Spread, Not Shed

16 October 2026. Filings, ratings and reporting are as of the dates stated in the text.

Interest disclosure: T-OpenAI and T-Kalshi are tokenized loan participation rights providing economic exposure linked to those companies' pre-IPO valuations, each issued through a dedicated issuing subsidiary. T-SpaceX, issued through its own dedicated issuing subsidiary, provides economic exposure linked to SpaceX; following the SpaceX listing, the investment underlying T-SpaceX is being divested. Redemption opens only once a Redemption Start Date is announced, and as of 16 October 2026 none has been. This piece discusses CoreWeave, whose filed customer contracts include OpenAI, and which recorded in its IPO prospectus of March 2025 that it would issue $350 million of its shares to OpenAI under their contract; we have not established whether OpenAI still holds them. It also discusses a CoreWeave facility for what the company calls non-investment-grade customers, which its chief financial officer said was priced to finance contracts with OpenAI and Cohere, and an earlier facility financing CoreWeave's contract with OpenAI, so how lenders treat credit backed by OpenAI's contracts bears on the company T-OpenAI is linked to. It also discusses AI labs as customers of the borrowers, including Anthropic, which competes with OpenAI, so Tessera's interest in how Anthropic's credit is received does not run in one direction: a competitor's valuation sits on one side, and on the other sits SpaceX, which its prospectus records as supplying compute to Anthropic. SpaceX's quarterly report also records AI hardware it leases, so how lenders treat chip-backed credit bears on it too. Tessera has no product, listing or arrangement linked to CoreWeave, Nebius, IREN, Nscale, Lambda, Crusoe or Fluidstack, and no Tessera product references Nvidia, Broadcom, AMD, Meta, Anthropic or Cohere; none is planned. Tessera's founder, and entities he controls, hold related exposure. Read what follows with that in mind.

This week a CoreWeave co-founder told the Financial Times where the next round of AI-chip financing could come from. Brannin McBee, who leads CoreWeave's financing, said that winning investment-grade ratings for more of its chip-backed debt, and eventually for the company itself, would let it borrow from insurers whose capital "is unlimited. Like, that's all the money in the world that sits in the insurance tranche." The next step, the FT wrote, is getting insurers to buy loans backed by lower-rated customers such as AI labs, and McBee said that step is not there yet: "I don't think that non-investment-grade offtakes from us are ready to go into the insurance tranche yet. I do think it'll get there." He added, by the FT's account, that the big AI labs' coming stock market listings would help. We have read the interview only as excerpted by Naked Capitalism and relayed by Seeking Alpha on 14 October, and not in the FT itself.

The FT's insurance correspondent, Lee Harris, had already supplied the other half. Many large traditional insurers are already at their limit for exposure to AI companies, he said on the FT News Briefing podcast, as The Next Web reported on 29 September. As we read The Next Web's account, he meant insurers taking on risk as underwriters, which differs from buying loans; we return to that difference below.

Both can be true at once, and that is the subject of this piece. Insurance money is already in this market, in investment-grade paper and below it. What is new is the scale being discussed, and the ratings it would take.

Insurers Are Already In

The filings name them:

• IREN: the commitments schedule (Schedule B) to the note purchase agreement for IREN's $2.1 billion of 5.96% notes due 2031, agreed in May and funded in tranches ($525 million by 30 June), was filed with IREN's annual report on 27 August. It names 58 purchasers, more than half of them US life and annuity insurers, among them MetLife, Prudential, MassMutual, New York Life, Pacific Life, Equitable, American General Life, Genworth, Guardian, Lincoln, Brighthouse, Symetra, Securian and USAA Life. IREN's annual report ties the financing to GPUs and a Microsoft contract, and describes the wider GPU financing the notes belong to as investment-grade rated.
• CoreWeave's first two GPU facilities: Liberty Mutual and Peerless Insurance appear among the lenders consenting to a December 2025 amendment on the first; American General Life, Fidelity & Guaranty Life and Security Life of Denver appear among the lenders consenting to amendments on the second (exhibits to CoreWeave's 2025 annual report).
• CoreWeave's OpenAI facility: a facility under a credit agreement of 28 July 2025 finances CoreWeave's contract with OpenAI, and its signature pages name insurers among the lenders, among them Protective Life, Midland National and USAA Casualty (8-K, 31 July 2025).
• CoreWeave's March facility: the $8.5 billion facility was "anchored by Blackstone Credit & Insurance and included participation from a diverse group of global financial institutions, asset managers, and insurance investors" (8-K, 31 March 2026). Its insurers are not named.

