Forty-year money, four-year asset

Forty-year money, four-year asset

Somebody has to own the forty-year paper. That question sits underneath everything else this week — the refunding, the CPI whispering, the renewed argument about whether inference revenue justifies the capex — and almost nobody asks it out loud, because asking it means admitting the AI trade migrated out of the equity market months ago and set up shop somewhere the average macro strategist doesn’t have a Bloomberg page open for.

Meta sold 2065s (including the $4.5 billion 5.75% tranche in its October 2025 $30 billion multi-tranche deal, plus a €1.5 billion 6.40% 2065 in its early-November 2026 euro package). Oracle sold 2065s (the $2 billion 6.10% piece in its September 2025 $18 billion deal). Alphabet came to market in dollars and euros in the same broad window, most recently with a €6 billion multi-tranche euro deal in early November 2026. And the Blue Owl–anchored Hyperion vehicle put a six-handle coupon (6.581%) on Louisiana concrete via the $27.3 billion Beignet Investor senior secured notes due 2049, with Meta holding a fifth of the equity and the debt sitting politely off to the side of the earnings model everyone screens on. Add it up and you get the largest new source of investment-grade duration supply in a generation, arriving in the same window as the Treasury’s own long-end auction calendar, competing for the same balance sheets.

The equity market is still debating whether demand is real. Credit already answered. It wrote a check with 2065 on it.

Here is the tension I can’t get past. The paper is long. The collateral is short. A GPU has a useful life measured in the same units as a presidential term, generously. The financing assumes an asset that throws off cash for decades, like a toll road or a pipeline or a regulated utility. Somebody is going to be wrong about that, and it won’t be the silicon. The useful-life debate that flared last autumn — how many years you depreciate a server over, and what that does to reported earnings — was treated as an accounting spat. It’s the entire credit thesis. Stretch the depreciation schedule and you flatter the P&L. The bond doesn’t care. The bond just sits there, compounding, long after the H-series has been sold for scrap to a second-tier inference shop in Ohio.

Now look at who’s supposed to buy it.

For thirty years the marginal buyer of Western long duration was a Japanese life insurer or a European pension fund with a liability curve to match. That buyer is going home. Thirty-year JGBs with a three-handle do the job domestically, without the FX hedge that has made dollar assets structurally unattractive to Tokyo for most of the past decade. The Ministry of Finance has been trimming super-long issuance precisely because it noticed demand thinning at the top of its own curve. Nippon Life doesn’t need Meta’s 2065s when it can buy the Japanese government’s 2055 at home in its own currency and sleep fine. The BOJ spent two years walking the policy rate off the floor, and the second-order effect wasn’t a carry unwind in the way the 2024 tourists remember it. It was a repatriation of duration demand. Slow, boring, devastating.

So the West’s pension complex gets asked to absorb sovereign supply and quasi-sovereign tech supply simultaneously, at a moment when its funded status is the best it’s been in twenty years — which means it needs less duration, not more. That’s the trap nobody models. A well-funded pension buys the long bond and stops. It doesn’t reach for the extra thirty basis points in a Louisiana SPV.

Which brings us to the Fed, and to the most misread chart on everyone’s screen.

The front end has been priced all year for a central bank that cuts when asked. Fine. Suppose it delivers. Suppose CPI cooperates and the December cut is a formality and the successor’s dovishness is fully monetized in SOFR futures. None of that lowers the cost of the buildout by a single basis point, because the buildout doesn’t borrow at the front end. It borrows at thirty and forty years. Cutting into a backing-up long end gives you a steepener, cheaper rolling paper for levered floats and Treasury’s bills, and a data center financing cost that hasn’t moved. QT ended a year ago; reserve management purchases of bills are not duration removal, whatever the briefing slides imply. If the long end genuinely misbehaves, the only tool that touches it is buying it, and buying it has a name nobody at the Eccles Building wants to say into a microphone in an election-adjacent year.

Watch the curve, not the funds rate. Watch the tail on this week’s thirty-year. Watch the new-issue concession on the next hyperscaler deal — that’s the honest price, the one that tells you whether the buyer base is deep or just polite. Watch Oracle’s five-year CDS, which has spent the autumn trading like something with a rating it hasn’t been given yet. Watch whether the SPV paper starts trading like project finance instead of tech, because the day it does, the equity multiple has to come down to meet the bond.

And watch gold, which keeps getting explained away as a de-dollarization story when the simpler read is that it’s the only reserve asset with no duration and no counterparty. Central banks are buying an asset that cannot be refinanced. That’s a very specific fear.

Bitcoin was supposed to be that. It isn’t, and this cycle finally proved it. No coupon, no maturity, infinite sensitivity to the discount rate — BTC is the longest-duration thing most people own, and it has traded like a levered claim on the same data center buildout, right down to the liquidation cascades. When NVDA sneezes, ETH catches it. The digital-gold thesis has been quietly repriced into a high-beta Nasdaq expression, and the people who bought it as insurance are discovering they bought more of the same risk with worse liquidity.

I’m not calling a top. The compute is real, the revenue is arriving, and the depreciation is a genuine expense against genuine cash flow. But a boom financed with forty-year money against four-year assets has a specific failure mode, and it doesn’t announce itself in an earnings miss. It announces itself in a badly-covered auction on a Thursday afternoon when nobody’s watching.

Somebody has to own the forty-year paper. Right now the honest answer is: whoever’s left after the lifers go home.