One Rate, Three Landings

The Fed's first hike in three years lands on the levered AI builders and the unlevered pure plays four to eight times harder than on the index — and on Microsoft, Alphabet, Amazon and Meta not at all that the data can see. Three questions get there, and the tape had answered the last one before the meeting.

TON618 Capital Research · As of 18 September 2026 · Download PDF

On Wednesday the Federal Reserve raised its policy rate for the first time since 2023 and, in the same statement, named one of the reasons it could: "capital investment is robust." That put a single rate move on top of a roughly $710bn-a-year capital build, and the argument that followed — rates are now the headwind to AI; the hyperscalers fund from cash and do not care — treats the whole complex as one thing. It is not, and the note asks three questions that separate it — the third is the one the note is really about, and the first two are what the third rests on. First, has the market already paid for the rate? If it has, the summer's move is in the price and "rates versus AI" is last quarter's trade. Second, does the hike change what the build has to earn? The house published in July that 2026's capex falls ~$137bn a year short of its hurdle at an assumed cost of capital; if the rate is now the lever, that number moves. Third, who in the complex actually receives a rate move, and how many times? A hyperscaler funding from cash, a company reinvesting its cash into three-year hardware, and a company borrowing to do it are three different claims on the same asset — or, on the builders' reading, one claim priced off the customer's credit — and a reader deciding where a rate move lands needs to know which one is being held. The answers matter because they decide what the argument is about. The finding is the third one: the rate lands on the levered builders and the pure plays four to eight times harder than on the index, and on the four not at all that the data can see. The first two answers are why that is the right reading rather than a coincidence of the summer — the market has paid for the rate, and the hurdle barely moved, so what is left of the rate move is the part that lands on leverage. The rate is not a verdict on the build; it is a sorting variable inside it. The case against reading it that way comes first, because the answers have to beat it.

Terms used throughout, defined once. The discount rate is the rate at which a stream of future cash is converted into a value today; a higher rate makes distant cash worth less now. The cost of capital, or WACC (weighted-average cost of capital), is the blended return a company's lenders and shareholders require on the money it deploys. A hurdle is the minimum annual profit an investment must earn to be worth making. The ten-year is the yield on the ten-year US Treasury note, the benchmark risk-free rate. A basis point is one-hundredth of a percentage point. A DDTL (delayed-draw term loan) is a bank loan the borrower draws down as it buys equipment; SOFR is the overnight rate those loans float on. Non-recourse debt is owed by a subsidiary and secured on its assets alone; the lender cannot pursue the parent. A mandatory convertible preferred is a share that pays a fixed dividend and must convert into common stock on a set date. A rate beta is the average percentage move in a share for a given move in the ten-year yield, estimated from daily data.

The first question — has the market already paid for the rate? — is not an AI question, and the note takes it first because the answer sets the terms for the other two. A rate move reaches an equity first through the discount rate, and that leg belongs to the whole market. The FOMC's own projections are where the shock lives: the median participant now expects the funds rate to end 2027 at 4.1%, up from 3.6% in June, and 2028 at 3.9% from 3.4%. The 2027–28 path moved up half a point in one meeting; the longer-run rate moved only from 3.1% to 3.2%. Sixteen of eighteen participants pencilled in another increase this year. The curve responded as a bear flattener — the two-year rose 7 basis points on the day to 4.74%, the ten-year 1 to 5.01%, the thirty-year fell — which is the shape a curve takes when it believes the tightening will work. Two-year/ten-year is 25 basis points, from 47 at the end of July.

The market-wide question is whether equities have paid for the summer's move in the risk-free rate, and the honest answer is: mostly.

FactSet's forward twelve-month price/earnings multiple for the S&P 500 was 20.4× on 30 June and 19.1× in its 18 September report — a 6.4% compression while the ten-year rose from 4.44% to 4.94%. In yield terms, the index's forward earnings yield rose 33 basis points against a 50-basis-point rise in the risk-free rate: a pass-through of two-thirds. The gap between the earnings yield and the ten-year narrowed from 46 basis points to 30. That is the entire market-wide story, and it has two readings that the note cannot separate in a single quarter: either one-third of the move was absorbed by a falling equity premium — investors demanding less extra compensation for owning stocks over bonds — or the last third is still owed. Both are consistent with a market that has, on net, taken the rate seriously.