Where insurers are named, the amounts are redacted (IREN) or not given (CoreWeave's amendments). The filings for CoreWeave's May and August facilities name no insurers.

For scale, the NAIC's Capital Markets Bureau estimates US insurers' exposure to all private credit at "$1.2 trillion as of year-end 2025, representing approximately 13% of cash and invested assets ..." (private-credit analysis, released in September 2026). It does not break out data centers or AI, and no official figure we have found does.

So Far, the Investment-Grade Rating Has Followed the Customer

Tessera's piece of 28 August, The Money Goes Round, said of CoreWeave's facilities: "The structure is not what earns the rating. The customer is." Its piece of 26 August, What Is the Collateral Actually Worth?, put it more broadly: the ratings were "tracking the customer". At the investment-grade line both were closer to right, because the one investment-grade facility rests on a customer CoreWeave describes as investment grade. Below that line both overstated things, as the facilities show:

• March facility: backed by a single contract with a customer CoreWeave describes as investment grade (investor presentation, 8-K, 17 September) and, as CoreWeave describes its facilities before August, by a contract "extending through the maturity of the debt ..." (8-K, Exhibit 99.1, 10 August). It advances 90% of capital spending and is rated A3 by Moody's and A (low) by DBRS (8-K and credit agreement, 31 March).
• May facility: for what CoreWeave calls "two large, non-investment grade customers", it advances 71.42%, and is rated Ba2 by Moody's and BB+ by Fitch (8-K, 18 May).
• August facility: its contracts average about three years against a maturity of about five, and it advances 70%. It is also rated Ba2 and BB+ (8-K, 10 August).

Below investment grade, structure and security do some of the work. CoreWeave says the August facility's ratings reflect "the strength of the collateral and structural protections supporting the transaction" (8-K, Exhibit 99.1, 10 August). The step into investment grade is different. On CoreWeave's own account, investment-grade ratings have come "when the underlying customers are investment grade" (investor presentation, 8-K, 17 September 2026).

That matters more for an insurer than for most lenders, because insurers' capital rules run on ratings. Under the NAIC's risk-based capital factors for life insurers (adopted June 2021), the base charge on a bond rated A3, where the March facility sits, is 1.016% of its value. At the lowest investment-grade designation it is 2.168%, and at the designation corresponding to Ba2 it is 4.537%. That is a curve rather than a cliff, but every notch costs capital. Whether loans like CoreWeave's are held as bonds, where those factors apply, is something we have not been able to establish. The IMF, discussing the growth in pension funds' and insurers' illiquid investments, names private credit as one of its main drivers, and adds: "The investment-grade status of many of these exposures supports favorable capital treatment. However, a deterioration in credit quality could trigger forced sales, amplifying losses and market stress" (Global Financial Stability Report, October 2026). The Financial Stability Board warned in May that private credit ratings "may be used to attract insurers who rely on ratings for regulatory purposes". The ratings on CoreWeave's facilities come from Moody's, DBRS and Fitch, and CoreWeave has announced them in its filings, so that warning concerns the wider market rather than this paper.

So the two parts of the interview are the ones to watch together. Lending against contracts with customers below investment grade is already happening: the May facility does it, at Ba2 and BB+. CoreWeave's chief financial officer described it, the day after pricing, as "the first to be syndicated in the public loan market to finance contracts with OpenAI and Cohere" (first-quarter call, 7 May 2026). What McBee said is not ready yet is putting paper like that into the insurance tranche. If, as we read it, getting there means an investment-grade rating on paper backed by lower-rated labs, and if an investment-grade customer is what has taken a rating across the line so far, that rating has to come from somewhere else: more structure, a guarantor, or CoreWeave's own credit, which it says it is working to migrate "toward investment grade" (8-K, 17 September). None of those is independent of how the labs fare.

A Precedent from 2008

The financial guarantors, the monoline insurers, are the useful precedent. Today's market differs from theirs. What carries over is how theirs failed.

