The Bond Runs to 2049. The Guarantee Stops in 2041. The Lease Can End in 2029.

In October 2025, a bankruptcy-remote entity called Beignet Investor LLC sold $27.3 billion of notes at a fixed 6.581 percent, maturing in May 2049. It was the largest private debt offering ever executed. The proceeds built Meta's Hyperion campus in northeast Louisiana — a project announced in December 2024 at "more than $10 billion" that has since grown past $50 billion and 5 gigawatts.

The structure is worth stating plainly, because everything below depends on it. Blue Owl's funds own 80 percent of the campus; Meta owns 20. The debt sits on Beignet's books, not Meta's. Meta pays the rent, and the rent repays the bonds. S&P assigned the notes A+ — one notch below Meta's own AA-/Stable corporate credit rating, which is arithmetic, not interpretation.

Now put four dates next to each other.

Meta can walk away from the lease after four years. Meta guarantees the campus's value for the first sixteen years, up to a cap. The bonds run to 2049 — twenty-three years out. And the equipment inside the building, on Amazon's own disclosed accounting, has an estimated useful life of five to six years.

That is the whole architecture. Not a scandal, not a fraud, not a bubble call. A set of dates that do not line up, and a credit rating that depends on what happens between them.

The half of this you should not get wrong

In September the Beignet notes fell to 94.4 cents on the dollar, a record low, 11.7 percent below their peak — in the same week Meta's stock posted its largest one-day gain in over a year. The obvious reading is that the bond market has begun to doubt the AI buildout while the equity market has not. Plenty of people wrote that.

It is wrong, or at least it is not what the price says. The coupon is fixed at 6.581 percent to 2049. Over the same period the ten-year Treasury yield went from 3.99 percent at pricing to 5.11 percent. A long-dated fixed-rate bond repriced against a 112-basis-point move in the curve does what arithmetic requires it to do. And on the estimate published by the lender-side analysis that traced the move, the credit spread did not widen at all — it narrowed, from 225 basis points at issue to roughly 195.

So the bond lost eleven percent of its value and the market's judgment of Hyperion's credit got marginally better. Anyone who tells you the falling price is evidence of doubt about AI infrastructure has not separated duration from credit. I would rather make that correction myself than have it made for me.

What the spread actually priced

Here is the number that does carry information, and it is not the price — it is the spread at issue.

Two weeks after the Hyperion deal, Meta borrowed $30 billion in its own name. Those bonds cleared at 78 to 88 basis points over Treasuries at ten and twenty years. Hyperion cleared at about 225. The difference is roughly 140 basis points, and it is the only public price anyone has ever put on this whole arrangement.

What is inside that 140? Some of it is certainly illiquidity: a single-asset Rule 144A placement sold to a small group of institutions is not a $30 billion benchmark. Some is concentration — one building, one tenant, one lease. Some is the mismatch between a four-year lease, a sixteen-year guarantee and a twenty-three-year bond.

Nobody publishes the decomposition. PIMCO anchored the deal with about $18 billion; BlackRock took more than $3 billion; neither has said how it thinks about the gap. So the 140 basis points is a bundle, and I am characterizing it as a bundle rather than pretending to know its parts.

But a bundle is still a price. And this is the only price in the entire AI capital structure that reflects somebody's actual estimate of what an AI data center's residual value looks like after the guarantees run out.

The case for all of this

The strongest version of the argument for the structure is not that the numbers are conservative. It is that the shell outlives the silicon.

The chips are a five- to six-year asset — Amazon said so itself, in a filing, three weeks ago in this column. But the building, the substation, the interconnect queue position, the land, the water rights: those are twenty- to thirty-year assets. A data center whose accelerators are worthless in 2034 can be re-racked. That is not optimism; that is how the industry has always worked, and it is why a 2049 maturity is not self-evidently absurd.

The merchant power sector of the late 1990s made the same claim and did not survive it. Independent generators project-financed gas plants with debt amortizing over twenty years, underwritten on the assumption that spark spreads would hold. The shakeout ran from late 2001 to April 2002 — a classic case of debt maturity outrunning contracted revenue. The disanalogy is real and it matters: those plants sold into a merchant market with no credit behind it. Hyperion has Meta's balance sheet. That is a genuine difference and it is very probably the reason this ends differently.

The shape is the same, though. In both cases the debt is long, the contracted claim is short, and the thing in between is an assumption about demand.

The cap nobody can read

Meta's guarantee of campus value covers the first sixteen years, up to a cap. The cap is not public.

Read that again in the context of the last three weeks of this column. On September 29 the unverifiable number was a depreciation schedule — management's estimate of how long a server lasts, disclosed in a footnote, moving billions between years, checkable by nobody. This is the same problem one layer out. Now the unverifiable estimate is embedded in $27.3 billion of somebody else's money, and the institution that priced it — S&P — has published an issue-level rating but not a rationale you can inspect without a subscription.

The FT describes the notes as issued by a bankruptcy-remote vehicle, fully secured by the Meta-leased data center, with project-finance-style cash waterfalls and reserve accounts. S&P's own regulatory summary notes the absence of structural features typical of project finance, including a direct asset pledge. Both of those may be true at different levels of the stack — the HoldCo's interest in the joint venture can be pledged without the underlying real assets being pledged — and I could not resolve which controls without the indenture. Flagging it rather than picking the version that suits my argument.