What the note can say is that the four largest capital spenders did not behave like the index. Alphabet fell 2.5% from 31 July and Amazon 7.5%; Microsoft rose 7% and Meta 23%. The dispersion is the point: their idiosyncratic earnings news dominated whatever the rate did to them. The regression in §3 makes this precise — the four's realized sensitivity to the ten-year over the window is indistinguishable from zero — and it is the first fact the note concedes to the fourth point of the case above. Whatever the discount rate did to Microsoft this summer, the tape did not record it.

The Gordon arithmetic — 8.5% to 9.4% against 3% growth, 18.2× to 15.6× — belongs here as a statement of what a fixed equity premium would imply, not as a forecast. The market has not paid it for these names, and the note does not claim it will.

The first answer, and why it matters. Mostly paid at the index level; not visibly paid at all for the four largest spenders, whose own earnings drove their prices. That removes "will the market reprice the AI trade for rates" as the live question for the index — two-thirds of it already happened — and moves the question to where the discount rate is not what moves the stock: the builders, in §3.

The second question — does the hike change what the build has to earn? — is the one the house owes an answer to, because it published the hurdle. The July note computed the annual operating profit that 2026's capital expenditure must earn to clear its cost of capital and replace itself as the silicon wears out. The identity is one line: required annual profit equals capital deployed times the sum of the cost of capital and the economic depreciation rate. On a 60/40 split between three-year IT hardware and fifteen-year buildings and power, and a cost of capital of 8.5%, the 2026 vintage needed $221bn a year and, after crediting every dollar of merchant and internal AI profit, fell $137bn short.

The 8.5% was an assumption. The July note said so — "an assumed stress input, not a derived one." This note derives it from Wednesday's curve.

Ninety basis points on the cost of capital adds $6.4bn to the annual hurdle on a $710bn vintage. Because the identity is linear in the rate, the sensitivity the July note could not publish is now a single number: every 50 basis points of cost of capital demands $3.55bn a year more from the 2026 vintage, $6.8bn across the $1.36tn of 2024–26 capital still inside a three-year life, and $4.25bn on the ~$850bn the market expects for 2027.

Set that beside the July note's other lever. Booking the short-life hardware over six years rather than three moves the same hurdle by $71bn a year — twenty times the effect of a 50-basis-point rate move. The reinvestment landing is real, it is measurable, and it is second-order. The note states this in its highlighted layer because it is the house's own finding and because it settles the sixth point of the case above: the payback clock's conclusion — that payback requires the capex to plateau — was never a rates argument, and Wednesday did not make it one.

The second answer, and why it matters. Almost nothing changes in the arithmetic: the hurdle rises $6.4bn on a $228bn base, and the shortfall the July note found is intact at ~$143bn for reasons that have nothing to do with the rate. It matters because the market's argument this month — that the hike is the thing that decides whether the build pays — has the lever wrong by a factor of twenty per 50 basis points — eleven times even against the full 90-basis-point move — and the lever is the life of the silicon; and the reader who wants to know whether the build pays should watch the depreciation schedules and the 2027 capex guides, not the funds rate.

Two things do change. First, the vintage that prices the new dot path is 2027's, whose size does not exist until the four guide in late October; the note leaves that cell open rather than assume it. Second, the Committee itself named the reinvestment leg as a cause of the rate: "capital investment is robust" is, in the statement's own logic, one of the reasons the economy can carry 4% money. The build is raising the rate that reprices the build. That loop is real, and it is the subject of §4.

The third question — who in the complex actually receives a rate move, and how many times? — is where the builders' case meets the filings, and the filings settle some of it and leave the rest to the tape. The answer has two parts: who borrows and at what price, from the 10-Qs; and how each name has traded against the ten-year, from the tape.

Three of the four now borrow to build. Microsoft issued no debt in fiscal 2026; its cash-flow statement records proceeds from debt issuance of zero and repayments of $3.0bn, against $40.3bn outstanding. It is the only one of the four for which the third point of the case above holds. Alphabet issued $51.8bn of senior notes in the first half of 2026 — $20bn in dollars at a 4.80% weighted coupon and fifteen-year weighted maturity, the balance in sterling, Swiss francs, euros, Canadian dollars and yen — and then, in June, raised a further $49.6bn net in common stock and a 6.25% mandatory convertible preferred, "including capital expenditures to scale AI infrastructure," and established a $40bn at-the-market equity programme. Amazon issued $37bn of dollar notes in March at 3.85–6.05% with maturities to 2076, roughly $30bn more in euros, francs and Canadian dollars through June, and a further $25bn of dollar notes after quarter-end at 4.60–6.25%; it also signed a $17.5bn three-year term loan at SOFR plus 0.625–0.875%. Meta issued $25bn in May at 4.55–6.45% with maturities to 2066, taking its notes to $84bn, alongside $349bn of non-cancellable purchase commitments and $279bn of data-centre leases not yet commenced.