Their core business was guaranteeing municipal bonds. New York's insurance regulator later wrote that "both the market and the NRSROs viewed FGIs as having a near 'zero underwriting loss' business model ..." (Circular Letter 19, September 2008). On that reputation they wrote more, and riskier, structured finance. At the end of 2007 MBIA's collateralized debt obligations were $130.6 billion of $678.7 billion insured, and Ambac's were $66.9 billion of $524.0 billion (10-Ks). MBIA's book did not look concentrated by name: its ten largest structured-finance credits were "$24.2 billion, representing 3.6%" of net par (10-K 2007).

The concentration lay in what those names had in common. New York's regulator wrote that guaranteeing pools of asset-backed securities had "made possible additional concentration risks", among them the original lender, the servicer and the year of origination, so that a focus on the "nominal issuer" had "proven inadequate", and that strict adherence to bright-line requirements was "a poor substitute for diversification ...". The Financial Crisis Inquiry Commission put it more simply: the mortgage securities inside those CDOs "were less like coins than like slices of bread."

The losses also arrived before the claims did. "Claim payments in 2007 were negative $2 million and we paid no claims related to our CDO of ABS portfolio," Ambac's interim chief executive said in January 2008, in a results release that booked mark-to-market losses on credit derivatives of $5,211.0 million for the fourth quarter (8-K, 22 January 2008). Fitch, S&P and Moody's cut both firms from AAA in the first half of 2008: Fitch cut Ambac on 18 January and MBIA on 4 April, S&P cut both on 5 June, and Moody's cut both on 19 June (the firms' 8-Ks; for Fitch's MBIA action, MBIA's 10-Q for the second quarter of 2008). MBIA's response was that "This is an issue of ratings and not solvency" (press release of 19 June 2008, furnished on Form 8-K, 25 June 2008). For a business whose product was its rating, that was the problem. The banks that had bought protection took the losses. Merrill Lynch recorded credit valuation adjustments on hedges with financial guarantors of $3.1 billion for 2007, $2.6 billion of it on hedges of US super-senior CDOs of asset-backed securities, and $10.4 billion in 2008. Citigroup recorded credit valuation adjustments on its exposure to the monolines of $967 million for 2007 and $5.7 billion for 2008 (Merrill 10-Ks for 2007 and 2008; Citigroup 10-K for 2008).

The municipal book was not what failed. In February 2009 MBIA moved about $537 billion of US public-finance par, as of 30 September 2008 and including business reinsured from FGIC, into a separate insurer (8-K, 20 February 2009). What failed was a slice of the book large enough, with leverage on top of it, to take down the rating of everything else.

The line worth carrying forward comes from the Basel Joint Forum, in July 2008: it "would seem prudent to ask whether there is any counterparty whose creditworthiness would be unaffected by the stress events that impose losses on super-senior tranches."

Where the Analogy Breaks

It breaks in an important place. The monolines wrote unfunded guarantees: at the end of 2007 Ambac had $524.0 billion insured against what it called claims-paying ability of $14.5 billion (10-K; 8-K, 22 January 2008). In the filed record today, insurers mostly buy GPU-backed debt. A buyer can lose what it lent, but it owes nothing to anyone else when the borrower defaults. For insurers that buy, the nearer 2008 precedent is the buyers of rated CDO paper, whose capital also ran on ratings. The monolines are the template for what follows if insurers start writing protection, which is one of the three conditions below.

The one insurer guarantee we have found in the filings is contingent and limited. Athene, Apollo's insurer, has guaranteed 15% of an Apollo fund's outstanding obligation to buy compute infrastructure under the $35 billion Broadcom arrangement that facilitates Anthropic's build-out. It will be reimbursed by a third party for 35% of anything it pays, and it records no liability because it judges payment not probable (Apollo and Athene 10-Qs, second quarter 2026). That guarantee covers a purchase rather than GPU-backed paper, and we have seen no GPU-backed paper whose rating rests on an insurer's guarantee.

Nvidia has reportedly talked to insurers about taking on some of the risk of loans backed by its chips, working with the broker Howden Re on one possible structure, but the talks were described as early, and "may not result in any deals" (Investing.com, 29 September, reporting the FT; The Next Web reported the same). On the same FT podcast, as The Next Web reported, Harris said start-ups are now selling residual value insurance, which protects big chip buyers against their chips losing value faster than expected. That protects the hardware rather than the loans, and we have not found how much has been written, but the unfunded side of this market already exists. We have found nothing filed that shows an insurer writing protection on the loans at scale.