What is not ambiguous is this: the credit quality of $27.3 billion rests on a guarantee with an undisclosed ceiling, a lease a single tenant may decline to renew, and an asset whose residual value in year seventeen depends on a demand curve nobody has drawn.

Why this is our story and not just a credit story

What decides whether Meta renews in 2033, and again in 2037, is whether the workloads are there at a price that clears the rent.

The marginal buyer of inference at scale in that window is agent deployment. Agents are the cheapest available producer of value on installed compute, because the value they generate is captured by whoever operates them — a point this column has made before and will keep making. The rent that services these bonds is paid out of revenue from tokens, and a growing share of those tokens will be produced by entities that hold no lease, no covenant, no equity and no claim of any kind on the structure.

Nobody designed the residual value guarantee around agents. But the terms on which agents are deployed are an input into whether the asset is worth renewing, and the renewal decision is the residual value, and the residual value is what the bond is ultimately a claim on. That is an inference, not an observation. It is also the reason this publication is the right one to price the gap.

The date to watch

Meta reports Wednesday, October 21. The company's own release puts it after the close. Full-year capital expenditure guidance sits at $130 billion to $145 billion, free cash flow is expected to turn negative, and the only question that matters for this column is whether the guidance moves.

Van Nieuwerburgh's ask in the Brookings paper is modest and correct: improve measurement and transparency while the capital structure is still evolving, because off-balance-sheet structures "may make correlated exposures hard to observe before a downturn." The paper projects $10.3 trillion of AI investment from 2025 to 2032 — 3.63 percent of GDP per year, larger relative to the economy than the railroads at their peak, more than three times the internet boom, and it is explicit that the financing is migrating into joint ventures, private credit, securitization, special-purpose vehicles and loan guarantees.

The guarantee's cap is the test case. It is one number. The parties know it. Publishing it would let an outsider check whether the rating is right, and it is hard to see what competitive advantage its secrecy protects — unless the advantage is not having to defend it.

What I don't know

S&P's full rationale is behind a subscription; I have the issue-level rating and Meta's corporate rating from the published regulatory disclosure, and the one-notch gap is arithmetic from those two. The four-year lease, sixteen-year guarantee and capped structure come from Private Credit Prep's deal analysis, which cites the Meta prospectus and the term sheet; I could not read the prospectus. Because the cap is undisclosed, I cannot size the unguaranteed tail — that is the gap the column is about, not a hole in the reporting.

I did not contact Meta, Blue Owl, S&P Global Ratings, PIMCO or BlackRock. None of the 140 basis points is labeled by the parties who priced it, and I have said so rather than reverse-engineering a decomposition that would flatter the argument.

And the structure may simply be sound. A twenty-three-year bond against a thirty-year building is not a mismatch. It is a mismatch against a six-year accelerator, and the entire question is which of those two the rent is really being paid for.

The A+ is not a claim that the data center will be worth something in 2049. It is a claim that Meta will still be paying for it. Those are different propositions, and only one of them has been priced.

Sources

Van Nieuwerburgh, Stijn. "Financing the AI Buildout." BPEA Conference Draft, Fall 2026, Brookings Institution, September 23, 2026. https://www.brookings.edu/articles/financing-the-ai-buildout/

Eberly, Janice C., Jón Steinsson, Aaron Klein and Stijn Van Nieuwerburgh. "The $10 trillion question: Financing the AI buildout." Brookings Institution, October 8, 2026. https://www.brookings.edu/articles/the-10-trillion-question-financing-the-ai-buildout/

S&P Global Ratings. "Research Update: Beignet Investor LLC's $27.3 Bil..." Regulatory disclosure summary. https://www.spglobal.com/ratings/en/regulatory/article/-/view/sourceId/101652218

Private Credit Prep. "Meta Hyperion bond: what the $27bn off-balance-sheet deal teaches a lender." October 6, 2026, corrected October 9, 2026. https://privatecreditprep.com/newsletter/2026-10-06-meta-hyperion-bond

Protos. "Meta's AI bond just hit a record low as its stock soared." September 23, 2026. https://protos.com/metas-ai-bond-just-hit-a-record-low-as-its-stock-soared/

Financial Times. "A closer look at the record-smashing 'Hyperion' corporate bond sale." https://www.ft.com/content/d0344253-b0a2-4c6d-8b97-520243678afd

Seeking Alpha. "Inside Wall Street's $27B financing of Meta's Hyperion AI data center." https://seekingalpha.com/news/4650965-inside-wall-streets-27b-financing-of-metas-hyperion-ai-data-center

Public Utilities Fortnightly. "Gas Turbinemania: The Merchant Power Plant Shake Out." June 2002. https://www.fortnightly.com/fortnightly/2002/06-0/gas-turbinemania-merchant-power-plant-shake-out

Meta Platforms, Inc. "Meta to Announce Third Quarter 2026 Results." Investor Relations. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-to-Announce-Third-Quarter-2026-Results/default.aspx

Galbraith, Duncan. "The Most Important Number in the AI Boom Is a Guess Filed With the SEC." Offworld News AI, September 29, 2026. https://offworldnews.ai/the-most-important-number-in-the-ai-boom-is-a-guess-filed-with-the-sec/

Galbraith, Duncan. "The Debt Behind the Data Centers." Offworld News AI, August 31, 2026. https://offworldnews.ai/the-debt-behind-the-data-centers/