Oracle, one tier down, issued $43bn of senior notes in fiscal 2026 — $18bn in September 2025 at 4.45–6.10% and $25bn in February 2026 at 4.55–6.85% — plus a 6.50% mandatory convertible preferred. Its interest expense rose 55% year on year in the quarter to August, its fiscal-2026 operating cash flow of $32bn covered 57% of its $55.7bn of capex, and its senior notes trade at $105.7bn against $125bn of face.

The four are not a tier. They are a range, and "they fund from cash flow" is true of one of them. But the finding cuts the other way too, and the builders' case is entitled to it: three companies issued between $25bn and $92bn of debt this year at 3.85–6.45% and show a rate beta of zero. Borrowing, as such, is not the third landing. The landing is what the debt is secured against and how much of the equity it leaves — and on that measure the four and the builders are not doing the same thing.

The builders borrow at a different price, and the price is rising within the year. CoreWeave's Note 10 lists $35.6bn of principal at effective rates from 2% to 15%: its two convertible issues at 2% and its non-recourse March facility at 7% sit at the bottom; its recourse term loans run from 9% to 15% (the first at 15%, the second at 11%), its unsecured senior notes at 10% (the April and June issues carry 9.75% and 9.625% coupons), its equipment-vendor financing at 11%. The company disclosed $640m of interest expense in the second quarter and guided the third to $860–940m; its first-half debt service — $5.2bn of principal and $982m of interest — exceeded its $3.7bn of operating cash flow. The instructive series is its own loan ladder in 2026: the March facility priced at SOFR plus 2.25%, the May facility at SOFR plus 4.50%, and the August facility, the first to finance shorter-duration customer contracts, at SOFR plus 5.50%. Between the first and the last the policy rate had not moved; the spread had widened 325 basis points because the collateral had changed. That is the third landing in one table: the marginal AI dollar at a levered builder costs what the customer behind it costs, plus what the lender demands for the term. Two facts from the same filings pull the other way and belong beside it: CoreWeave says its weighted cost of debt fell almost 300 basis points over the year as cheaper facilities replaced its 15% first loan, and its disclosed sensitivity to floating rates after hedging is modest — roughly $30m a quarter per 100 basis points — because its facilities require at least 95% of floating borrowings to be swapped. The hike's direct cash cost to CoreWeave's existing book is small. Its cost to the next facility, and to the $6.2bn maturing in 2027 and $4.4bn in 2028, is what the ladder is measuring.

IREN made the counterparty point in a single sentence of its fiscal-year results. Its $3.6bn of GPU financing for the Microsoft contract carries a weighted rate of 6.0%; its new $2.4bn facility from Blue Owl and PIMCO for non-investment-grade customer deployments carries 9.0% fixed. Three hundred basis points for the same hardware, the same site, a different counterparty. This is the first point of the case above, and it is correct — the coupon is a function of the customer, and IREN's Microsoft paper is priced inside Meta's own — but it does not remove the landing. The counterparty's credit sets the level of the third landing; the level is 300 basis points wide inside one company's own book; and the non-investment-grade customers are the ones the builders are adding fastest.

Nebius sits in a third place. Its $9bn of debt is entirely convertible — 0.50% to 4.50% coupons — with the cost carried by the equity option rather than the cash coupon; its first asset-backed facility, in July, priced at SOFR plus 2.50%; and 70% of its second-quarter contracts carried customer prepayments covering half to three-fifths of the associated capex. The prepayment is the mechanism by which a builder lends its customer's credit to itself, and it is the strongest structural answer the builders have.

The market has ranked the three landings already — and paid the builders while doing it. The realized sensitivity of each name to the ten-year — daily returns regressed on daily changes in the yield, 31 July to 18 September, thirty-four sessions — sorts the complex by funder, with the exceptions the table shows:

Long-dated Treasuries (TLT) register −1.2% per 10 basis points with an R² of 0.78 — the regression is measuring what it should. One threshold applies to every row: a coefficient is treated as established where its t-statistic is at or beyond 2 in absolute value. IREN, CoreWeave, Marvell, Dell, Oracle and Nvidia clear it; Nebius (−1.7) does not, and is reported as pointing the same way at lower significance; none of the four comes close (the largest, Meta, is −0.9). Against the Treasury ruler, the builders that clear the bar carry four to eight times the index's rate sensitivity; the four carry none that the data can distinguish from zero.