The monolines are therefore a warning about what this market would become if three things happened together:

• Ratings that cross into investment grade without the customer: ratings that reach investment grade on the structure, or on a guarantor exposed to the same downturn, when the customer is below it and the collateral is hardware likely to be worth least when that customer fails: one or two labs, one kind of hardware. That is the step from an investment-grade rating on a single strong contract to one on debt backed by a lower-rated lab. US insurance regulators have a tool aimed at part of this. Under a process the NAIC's manual makes effective from 1 January 2026, its Investment Analysis Office can propose removing a rating it judges three or more notches out of line from the exemption that lets insurers rely on it (Purposes and Procedures Manual, December 2025). We have found no published case of its use; the manual says anonymized summaries of any removals are to be published on an insurer-accessible web location (Purposes and Procedures Manual, December 2025).
• Insurers writing protection at scale: a guarantee on chip-backed loans pays out when the borrowers fail, and if what broke them was the market rather than their own mistakes, that is when the hardware is worth least. That is the correlation Tessera's 12 October piece, In Exchange for the Guarantees, described for suppliers' guarantees, moved onto an insurer's balance sheet. Harris's remark cuts against this one for now: if large traditional insurers are already at their limit as underwriters of AI risk, protection at scale would have to come from elsewhere, such as the start-ups he described or reinsurers.
• Limits set by name while the common factor goes unlimited: the neoclouds are many names, but they share collateral, supplier support and, in places, customers. Anthropic has signed with Fluidstack (company announcement, November 2025) and Nscale (S-1), and reportedly with Lambda (Reuters, 31 August). Nvidia had $36 billion of commitments to buy cloud capacity from AI clouds at 26 July. The clouds can sell that capacity to others instead, and "if AI clouds do not successfully sell committed capacity to third-party customers, we have agreed to purchase that capacity" (10-Q). Inside a single borrower the concentration is plainer: CoreWeave's three largest customers were 36%, 26% and 10% of its second-quarter revenue (10-Q). A portfolio of ten neocloud loans can pass every test set by borrower name and still behave much more like one position than ten. Size is what turned the monolines' slice into the whole firm, and nothing we have found shows that for any insurer today. If the money McBee calls "unlimited" arrives at the scale he implies, that could change.

Syndication Spreads the Risk Without Shedding It

Every new holder makes the market look deeper without making the risk smaller. Arrangers may keep a share, and the rest sits with whoever bought it. Spreading it from banks to insurers that are not levered on it the way the monolines were is, taken alone, safer than leaving it on bank balance sheets. The catch is that for the insurers among the holders, capital keys off the same ratings that sold the loan, at least where the paper is held as a rated bond, so a downgrade would raise the cost of holding it for all of them at once. The holders also overlap. American General Life appears among the lenders on CoreWeave's second GPU facility and on IREN's commitments schedule. Neither filing gives its amount, so this shows the same name across borrowers rather than how much it holds. OpenAI's contracts stand behind at least two of CoreWeave's facilities, July 2025's and May's. Where ratings rest on the same few customers, one customer's downgrade, or its decision to terminate or not renew, would reach several facilities, and the same insurers, together.

Demand for the secured paper has been strong. CoreWeave's May facility tightened by 50 basis points from initial discussions, to SOFR + 4.50% (8-K, 18 May). Reuters reported on 1 October, of Nvidia's chip-backed deals and citing banking sources: "Despite the concerns over guarantees, there remains high demand to finance the deals, the sources said." The same story reported that some lenders want higher guarantees from Nvidia than it had first outlined. If that demand is pro-cyclical, it will be thinnest when it is needed most. This week gave a glimpse of that cycle in CoreWeave's own unsecured debt: the yield on its bonds maturing in 2031 hit 15.78%, the ABC reported on 16 October. Those bonds are separate from the GPU-backed facilities, and we have found no announced rating action on the facilities this week, but it is the same company's credit.

The August facility carries a maturity of about five years against contracts averaging about three (8-K, Exhibit 99.1, 10 August). Tessera's 26 August piece called those extra two years "a bet on what a fleet of accelerators is worth in 2029 ...". The facility's terms put that bet on CoreWeave first. If the projected value of the contracts behind the loan falls below 2.40 times the amount outstanding for three monthly payment dates in a row, the cash in its collateral accounts goes to prepaying the loan, and if an expiring contract is not renewed or replaced by its renewal deadline, CoreWeave must prepay enough to restore its coverage ratios within five business days (credit agreement). Much of the last two years therefore rests on CoreWeave's ability to re-contract in whatever compute market exists then, or to find the cash if it cannot.

