Insurance funds become new sponsors for AI computing power: CoreWeave (CRWV.US) and other AI infrastructure financing collectively turn to insurance capital pools

According to Zhitong Finance APP, CoreWeave (CRWV.US) is seeking to tap into the massive investment pools of insurance companies to finance its AI infrastructure expansion. Co-founder Brannin McBee plans to secure investment-grade ratings for more debt backed by chips, and even for the company itself.

When discussing the funds available from insurance companies, McBee said, "It's unlimited. It's like all the money lying in the insurance fund accounts in the world."

This year, CoreWeave has raised over $35 billion through chip-backed loans, high-yield bonds, convertible bonds, and stock offerings. As of the end of June, its total debt and lease obligations were about $52 billion.

An $8.5 billion loan, collateralized by computing infrastructure and a contract with Meta Platforms (META.US), received an investment-grade rating. CoreWeave is now seeking to expand this financing method to debt backed by contracts with lower-rated AI labs.

McBee stated that almost all of CoreWeave's computing capacity for 2027 has already been sold, and the company has also signed contracts extending into 2028 and beyond.

The focus of AI infrastructure financing is shifting to insurance capital pools

CoreWeave's financing pursuit is not unique—it is a rapidly replicated industry model. On August 10 this year, NVIDIA announced the signing of memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, aiming to leverage over $500 billion in third-party capital to build financing platforms dedicated to computing power assets.

The primary sources of capital are the general accounts and long-term mandate funds of insurance companies: Apollo owns annuity insurer Athene, KKR controls Global Atlantic, Brookfield has a wealth solutions division, while PineBridge and BlackRock manage large-scale insurance portfolios.

These platforms lend downstream, mainly to emerging cloud service providers like CoreWeave, Crusoe, and Nebius, as well as project companies building data centers. Borrowers use the funds to purchase NVIDIA hardware and repay debt with cash flow from computing power purchase contracts. In March this year, CoreWeave completed an $8.5 billion delayed-draw term loan, placed in a bankruptcy-remote special purpose vehicle (SPV), collateralized with GPUs and customer contracts, and rated A3 by Moody's, priced at SOFR plus 225 basis points—an effective fixed rate of about 5.9%—led by Blackstone's credit and insurance unit.

Meta Platforms (META.US) used the same logic on an even larger scale at its Hyperion data center in Louisiana: a $27 billion SPV debt received an S&P A+ rating, with a yield of 6.58%, maturing in 2049. PineBridge reportedly subscribed for about $18 billion and BlackRock for over $3 billion. Meta holds only a 20% stake in the joint venture, so the debt does not appear on its balance sheet.

The asset securitization market is also expanding. By August this year, data center ABS issuance had reached $38.4 billion, up from $24.1 billion in the same period last year and about triple the full-year total of $13.6 billion in 2023. Blue Owl’s digital infrastructure credit platform has issued over $14 billion in data center loans since 2024; Apollo’s institutional credit division in July completed $4.75 billion in financing for a large-scale campus in Texas, priced at SOFR plus 340 basis points.

From chip-backed lending to investment-grade bonds

The key factor that brings these debts into the scope of insurance capital is the "credit substitution" in rating logic. Rating agencies evaluate not the chips, but the payment obligations of the off-takers: as long as the tenant is an investment-grade, hyperscale operator such as Microsoft or Meta, and the contract is a fixed-payment type, the cash flows are essentially considered quasi-sovereign receivables.

The risk is that this substitution quietly fails when the off-taker is an AI lab. OpenAI's annualized recurring revenue is about $25 billion, but its expected cash burn in 2026 may reach nearly $27 billion, with committed cloud deals with Microsoft, Oracle (ORCL.US), and Amazon approaching $590 billion. Oracle itself illustrates this risk: Of its $638 billion order backlog, about half is attributed to OpenAI, and in fiscal 2026 it is projected to have negative free cash flow of $23.7 billion. S&P downgraded its rating to BBB- in July, just above high-yield territory.

Another frequently discussed variable is depreciation. The economic life of GPUs is generally considered to be only two to three years, whereas data center buildings are designed to last 20 to 30 years. Michael Burry has built a public short thesis on this, estimating that hyperscale operators will understate about $176 billion in depreciation from 2026 to 2028.

This aligns with critics' concerns: that technology risk is entering insurance companies' balance sheets disguised as investment-grade fixed income. If AI infrastructure cash flows fall short of expectations, losses could transmit backward through the chain of AI projects, private credit, insurance companies, annuity holders, and state guarantee funds. While Wall Street profits from the spread during upswings, part of the losses during downturns may ultimately be borne by state finances.