Goldman's fund desk is underweight the $420 billion AI debt wave Goldman itself forecasts

The desk that stepped aside

Lindsay Rosner, head of multi-sector fixed income investing at Goldman Sachs Asset Management, said the quiet part on television on September 24↗: "For that sector at large we are underweight knowing more issuance will come" ¹. Goldman's researchers tally what is coming: a record $420 billion of hyperscaler bond sales in 2027, marked up from roughly $400 billion in July and 60% above this year's pace ²³. One Goldman arm counts the wave. Another is declining to ride it.

The demand math moved first

This is not a default call. Portfolio managers are not worried the giants will fail to pay; the worry is the sheer volume and unpredictability of borrowing, and concentration ². BlackRock's Russell Brownback describes some deals as double-A credits pricing closer to triple-B spreads, investors lending to highly rated companies at compensation normally paid to weaker borrowers ².

Orders for new hyperscaler bonds fell from nearly five times deal size in February to below two in July, Apollo data show↗ ³. Median new-issue concessions rose to 12 basis points this year from 2.25 in 2025 ³. By late July, 78 of 91 hyperscaler bonds sold in 2026 were trading wider than they priced ³. AI issuers pay about 115 basis points over risk-free rates, against 78 for the wider investment-grade market ².

The tally keeps outrunning the forecast

Goldman's own tally keeps climbing. AI-related borrowing has already reached $489 billion this year, against the $322 billion the bank forecast ⁴. Morgan Stanley sees nearly $570 billion for the full year ⁴⁵. Gross leverage has doubled to about 1.8x in six months, with some $3 trillion of lease and purchase commitments off balance sheet ⁴.

The buyers of last resort

JPMorgan Asset Management's Stephanie Aliaga sits on the other side. The six largest hyperscalers could raise another $1.5 trillion, she says; their debt burden is well below the market average, and the top three are signing client contracts faster than they spend ⁶. Her math holds so far, with 2026 hyperscaler capex of about $750 billion still inside roughly $778 billion of operating cash flow ³.

The ECB's objection is for everyone else. US big tech is already just shy of 10% of this year's euro bond issuance by non-financial companies↗, and its borrowing could raise financing costs for issuers in unrelated sectors ⁷.

The Fed cannot referee this

Goldman vice chairman Rob Kaplan, on the bank's own podcast↗, notes the AI boom "is being done with debt and in some cases equity along the yield curve," borrowing a policy rate sitting at 3.75 to 4% never touches. He expects one more increase, to 4 or 4.25%, then a pause ⁸. Tightening lands on the lenders, not the builders.

The break-even revenue has not arrived

The strain lands on revenue that has not arrived. Goldman's break-even math wants roughly $300 billion of annual AI revenue; cloud sales are running about $70 billion above the pre-AI trend ⁹. The next marker is the FOMC on October 28, with the heaviest issuance still to be priced ¹⁰. Until then, an underweight looks like what Thornburg's Lon Erickson describes: skip the new issues, hold cash, buy wider later ². What the bank's underwriters earn selling the wave is broken out nowhere; that is the one number Goldman does not publish.