The 26 August piece also said that the exposure that matters "sits in private credit, which nobody is required to report." For insurers that needs narrowing: they report their holdings to their regulators line by line. What no official source we have found does is add up how much of it is GPU-backed or data-center debt. The best public evidence we have read is still a list of names on a commitments schedule.

The Next Question

This piece does not predict a loss, and it does not say that any of these loans is badly underwritten or that any insurer named here has taken too much. The March facility is close to the opposite of a monoline risk in the sense that matters here: its rating rests on one contract, running to maturity, with a customer CoreWeave describes as investment grade, rather than on a guarantor or on assumed diversification.

Insurance money is long-dated, rating-driven and large. It is the natural buyer of investment-grade paper, and that is exactly the pressure: to produce investment-grade ratings for the part of this market that is hardest to price, lower-rated customers and contracts shorter than the debt. The monolines are what happens when ratings rest on an assumption of diversification, a guarantor levered on the same assumption stands behind them, and the holders discover that what looked like many risks was one. The test for insurers adapts the Joint Forum's, asked of the whole book rather than of one counterparty: which of these exposures would be unaffected by the stress that hurts the others?

Sources: Financial Times, interview with Brannin McBee, as excerpted by Naked Capitalism (14 October 2026) and relayed by Seeking Alpha (14 October 2026); The Next Web, 29 September 2026, citing the Financial Times and its FT News Briefing podcast; Investing.com, 29 September 2026, citing the Financial Times (as republished by Yahoo Finance UK); Reuters, 1 October 2026 (republished by The Star, 1 October, and Gulf Times, 3 October) and 31 August 2026 (republished by The Star); ABC, 16 October 2026. IREN Limited, Form 10-K for the year ended 30 June 2026, filed 27 August 2026, and the note purchase agreement filed with it (Exhibit 10.36, Schedule B). CoreWeave, Inc., IPO prospectus (Form 424B4), March 2025; Form 8-K and credit agreement of 31 July 2025; Form 10-K for 2025 and its exhibits; Forms 8-K and credit agreements of 31 March, 18 May and 10 August 2026; Form 8-K and investor presentation of 17 September 2026; Form 10-Q for the quarter ended 30 June 2026, filed 12 August 2026; first-quarter 2026 earnings call, 7 May 2026 (FactSet corrected transcript, as hosted by CoreWeave). Nscale Ltd, Form S-1, filed 18 September 2026; Fluidstack, announcement of 12 November 2025. Apollo Global Management and Athene Holding, Forms 10-Q for the quarter ended 30 June 2026, filed 10 August 2026. NVIDIA Corporation, Form 10-Q for the quarter ended 26 July 2026. SpaceX, prospectus and Form 10-Q for the quarter ended 30 June 2026. NAIC Capital Markets Bureau, private-credit analysis of the US insurance industry, September 2026; NAIC life risk-based capital proposal 2021-11-L, adopted 30 June 2021, and Purposes and Procedures Manual of the NAIC Investment Analysis Office, December 2025; IMF Global Financial Stability Report, October 2026; Financial Stability Board, Report on Vulnerabilities in Private Credit, 6 May 2026. MBIA Inc. and Ambac Financial Group, Forms 10-K for 2007 and 2008 and 8-Ks of 2008 recording the rating actions, and MBIA's Form 10-Q for the second quarter of 2008; Ambac's 8-K of 22 January 2008 (results release); MBIA's 8-K of 20 February 2009; MBIA's press release of 19 June 2008, furnished on Form 8-K; Merrill Lynch, Forms 10-K for 2007 and 2008; Citigroup, Form 10-K for 2008; New York State Insurance Department, Circular Letter No. 19, 22 September 2008; Financial Crisis Inquiry Commission, Final Report, January 2011; Joint Forum, Credit Risk Transfer: Developments from 2005 to 2007, July 2008. Tessera, "What Is the Collateral Actually Worth?" (26 August 2026), "The Money Goes Round" (28 August 2026) and "In Exchange for the Guarantees" (12 October 2026), as published at blog.tessera.pe.

This is market commentary, not investment advice. It is not a recommendation to acquire, hold or redeem any Tessera product, or to take or avoid exposure to any company mentioned.

T-Tokens are tokenized loan participation rights, not equity. High risk. DYOR. Not financial advice. Not available in the US or other restricted territories. tessera.pe/terms