The AI Buildout Has Become a Circular Credit Machine

Jensen Huang named the problem himself last year, describing what happens to the previous generation of chips when a new one ships.

When Blackwell starts shipping in volume, you couldn't give Hoppers away. — Jensen Huang

That is the chipmaker's own admission that the hardware at the center of the AI credit machine depreciates on roughly a two-year cycle. The loans backed by those chips run for six. Nvidia has built a multi-billion-dollar investment portfolio that includes a $3.96 billion stake in CoreWeave, one of its biggest GPU customers, plus positions in Arm, Applied Digital, Nebius, and 27 private startups. ^[1] It proposed a $100 billion partnership with OpenAI structured so that OpenAI pays for GPUs in cash and Nvidia reinvests that cash back into OpenAI as equity. ^[2] The arrangement drew a question from short seller Jim Chanos that has not been answered.

[Don't] you think it's a bit odd that when the narrative is 'demand for compute is infinite', the sellers keep subsidizing the buyers? — Jim Chanos

The customers use that capital to buy Nvidia GPUs. Then Wall Street repackages the GPU-backed contracts as debt. In March, CoreWeave secured an $8.5 billion loan rated A3 by Moody's, the first investment-grade deal collateralized by GPUs, arranged by Morgan Stanley and MUFG with Blackstone Credit as anchor investor; the yield was 7.5 percentage points lower than its 2023 financing. The rating rested substantially on a $19.2 billion contract backlog from Meta, which runs through 2032. ^[3] Then Meta became the competitor. In July, the company launched Meta Compute to sell excess GPU capacity to external customers. ^[4] CoreWeave's stock fell 50% from its 52-week high. ^[5] The same customer whose contract secured the investment-grade debt was now building a competing cloud, and the borrower whose balance sheet backed the loan was suddenly half its former value. Apollo Chief Economist Torsten Slok identified the structural dynamic beneath this sequence. The silicon and equipment companies maintain 41% profit margins. The model developers operate at negative 59% margins. ^[6]

AI boom’s profits are currently being funded by investors rather than earned from customers. — Torsten Slok

The upstream profits are not funded by end demand. They are funded by capital raised at the layer that is losing money. The entire chain depends on model developers continuing to attract investment at negative margins, which they then pass upward to the chipmaker and outward to the debt arrangers. The debt does not disappear. Citadel Securities forecasts more than $500 billion in new debt by 2028 specifically to finance AI chips, with hyperscalers having already issued roughly $570 billion. ^[7] JPMorgan raised its 2026 tech debt forecast to $540 billion and identified chip-backed financing as a frontier that could reach trillions by 2030. ^[8] The paper lands somewhere: with pension funds, insurers, and credit investors who bought it as infrastructure-grade. Andromeda Capital Management released a report this week that named the asymmetry directly, describing the setup as "heads I win, tails you lose" for credit investors who accept record-tight spreads and capped upside while facing uncertain returns. ^[9] The collateral itself ages faster than the debt it secures. Huang's own product roadmap makes this explicit: annual chip releases that render the previous generation nearly worthless. Short seller Michael Burry has argued that companies are overstating chip useful life at six years when the real figure is closer to two or three. ^[10] The loans, meanwhile, are structured for six. A pension fund holding A3-rated paper backed by GPUs that will be obsolete before the loan matures, secured by contracts with a company now competing with the borrower. That is not infrastructure risk. It was sold as though it were.