OpenAI and Anthropic's bankers are pre-clearing the labs for insurance and pension bond money

For a large part of the buyer base, a rating is an eligibility test, not advice that a bond buyer weighs and may ignore, which is exactly why you go and get one before you need to issue anything. Pension funds and insurance companies run mandates capping how much below-investment-grade paper they may hold ^[3], so the distance between a high-yield notch and an investment-grade one is not a few basis points of spread, it is whether a pool of money is allowed to open the prospectus.

The scale being financed is the part that rewards arithmetic. S&P has the six largest US hyperscalers spending more than $7 trillion on data centres and AI-related capex from 2025 through 2030 ^[8], an average above $1.17 trillion a year ^[13], against LSEG's roughly $720 billion for the five biggest in 2026 ^[7]; different baskets, so read it as direction rather than like-for-like, but the direction is something like 62% above next year's run rate, sustained for six years ^[14]. Goldman Sachs Research puts more than $1 trillion of global AI investment into 2026 alone ^[5], while its economist Joseph Briggs has the cumulative total since 2022 above $1.8 trillion by the end of that year ^[6], which leaves roughly $800 billion for the four years before it ^[15]. Funding a single year that outspends its four predecessors combined takes debt, not operating cash, which is close to what the Bank for International Settlements said in January when it noted the turn toward debt and private credit ^[4].

Then there is the collateral problem: durable collateral is largely absent. PwC has servers and GPUs replaced every four to six years inside a $31.6 trillion data centre capex path to 2050 ^[10]. Lend for longer than that and you are underwriting the refinancing rather than the hardware, which is a decent business while adoption holds and a bad one otherwise, and PwC's own caution is that slower adoption or weaker pricing complicates the financing of the later stages ^[11]. S&P's September 3 temperature check said capex is outrunning expectations, financing structures are getting more complicated and less transparent, and returns may take years ^[9]. That warning is published by a ratings agency, and ratings agencies are the constituency being lobbied ^[1].

This can break the other way without anything going wrong at the labs: the agencies may decline to hang an investment-grade notch on issuers that have no ratings yet ^[1], or the listings may raise enough that debt stays a standby facility and the rating is just a cheap option. Cryptopolitan sources the push to the Financial Times, and reports no agency response and no listing date ^[16]. The structural point survives either outcome, because cheap and repeatable borrowing lets a frontier lab buy compute without selling shares, and widens the gap to competitors who must sell them ^[12]. Equity holders in a private lab chose the risk they hold, whereas an insurance general account would hold this paper because a rule said the notch was acceptable, and that transfer is the thing to price.