THE SECOND OPINION
A guarantee from an investment-grade counterparty is worth exactly what its timing says it is worth. The Hubbard financing is the cleanest illustration the market has produced this cycle, and it is about to be sold to people who will read the guarantor and skip the trigger.
Google is backstopping Anthropic's obligations on a $15 billion data center campus in Hubbard, Texas. The paper is expected to carry a speculative-grade rating anyway. That is not a ratings error. It is the agencies pricing the one thing the headline does not carry: when the support attaches.
The reporting contains the tell, and it is a single clause. Google's backing was set at the minimum level lenders required to complete the financing. Not the level that made the credit safe. The level that cleared a committee. Everything else in this structure follows from that sentence.
THE ANATOMY
Start with the borrower, because almost every write-up gets this wrong by omission. The debt is to Nexus Data Centers, the developer. Not to Anthropic, the tenant. Not to Google, the guarantor. A creditor of this paper has a claim on a development company whose principal asset is a construction site and a set of leases.
The package is a $14 billion bridge plus a revolver, arranged by a syndicate led by Morgan Stanley with JPMorgan, RBC, TD Securities, and Mizuho. The banks intend to syndicate it out through multiple bond issuances as the borrower draws against construction milestones. That is the originate-to-distribute shape, and it means the arrangers' incentive is to move the paper, not to hold the risk they underwrote.
Layered on top: Apollo and Blackstone extend private credit through an SPV that leases the hardware. Broadcom provides vendor financing on the chips with a residual backstop. Google guarantees four leases and the associated power purchase agreements, and takes roughly twenty percent equity in both the data center and the power project.
Six distinct claims against one set of megawatts. Bank loans, bonds, private credit, vendor paper, a lease guarantee, and equity — each sold to a different buyer, each booked as a different exposure, all of them levered to whether one campus in Limestone County gets built on time and gets used.
WHERE THE RISK ACTUALLY SITS
The guarantee activates when the facility is operational and the lease commences. Before that: delay, cost overrun, equipment lead times, and completion risk sit with debtholders. After that: lease and power payments sit with Google.
So the window in which a data center is most likely to disappoint is precisely the window in which the investment-grade name is not standing behind it. This is not hidden. It is in the structure, it is why the rating is what it is, and it will be disclosed. It will also be the first thing forgotten when the bonds are marketed on the strength of the guarantor's name.
Anyone who has worked a prepay deal recognises the shape immediately. The guarantor exists, the rating agencies see it, and the only question that has ever mattered is what has to be true before it performs. Everything else is decoration.
THE PART NOBODY IS TOTALLING
Hubbard is not the first of these. Roughly $35 billion of prior Anthropic-linked data center debt already exists across five US facilities developed with Fluidstack, under the same structure: Google repays bondholders on a tenant default, with support activating once the facility is operational. TeraWulf, Cipher, Hut 8, and a Next Frontier joint venture have raised bonds against it.
So the template is already at scale, and the aggregate contingent liability sitting behind a single guarantor across all of it is a number no one has published. Add Hubbard and the figure is meaningfully north of $50 billion of leases and power obligations guaranteed by one balance sheet.
That balance sheet is being watched. Alphabet's five-year CDS has widened this year as bond investors price AI buildout risk. The guarantor's own spread is the cleanest available read on how the market is scoring the aggregate, and it is not improving.
THE RATEPAYER QUESTION
Friday's letter argued that load-forecast risk in the AI buildout lands on ratepayers, because the transmission and generation that serves the load is funded through a regulated rate base. Hubbard is the structure that answers back.
A 1.6 gigawatt gas plant on site, inside the fence, is a deliberate exit from the interconnection queue and from the rate case. No transmission upgrade, no cost allocation proceeding, no county commission with a vote on the electric bill. The developer buys its way out of the political layer entirely.
Which relocates the residual rather than removing it. Behind-the-meter generation converts a ratepayer problem into a bondholder problem: the debt now carries power plant construction risk, gas supply risk, and merchant exposure on any surplus, on top of the data center. Both letters describe the same question. Only the holder changes.
WHAT WOULD MAKE THIS WRONG
The guarantee is broader than reported. If final documents show Google support attaching at financial close rather than at completion, the construction-risk argument collapses and the paper is mispriced cheap. Watch the offering memorandum.
Completion is de-risked contractually. A hard-dollar EPC wrap with a creditworthy contractor and liquidated damages would transfer most of what I am describing away from bondholders. That would be the single most important term in the deal and it has not been reported.
The bonds clear tight anyway. If this prices through comparable single-tenant paper, the market is telling me it reads the guarantee as effectively unconditional, and the structural point is academic. Pricing is the test.
Timeline. Syndication is expected as draws hit milestones. First take on this is checkable within two quarters.
THE WEEK AHEAD
Tuesday — July CPI. The whole week. A soft jobs print took the next hike off the table for now; a hot core CPI puts it back on. The long end at 5.20 with 30s trading through 20s says duration is being rationed, not bid. Record equity highs into a live inflation print is a market long convexity it has not paid for.
Thursday — July PPI; 30Y auction settles into it. Pipeline read on the tariff pass-through. Watch the goods/services split against that 82% YTD heating oil move.
Friday — Retail sales, industrial production, Michigan. The consumer against 6.7% y/y retail sales momentum and a Fed still talking tightening. Michigan 1Y inflation expectations are the sleeper print.
All week — Hubbard syndication watch. Any OM language on completion support or an EPC wrap is the falsifiable-claims section resolving in real time. First bond tranche pricing is the test.
All week — Fed speakers reacting to the jobs/CPI sequence. The market cut hike odds on one print. The speakers will tell you whether the committee agrees. Listen for 'patience' versus 'vigilance'.
THE BOARD
Structure per public reporting on the Nexus Data Centers financing; terms are pre-syndication and subject to change. Levels: Koyfin, Friday 8/7/26 close.
Disclosure: this letter is drafted with the assistance of AI tools, including models built by Anthropic, which is the tenant in the transaction discussed. All analysis, structural interpretation, and conclusions are the author's own.
The Bond Bro Dispatch is published by Positive Carry LLC. All content is general market commentary provided for informational and educational purposes only and does not constitute investment advice, a recommendation, an offer, or a solicitation to buy or sell any security. Nothing herein is tailored to the circumstances of any recipient. Data are drawn from sources believed reliable; accuracy and completeness are not guaranteed. [email protected] · © 2026 Positive Carry LLC. All rights reserved.