AI capital markets
Hyperscalers spent from cash flow until they couldn’t. The next dollars route through Wall Street — and the same issuance wave that pays the banks is what the literature flags as a late-cycle warning.
For years the AI build-out was self-funded. Google, Amazon, Meta, Microsoft — they paid for chips and data centers out of operations. That ended toward the end of 2025. Oracle borrowed. Meta followed. Alphabet sold stock after a decade of buybacks. Michael Cembalest, JPMorgan Asset Management’s chair of market strategy, put the shift plainly: the free-cash-flow projections for the hyperscalers are heading very low. The question is no longer whether AI needs more capital. It is who intermediates the raise once internal cash is gone — and whether the wave of deals that pays the intermediaries is also the tell that sentiment has peaked.
David Solomon, Goldman Sachs’s CEO, framed the demand on the bank’s July 2026 earnings call as a three-to-five-year infrastructure cycle whose financing needs are “outstripping what we think is the appropriate quantum.” The numbers on the same call bear him out. Denis Coleman, Goldman’s co-head of global banking and markets, reported record $59 billion of fundraising in the second quarter — $31 billion in private credit alone — and raised the full-year alternatives target to more than $125 billion. Equity financing was also a record, up 91% year over year; financing revenues of $4.5 billion rose 62%. Investment-banking backlog sits at its highest level in five years; Goldman became the first bank to cross $1 trillion of announced M&A volumes in a half-year.
We’re working our way down the capital curve. We started with free cash flow. The speed with which the tech companies blew through the debt markets is kind of incredible. It took like a year. And now Google’s issuing equity. What’s after that?
Ben Thompson, August 2026
The bank at the toll booth
Goldman is the cleanest listed expression of the intermediary seat. Record net revenues of $20.3 billion, earnings per share of $20.98, and a return on tangible equity of 25.5% in the second quarter are the realized print. The re-rate — whether the market awards a higher multiple on the financing franchise — is still forward-looking and capped. Goldman’s supplementary leverage ratio is 4.3%, the lowest among peers; Coleman called it “a governing constraint on growing the financing business.” You can deploy $31 billion of private credit in a quarter and still hit a balance-sheet ceiling.
The private-credit boom sits beside a retail stress that the wiki records but does not resolve. Four large retail BDC vehicles hit redemption caps in the same quarter Goldman deployed a record $31 billion institutionally. Goldman Sachs Private Credit Corp received repurchase requests of roughly 3.24% of shares in the second quarter — below its 5% cap, so all were fulfilled — while BCRED, Apollo ADS, and Blue Owl saw requests of 10% to 40%, all capped at 5%. Either Goldman’s institutional origination is a genuine edge or it is late-cycle risk accumulation. Management offered no portfolio-default disclosure on the call. The tension is narrowed, not closed.
Debt as a share of hyperscaler capex
From the stock-market wiki’s mega-issuance chain. Big-four capex exceeded $700 billion in 2026, up 77% year over year, with more than $1 trillion projected for 2027.
The issuance peak nobody wants to call
The financing supercycle has a mirror image. The simultaneous mega-raise cluster — SpaceX, OpenAI, Anthropic — may surpass the 1998–2000 IPO-mania haul, Michael Batnik said on The Compound in May 2026: “if this is a top, this would be with the benefit of hindsight the most obvious top we’ve ever seen.” Peer-reviewed work backs the contrarian read, not the timing. Baker and Wurgler showed that the equity share in new issues predicts lower subsequent market returns; Arif and Lee found aggregate investment peaks in high-sentiment periods and is followed by lower equity returns and earnings disappointments. Equity issuance is a warning, not confirmation — but Butler and colleagues argue the predictability may be a statistical artifact, and the lead time from peak to de-rate is imprecise.
The credit market is where the warning is becoming machine-checkable. Morgan Stanley and JPMorgan estimate the sector needs roughly $1.5 trillion of new debt; about $100 billion of AI-capex bonds had already printed in 2026, with investors demanding record credit-default-swap protection. S&P downgraded Oracle from BBB to BBB- in July 2026, citing capex and cash-flow concerns — Oracle’s 2026 capex is approximately twice operating cash flow — while Microsoft, FCF-positive with capex held steady, was untouched. Bank of America’s Michael Batnik cited a path for hyperscaler US investment-grade index debt from $288 billion at year-end 2025 to $659 billion by year-end 2027. Hyperscaler credit spreads were widening by mid-July, and equity issuance was substituting for debt — Amazon doing an at-the-market program, Google reportedly considering one.
Dylan Patel, on Dwarkesh in late August, put numbers on the same curve. OpenAI and Anthropic take as much as 40 to 50 percent of compute next year. To get toward 100 gigawatts by 2028 they have to start paying 25, 30, 50 million dollars a megawatt. Three to four trillion of implied capex is not funded from operating cash — hyperscalers spend everything on capex and then raise debt. SpaceX, he said, will lease quite a bit of new compute to the two labs. Same spine as Thompson. It does not graduate the derate. Josh Brown, relaying a former Fed chair’s 2027 setup, said the increase in AI investment, not the level, is what prices growth — “that’s without the bubble bursting.” Color on a second derivative. Not a print.
Ed Zitron, on the same Compound episode as Brown at the end of August, named the concentration another way. Nvidia’s latest quarter was 16 percent from one customer; three customers were 44 percent of the first half of fiscal 2027; five were 70 percent. OpenAI, he said, is about 60 percent of AI data-center demand, and without it there is not another OpenAI-sized spender except through venture capital. Brown, on that tape, had Nvidia guiding about 70 percent revenue growth with about 90 percent of the business data-center sales to the same four or five hyperscalers. Same issuance and concentration spine. Color. It does not graduate the derate, and it is not a new infrastructure name.
Santoli, on an earlier Compound tape, put the same concentration in a sentence about capitalism. Everyone all at once spending free cash flow on the same thing, to build the same thing, and everyone getting a great return — “I still think it’s a little bit of a threat.” He hedged immediately: “maybe that’s not today.” Color. It does not graduate the derate.
Dan Skelly, a Morgan Stanley Wealth Management portfolio manager, sat on The Compound four days later and named the second-derivative risk the wiki already tracks. Servers, chips, electrical gear, industrial components — ordered ahead of forty or fifty gigawatts of projected data-center construction in the next two to three years. “Are all of those buildings going to get built?” Double-ordering in the supply chain of semis and servers is real. “Is it a risk today? No. But is a risk from here for now? Could be.” Same episode: his hedge-fund clients have not re-grossed in memory or in the first-half AI-capex winners; they have been waiting. Through June, he said, ninety percent of the S&P’s attribution came from three industry subgroups — semis, IT hardware, and power — all tethered to data centers. Practitioner color. It does not graduate the derate. It is not a new infrastructure name.
Brad Gerstner, on All-In in mid-September, pushed the same curve from the other side of the table. Patel’s 43 gigawatts of add next year — about 14 of that to the leading labs — is “too aggressive.” Altimeter’s stand-up is “somewhere closer to 25 gigawatts,” half of it for Anthropic and OpenAI. The labs, he said, still need at least $180 billion of offtake by year-end “just to keep the AI trade intact.” Semiconductors were 70 percent of the Nasdaq’s return. Same second-derivative spine as Skelly’s 40-to-50-gigawatt double-order risk. Color. It does not graduate the derate.