Goldman's Desk Called It Carnage

Goldman Sachs credit desks are openly warning that the AI rack bill has outrun the bond market's stomach — six hyperscalers issued $244 billion in debt in the first half of 2026, leverage doubled to 1.8x in six months, and traders say $25 billion of fresh supply now stresses what $75 billion used to.

The numbers that moved:

• Per The Wall Street Journal and Dealogic data cited across Goldman Sachs research coverage, Alphabet, Amazon, Meta, Oracle, Nvidia, and SpaceX collectively issued roughly $244 billion in bonds globally through mid-2026.
• That is more than double 2025's $108 billion total and roughly 14 times the $17 billion issued in 2024.
• Morgan Stanley data cited in market reporting shows aggregate hyperscaler leverage surging from 0.9x in Q3 2025 to 1.8x now — surpassing the entire energy sector's leverage in just over two quarters.

What Goldman is saying:

• Brian Garrett, head of equities on Goldman's Cross Asset Sales desk, told Goldman's Markets podcast on July 15 that hyperscalers "need to raise somewhere between $5.5 or $6 trillion" between 2025 and 2030 — and that they are "transitioning into asset heavy businesses" after decades as buyback machines.
• Garrett said Goldman's hyperscaler bond basket widened 22 basis points in a single week — "which doesn't sound like a lot, but it actually is."
• Credit strategist Jeffrey Papai, in a Goldman note reported July 13, said $75 billion of recent AI bond supply left the market "struggling to absorb" — and that the stress threshold has collapsed: where $75 billion once pressured spreads, now just $25 billion puts the market "on the defensive."
• Papai wrote that "incremental investor absorption capacity has significantly diminished."
• Goldman traders described conditions on credit desks as "brutal," per reporting on the note.

Why spreads are diverging:

• Brian Garrett flagged two blockers: the sheer size of issuance already in the pipeline, and concentration — "there's only so much tech you can own" if you run a diversified fixed-income book.
• Per MarketAxess data cited in WSJ reporting, Alphabet's 10-year spread widened 12 basis points last week. Meta's climbed 16 basis points. The average investment-grade spread moved only 2 basis points, per Bloomberg.
• Nvidia issued $25 billion in June. Amazon followed with another $25 billion in July — after a $37 billion sale in March, the fourth-largest corporate bond deal on record.
• SpaceX, a first-time issuer, tapped the market for $25 billion days after its IPO; its 10-year spread has jumped nearly half a percentage point since pricing, per WSJ.
• Cover ratios — orders relative to supply — fell from roughly 5x in February to about 2x by July, per CryptoBriefing's analysis of recent deals.

The cash-flow squeeze:

• Goldman equity research estimates the five largest hyperscalers face $5.8 trillion in combined AI capex from fiscal 2025 through 2030, per market reporting.
• Amanda Lyman, Goldman's head of credit strategy, noted only about $510 billion of additional debt capacity would unlock if hyperscaler leverage matched large U.S. banks — a fraction of what the buildout requires.
• Torsten Slok, Apollo's chief economist, warned that if AI productivity gains take five years instead of five months, equity markets already pricing immediate margin expansion face a painful repricing — with falling token prices and rising Chinese model share adding pressure on hyperscaler free cash flow.
• Michael Hartnett, Bank of America's chief investment strategist, flagged "AI capex cuts" as the market's largest tail risk — with a trigger chain running from bond vigilantes cutting liquidity to equity financing and layoffs.

What changed the narrative:

• In May, banks reportedly sought ways to spread data-center debt risk beyond core buyers as AI bond supply surged.
• Roughly 48 hours later, Goldman published "Decoding the Agent Economy," arguing agentic AI would lift LLM margins and reassure bondholders — fueling an even larger issuance wave.
• Papai now expects the AI bond complex to "further underperform the broader market" over the medium term and recommends credit-volatility hedges against "contamination" into wider investment-grade paper.
• AI-related issuance has grown from roughly 1% of the investment-grade market in 2024 to a projected 18% in H1 2026 — and 42% of long-duration paper exceeding 15 years, per Papai's note.

What to watch:

• Whether hyperscalers slow issuance, diversify maturities, or accept wider spreads as the price of admission.
• Q2 earnings from Amazon, Alphabet, Meta, and Microsoft — the cash flows underwriting every new bond.
• Kimi K3's open-weight drop July 27: cheaper inference compresses the revenue assumptions behind $244 billion of borrowing.

Convina's view: The AI trade just got a credit committee. For two years, enterprises watched hyperscalers spend like scarcity was permanent and bond investors cheered the yield. July 2026 is the month both sides admitted the math doesn't close in cash flow — only in debt. That doesn't mean AI demand is fake. It means the rack bill now competes with every other borrower for a finite pool of investment-grade appetite, and Goldman is pricing the friction in real time. If you're negotiating a multi-year AI contract, your vendor's cost of capital just became part of the quote — and the quote is getting wider every Monday.