AI Credit

State (current)

The financing layer of the AI buildout has become a separately priced risk, and it is largely invisible in the places allocators normally look for credit stress. The defining event so far is 14 August 2026: Bank of America analyst Tom Curcuruto estimated that Broadcom's chip-financing vehicle could reach $370B of senior debt by mid-2029 at 20-gigawatt scale, including roughly $150B of new issuance in 2027 alone, and Broadcom shares fell about 6% on a day with no company news. The structure matters more than the number. Per Broadcom's latest 10-Q an investor partner assumed the purchase and lease agreements; Broadcom agreed to backstop lease payments for five years with maximum exposure up to $29B on the initial transaction. So the $370B is not Broadcom's debt. It is capital raised by a vehicle that buys custom AI accelerators and leases them to customers, with a vendor backstop sitting behind part of it. The vehicle launched in June when Apollo and Blackstone led a $35B financing of Broadcom's AI XPV Platform, funding more than 1 gigawatt of compute for Anthropic and sized to support more than 20 gigawatts for frontier AI labs through 2028; Blackstone has since sounded out investors on a further transaction exceeding $30B.

Broadcom is not the only one. NVIDIA announced in the same week a framework to collectively finance roughly $500B of AI computing deals alongside Goldman Sachs, Blackstone and Apollo, with KKR, BlackRock and Brookfield added days earlier; Jensen Huang said NVIDIA's support would cover as much as 25% of an opportunity and no deals were signed at announcement. NVIDIA did not sell off on 14 August. The difference between the two reactions was not structure, it was disclosure: one had a number attached that day. A Goldman Sachs note cited the same session says hyperscalers are using lease commitments and other debt structures to fund expansion, which makes this a theme-wide financing pattern rather than a merchant-vendor quirk.

Why this is its own entity and not a line in ai-capex: the leverage is accumulating off the balance sheets that investors read and outside the indicators that credit investors watch. Investment-grade and high-yield spreads and single-name CDS see funded debt at the issuer. They do not see a special-purpose vehicle's senior debt or a five-year vendor lease backstop. That means repricing in this theme arrives as a step function, on the day someone quantifies it, rather than as a gradual widening. Two channels connect it back to the macro regime: the vehicles are being funded by private credit and alternative managers (Apollo, Blackstone, KKR, Brookfield), so the risk sits in a part of the market that marks slowly; and their cost of capital is a function of the long end, where the US just auctioned 30-year bonds at 5.216%, described as the highest in 25 years.

Timeline

• 2026-08-15: The theme was priced for the first time. BofA (Tom Curcuruto) estimated Broadcom's chip-financing vehicle could reach $370B senior debt by mid-2029 at 20GW scale, incl ~$150B new issuance in 2027; AVGO -5.94% to $392.99 with no company news. Disclosed structure per AVGO 10-Q: investor partner assumed the purchase/lease agreements, AVGO backstops lease payments five years, max exposure up to $29B on the initial transaction. Vehicle launched June 2026, Apollo + Blackstone led a $35B financing of the AI XPV Platform, >1GW for Anthropic, sized for >20GW through 2028; Blackstone sounding out investors on a further >$30B transaction. NVDA's structurally similar ~$500B framework (Goldman, Blackstone, Apollo; KKR, BlackRock, Brookfield added days earlier; up to 25% of an opportunity, no deals signed at announcement) did not sell off. Goldman note cited: hyperscalers using lease commitments and other debt structures to fund expansion (https://247wallst.com/investing/2026/08/14/broadcom-sinks-6-as-bofa-flags-370b-in-ai-debt-amd-climbs-4-on-bairds-1250-call/ ; https://finance.yahoo.com/markets/stocks/articles/stock-market-today-aug-14-213018651.html)
• 2026-08-15 (context, carried from ai-capex): the long end is repricing at the same time. The Treasury sold 30-year bonds at 5.216%, described as the highest in 25 years, attributed to inflation, energy prices and the deficit (https://finance.yahoo.com/markets/stocks/articles/stock-market-today-aug-14-213018651.html)
• 2026-08-07 (carried from ai-capex): Oracle CDS hit a record, above its 2008 peak; Celestica raised $3.0B; cumulative AI spend ~$2.0T with depreciation schedules extended from 24 to 60+ months. This was the first visible credit-market signal in the theme.
• 2026-08-05 (carried from ai-capex): Oracle default risk record tops the 2008 peak; Apollo argued 60/40 is broken.

Our calls (journaled)

• 2026-08-15: This is a disclosure risk before it is a credit risk. The debt in question is real, but the repricing is triggered by quantification, not by deterioration, which means it arrives suddenly and in one name at a time. For an allocator the practical consequences are three. First, spreads and CDS are the wrong dashboard for this theme; the right questions are in filings, on lease commitments, vendor backstops and off-balance-sheet vehicles. Second, a vendor that finances its own customers is booking revenue against a contingent liability, so revenue growth and credit exposure are the same line read twice. Third, this concentrates risk in private credit and alternative managers who mark slowly, so the absence of visible stress is not evidence of its absence. Invalidation: the next leg of the buildout being funded from operating cash flow with these vehicles shrinking, or full disclosure arriving without a share-price reaction, which would mean it is already priced. Watch whether Broadcom's disclosed $29B backstop grows toward the platform's 20GW scale, whether NVIDIA's ~$500B framework converts into signed contracts, and whether any of these vehicles' debt starts trading publicly. Outcome: open

Hit-rate

• pending (entity created 2026-08-15)