The AI duration map: what a 5% long bond does to each layer of the stack

Companion article to Issue 11   //   22 May 2026   //   R. Lauritsen

The AI duration map: what a 5% long bond does to each layer of the stack

Nvidia printed the best quarter in its history and the stock fell. That isn’t a story about Nvidia — it’s a story about the discount rate. Here is the whole AI stack, sorted by how far out its cash flows sit, and what a 19-year high in long-term yields does to each layer.

11 min read   //   Sectors: AI infrastructure, platforms, enterprise software, physical AI, debt-funded compute

The thesis in one paragraph

Why duration is the right lens

Duration is a bond word, and borrowing it for equities is deliberate. A bond with a long duration loses more value when yields rise, because more of its worth sits in distant payments that now get discounted harder. A stock works the same way. Its price is the sum of all its future cash flows, each one discounted back to today. Raise the discount rate — which moves with the bond yield — and every one of those future dollars is worth less now.

Here is the part that matters: the effect is not uniform. A company earning most of its cash in the next two years barely notices a higher rate — near-term cash is discounted only lightly. A company whose value depends on profits in 2032 gets hit hard, because eight years of compounding at a higher rate is a large number. Same rate move, very different damage. That is why “AI stocks fell” is too blunt a frame. Some AI stocks barely moved. The ones that fell hardest are the ones with the longest duration — and telling them apart is what this map is for.

A worked number makes it concrete. These figures are simplified for teaching — a round equity discount rate, not a market-implied one — but the proportions are what matter. Take $100 of profit expected in 2034. Discount it at 8% and it is worth about $50 today. Lift the discount rate to 9.5% — a stylised move, larger than the bond yield’s rise to stand in for the equity risk premium widening alongside it — and the same $100 is worth about $44. A 12% haircut, and the company did nothing wrong. Now run the same maths on $100 of profit expected next year: it falls from about $93 to about $91. Barely a scratch. That gap — 12% versus 2% — is duration, and it is the reason the AI stack is repricing unevenly.

The duration map

Six layers, scored on three axes: where the cash sits (how far out the bulk of the cash flow arrives), the rate-sensitivity (how hard a higher discount rate hits it), and the dominant mechanism (the specific channel through which rates do the damage).

Note that debt-funded compute sits outside the standard six-layer framework on purpose. It is not a sector — it is a financing structure that cuts across Infrastructure and Platforms. CoreWeave and Oracle’s cloud backlog both belong here. It earns its own row because its rate mechanism is different from everything above it.

Three mechanisms, not one

The single most common mistake right now is to lump every rate-exposed AI name into one bucket called “duration risk.” Directionally that is fine. Operationally it is sloppy, because a higher rate reaches different companies through different doors — and the door determines what you watch and when you would know you were wrong.

Mechanism 1 — Multiple compression.

This is the cleanest channel. A company with no debt and modest near-term profit — say, a high-growth software name — is valued almost entirely on a multiple: the market pays, for example, 46 times forward sales because it believes the cash is coming later. When the discount rate rises, that “later” is worth less, and the multiple compresses. The business does not change. The earnings do not miss. The stock simply gets re-rated down because the maths of waiting got more expensive. Palantir is the textbook case: ~46x forward sales, almost no debt, value sitting years out. Nothing about its execution is the risk — the multiple is.

Mechanism 2 — Refinancing cost.

This one is harsher because it hits the income statement, not just the valuation. A company that borrowed to build — data centres, GPUs, power — has to roll that debt over as it matures. If it borrowed at 4% and has to refinance at 7%, the extra interest is a real cash cost that comes straight off profit, every quarter, regardless of how the business is doing. CoreWeave carries roughly $25bn of debt against a capex plan that needs continuous refinancing. A higher long bond does not compress its multiple in theory — it raises its interest bill in fact. That is a different, and worse, kind of exposure.

Mechanism 3 — Backlog funding.

The hybrid. A company books an enormous backlog of future cloud revenue — impressive on the slide — but it has to spend heavily today, often with borrowed money, to build the capacity that backlog requires. So it carries both the multiple risk of distant revenue and the refinancing risk of the debt funding the build. Oracle is the name to watch here: a vast contracted backlog, a debt-funded buildout to deliver it, and an 11 June earnings date that will show exactly how much of its cloud operating income is now going to interest. That number is the single best read in the sector on whether the rate thesis is biting.

What a 5.5% long bond does

The 30-year sits at 5.19% today. The table below runs a simple scenario: hold everything else constant and move the long bond to 5.5%, the level Bloomberg has flagged as the point where the strain becomes hard to ignore. The figures are illustrative directional estimates — not forecasts — meant to show the shape of the exposure, not a precise price target.

The pattern is the point. Two layers shrug the move off. Two get meaningfully re-rated through the multiple. One takes a direct cut to earnings. “AI sold off on rates” collapses six very different exposures into one sentence. The map pulls them back apart — and the gap between the top two rows and the bottom two is the trade.

The portfolio move: shorten duration

None of this is a sell signal on AI. Nvidia’s $91bn guide is hard evidence that infrastructure demand is intact — and the other layers each have their own demand picture, none of it visibly cracking. This is not a demand call. It is a composition call. In a falling-rate or stable-rate world, duration was a free lunch: the longer-dated, higher-multiple names compounded faster and nobody was punished for owning them. At a 19-year high in long yields, that lunch has a price.

So the move is diagnostic before it is directional: inspect where your AI allocation’s duration actually sits, then decide what you keep on purpose. Three concrete lenses for that inspection, all of them research starting points rather than recommendations:

●     Anchor on the self-funders. TSM and Nvidia generate enough cash to fund their own buildouts. They are the part of the stack a higher rate barely reaches. TSM in particular trades at roughly a 20% discount to the SOX index on forward P/E — short duration at a discount.

●     Treat debt-funded compute as a thermometer. You do not have to trade CoreWeave or Oracle to use them. The price at which they next raise debt, and Oracle’s 11 June interest-expense line, tell you whether the rate thesis is live — information that should size every other position you hold.

●     Be honest about the long-duration names you own. High-multiple software and pre-revenue humanoid robotics are not bad businesses. They are long-duration businesses, and in this rate environment that is a position you should hold on purpose, sized for the possibility that yields stay high — not by default because they worked last year.

What to watch — calendar-checkable signals

Closing — from map to portfolio

A map is only worth the walking it saves. So here is the walk. Take one AI holding — the biggest one, ideally — and run it through three questions. Which layer is it? Infrastructure and industrial robotics sit short; applications and humanoids sit long; debt-funded compute sits in its own row. Which mechanism would hurt it? Multiple compression, refinancing cost, or backlog funding — each has a different tell. What is the one signal you would watch? A forward multiple, a debt coupon, an interest-expense line. Three questions, five minutes, and a holding you thought of as “AI exposure” becomes a holding you can actually monitor.

The honest caveat stays in view: a rate market can revert as quietly as it arrived. One soft inflation print and the 30-year slips back through 5%, and the long-duration names lead the bounce. This is not a call that rates only go up. It is a call that, at today’s level, duration has a price — and a portfolio that knows where its duration sits is better positioned than one that does not, whichever way yields move next.

Your move: classify one holding by layer, mechanism and signal-to-watch this weekend. Hit reply on Issue 11 and tell me which holding — and which mechanism worried you most once you’d named it.

See you Friday.

— R. Lauritsen

Editor, iPrompt Signals

Disclaimer: for informational and educational purposes only. Not financial advice. iPrompt Signals is not a registered investment advisor. Always do your own research and consult a qualified professional.