CoreWeave has spent 2026 borrowing against its chips. Its co-founder now wants to borrow against something larger: the balance sheets of the global insurance industry.
Brannin McBee, CoreWeave’s co-founder and chief development officer, told the Financial Times in an interview published Wednesday that the company wants more of its GPU-backed loans to reach investment-grade ratings, and eventually to lift its own rating, so that insurers become a standing source of capital. He described that pool as effectively unlimited.
The mechanics explain the ambition. Insurers are permitted to hold far more investment-grade paper than speculative debt, and their liabilities are long-dated, which makes multi-year data center leases a workable match for pensions and annuity books. A bank loan is a relationship with a limited appetite; insurance capital is a market measured in trillions of dollars.
CoreWeave has already tested the thesis. In March it closed an $8.5 billion facility secured by its Nvidia chips and a customer contract with Meta worth at least $19 billion. Morgan Stanley and MUFG co-structured the deal, anchored by Blackstone Credit & Insurance, and it carried an A3 rating from Moody’s and an A (low) from DBRS Morningstar — the first chip-backed financing of its kind to reach investment grade.
The rating is the product. It converts a specialty lending discussion into an allocation decision that any large fixed-income investor can make, and it is the reason insurers can participate at scale rather than on a one-off basis.
The scale of the funding already in place is sizable. CoreWeave has raised more than $35 billion this year through chip-backed loans, high-yield bonds, convertible notes and share sales, according to the Financial Times. Its debt and lease obligations totaled about $52 billion at the end of June.
The market’s own reading of that credit is less settled than the ratings suggest. The yield on CoreWeave’s bonds maturing in 2032 rose nearly 15 percent on Oct. 13, and the spread on five-year credit default swaps widened from about 368 basis points on Oct. 6 to roughly 643 basis points by the end of that week, according to quotes cited in the report.
A gap of that size between an investment-grade grade and a double-digit yield is the tension at the center of the plan. Ratings agencies assess the probability of repayment over a horizon using the collateral and the contracts behind it. Bondholders price the possibility that the collateral loses value faster than the loan amortizes.
That risk is specific here. A GPU is not a building. Its resale value depends on which model generations it can still train or serve, and the useful economic life assumed at underwriting is a judgment rather than an observable fact. The Meta contract and other customer agreements provide the cash flow that services the debt, which means the credit depends on both the machines and the customers continuing to need them.
The company’s revenue is concentrated in a small number of customers with very large needs, including Microsoft and OpenAI, which gives it visible cash flow on long contracts and little diversification if one of them reallocates spending. That concentration is the reason lenders ask for the customer agreements as collateral in the first place.
The macro backdrop adds a second variable. The International Monetary Fund warned this month that debt financing for artificial intelligence infrastructure has grown fast enough to matter for financial stability, and that a pullback in hyperscaler spending could leave excess data center capacity idle. Hyperscalers issued $121 billion of bonds in 2025 alone, roughly four times their average annual issuance from 2020 through 2024.
If insurance money does open up on the terms McBee describes, the model will be copied. Other GPU cloud operators pursuing the same build-out have the same incentive to package chips and contracts into rated vehicles, and insurers searching for duration will have to decide how much of an untested asset class they want in portfolios built to pay claims decades from now.
Regulators have begun asking that question in other contexts. Insurance supervisors look at capital charges and asset concentration, and a growing allocation to debt secured by depreciating hardware would show up in both.
Blackstone’s insurance arm has already anchored the March facility, which gives McBee a reference point to sell to other carriers. Whether a single anchor investor becomes a standing market is a different proposition.
None of it has to go wrong for the arrangement to be uncomfortable. The loans require sustained demand, stable chip values and customers that keep paying, and two of those three depend on a technology cycle that has moved in one direction for three years.
McBee’s pitch is that the capital exists and the assets are real. The rating agencies have agreed for one facility. What the insurance industry has not yet said is how much of a machine-heavy credit it is willing to hold, and at what price, if the AI build-out stops surprising people.
