision.

The Fed held at 3.50%–3.75% on a 9-3 vote, with three regional presidents dissenting in favor of a hike, the most one-directional dissents since 2016.

Markets sold off anyway. The Dow fell 1,153 points, its worst day since April 2025. The 30-year yield hit its highest since 2007.

Kevin Warsh, nine weeks into the job, declined to signpost what comes next.

So the hike risk is not a debunked rumor. It is a live expectation, with Jackson Hole mid-month and the next decision on 16 September.

AMD reports after the close on 4 August. Consensus is around $1.61 against $0.48 a year ago, on revenue near $11.30bn, above the midpoint of its own $11.2bn ± $300m guidance. Options price a move of roughly 12% either way.

The long-term story is not in question. At Advancing AI on 23 July, Lisa Su raised the 2030 accelerator TAM to $1.4 trillion and server CPUs to above $200bn, citing agentic workloads.

The broader thesis makes sense. Still, holding single memory or GPU names through August requires navigating sharp downside risks. July demonstrated how easily these assets can drop 30% to 50% from external liquidity pressure or macro filings, independent of core operational performance.

• It is not defensive, and it did not sit out July: GS fell 5.1% on 29 July to $980.75, down 10.7% on the week, leading banks lower on AI concerns. If AI equities crater in August, this falls too. My argument is about the durability of the cash flow through the cycle, not the share price being insulated from it.

• It is not neutral: As a premier broker, Goldman services extensive institutional leverage across key market participants and ranks top-tier in leveraged lending, with a supplementary leverage ratio of 4.3%, the thinnest among peers. Extending credit into a leveraged market is distinct from simply tolling transactions.

• It is not cheap: Roughly 3.1x tangible book, near record levels, with Sell downgrades arguing this is a cyclical peak rather than a new baseline.

Restructuring fees have historically not been big enough to offset a major collapse in underwriting. In 2008 and 2022, both fell together. The two-sided model cushions a slowdown; it does not neutralize a credit event.

CEO David Solomon remains more cautious than the commentary surrounding him. He noted that the tech build-out is in its early stages while warning the path may be uneven, noting that enterprise AI deployment is proving slower and more complex than market expectations.

This is not an argument that institutional banking stocks preserve wealth during a market drop; nothing with high beta shields wealth during a broad drawdown.

Rather, for those seeking exposure to a multi-trillion-dollar capital cycle without concentrating risk on single-product suppliers exposed to sudden DRAM supply shifts, evaluating a fee engine that captures upside on issuance and downside on restructuring offers a different risk profile.

That is a relative structural thesis. Whether $980 represents the right valuation entry into a potential earnings cycle peak, against three dissenting Fed votes, remains a separate market decision.

Size position risk accordingly. August presents multiple binary events ahead, alongside central bank leadership that has explicitly declined to offer forward guidance.

This Week’s Featured Watch:

Two of the sharpest minds in macro and derivatives, Iggy Ioppe (CIO at Theo, ex-Credit Suisse) and Gordon Grant (Head of Derivatives at Bitwise), debate how to hedge uncertainty as the 40-year stock-bond playbook fails, whether Bitcoin is still “digital gold,” and why the market is miscalculating AI capex.

Have a great weekend,

Brian McGleenon

Global Head of News, BeInCrypto