Blackstone Asked What a GPU Even Was Before Lending CoreWeave $2.3 Billion. Chip-Backed Debt Is Now Heading Toward $1 Trillion

When Blackstone's tactical opportunities group first evaluated a loan proposal from an obscure AI cloud startup in 2023, the questions were basic. CoreWeave co-founder Brannin McBee has recalled that Blackstone executives asked, point blank, what a GPU was and whether AI was real. After the diligence was done, Blackstone led a $2.3 billion private credit loan to CoreWeave, the first major deal of its kind, and per a Bloomberg Businessweek Daily podcast published this week, the structure it pioneered is now on track to become a trillion-dollar market.

The innovation was less the money than the shape of it. Rather than lending against CoreWeave's balance sheet, the deal used a special purpose vehicle to buy and hold the chips themselves, with debt repaid from the cash flows of a single target customer contract. The loan was collateralized by CoreWeave's agreement with Microsoft, and the debt tenor was matched to the five to six year depreciation cycle of the hardware, locking obsolescence risk into a predictable window. McBee described it as the hardest diligence of his career, and the first deal carried a mid-double-digit interest rate.

The structure was borrowed, consciously, from commodities. CoreWeave's model resembled the liquefied natural gas playbook of signing an offtake contract first and financing the plant second. As CoreWeave's customer list grew, the cost of money collapsed. In March, the company raised $8.5 billion in debt backed by its Meta contract at roughly 6 percent, with Wall Street pricing the risk off Meta's investment-grade credit rather than CoreWeave's junk rating, per the Bloomberg reporting.

The scale of the financing need explains why credit markets got creative. Chips cost two to four times as much as the data centers that house them, while depreciating in five to six years. Take Nscale's flagship project in West Virginia as an example: the associated power plant is expected to cost about $6.5 billion, the data center and equipment about $18 billion, and the GPUs about $47 billion. Hyperscalers can fund that from cash flow and cheap bonds. OpenAI, Anthropic and the neoclouds cannot, which is why project finance mechanics have migrated into AI.

The template has since been scaled well past anyone's 2023 expectations. NVIDIA has announced a $500 billion cooperation framework with six major lenders, Broadcom is reported to be negotiating a roughly $60 billion chip financing arrangement for Anthropic, and SpaceX has been reported seeking $40 billion led by Apollo to bankroll a massive NVIDIA chip order. JPMorgan analysts project that total chip spending could exceed $3 trillion by 2030, with most of it debt-financed.

The risks are the mirror image of the returns. These deals are private, so investors lack an easy exit, and the exposure is concentrated along a single chain: chipmakers guarantee buybacks, hyperscalers sign the contracts, and neoclouds operate the hardware, so stress at any layer propagates through all of it. CoreWeave's own corporate-level junk bonds slid below 80 cents on the dollar earlier this month with yields above 15 percent, a market verdict on the gap between contracted revenue and the cost of servicing the stack.

There is also a subtler credit question that rating agencies are only starting to formalize: the same GPUs cannot be collateral twice. When the chipmaker guarantees the residual value, the cloud contract secures the SPV debt, and the corporate entity pledges everything else, the true seniority of any single claim is harder to establish than the investment-grade rating on the paper suggests. Bloomberg's framing, that chip-backed debt is eclipsing even the borrowings used to build the data centers housing the chips, is the cleanest statement of how much of the AI buildout now rests on this one structure.

None of this means the edifice is unsound. Contracted cash flows from Microsoft, Meta, Google and Amazon are real, long-dated and mostly investment-grade, which is precisely why lenders accept GPU collateral at 6 percent. But the 2023 Blackstone loan worked because one lender did painful primary diligence on one borrower. A trillion-dollar market running the same template at industrial speed, with residual values guaranteed by the vendors whose chips it finances, is a very different risk object, and 2027 will be the first year the five-year depreciation clocks start expiring in volume.