The daily beta and the summer's level read opposite ways, and both are in the table. Every builder outran the index over the window — IREN +18%, Oracle +16%, Nebius +14%, CoreWeave +11% against 2% for the S&P — while carrying the highest sensitivity to each day's yield change. The ranking is about how the equity moves with the ten-year, not about which equity the summer punished; a reader who wants to know who got hurt should read the third column, and a reader who wants to know who is exposed should read the first.

The exposure ranking is largely, not entirely, a leverage ranking. IREN's FY27 capex guide is one and a half times its market capitalisation and CoreWeave's debt is four-fifths of its equity value; Alphabet's capex is 4% of its market value and its debt 2%. The identical rate move lands on a claim that is, in the builders' case, a thin slice of equity over a thick layer of debt over a three-year asset — and on a claim that is, in the four's case, almost entirely equity over a diversified cash-flow machine of which the AI build is a fraction. But Marvell (−5.0%, t −2.9) and Dell (−4.1%) carry builder-class betas with no builder-class leverage, and Nebius carries a builder-class beta on 15% debt to market cap. The AI-pure-play premium is duration of its own, independent of debt: leverage explains the ranking's shape, not all of its magnitude, and the note does not claim more.

Three structural caps on the table's denominators are stated rather than left silent. Debt to market capitalisation counts non-recourse debt as though it were full leverage on the common — for CoreWeave that overstates the recourse burden by about a tenth, for IREN's GPU financing by more. Capex to market capitalisation ignores the customer prepayments that fund 45–60% of the builders' GPU purchases before the hardware ships. And the per-50bp ratio below divides a burden on 2026 capex by 2026 operating income, when the capex is contracted to earn through 2027–29 and the builders' whole model is that the denominator grows into it. Each cap flatters the note's ranking; none reverses it, because the sorting variable the market has used is the daily beta, which needs no denominator.

On the day of the hike itself the builders rose 2–4% against a one-basis-point move in the ten-year: CoreWeave +3.0%, Dell +3.6%, IREN +2.5%, Vertiv +2.0%, Oracle +2.0%, while Microsoft fell 1.4% and Amazon 1.0%. Against a beta of −5% per 10 basis points, a +3% day on a +1-basis-point move is a positive residual of roughly three points — relief that the decision matched its pricing, not evidence of immunity. A −3% day on the same one-basis-point move would have been an equally unexplained residual; the session is one of thirty-four observations, not evidence either way, and the note reads it as neither.

The per-50bp arithmetic, per name. Because the hurdle identity is linear, 50 basis points demands 0.5% of capex in additional annual profit from every name alike. What differs is the denominator. For Microsoft that is $950m against $155.2bn of fiscal-2026 operating income — six-tenths of one percent. For CoreWeave it is $185m against a 2026 adjusted-operating-income guide of $960m–1.15bn — roughly 18%, with the caveat above that the guide is a 2026 number and the capex earns through 2029. Same rate, same identity, nearly a thirty-fold difference in what it means to the income statement that must absorb it in the year it lands.

The third answer, and why it matters. The rate lands on the levered builders and the pure plays, four to eight times harder than on the index, and on the four not at all that the data can see — and the market had ranked it that way for the seven weeks before the meeting. It matters because it changes what the rate question is a question about. It is not "does the hike hurt AI"; the index has paid and the hurdle barely moved. It is "which claim on the build is being held": a hyperscaler's equity, which the rate does not reach; or a builder's, which is a thin slice of equity over debt over three-year hardware, and which the tape has priced as a rate instrument for the seven weeks to the meeting — whether that beta is leverage or counterparty credit doing the work is what the next facility's spread (§5, condition 2) decides.

The three answers together raise a fourth point the note did not set out to ask, and the statement's causal sentences are why: they invert the framing the note began with. "Economic activity is expanding at a solid pace. While uncertainty remains elevated owing, in part, to geopolitical developments, domestic spending has been resilient. Productivity growth is strong, and capital investment is robust. Job gains have kept pace with the workforce, and the unemployment rate has changed little. Inflation remains elevated." The Committee is not describing a rate rise imposed on the AI build from outside. It is describing an economy whose investment boom is one of the reasons it can bear a higher rate, and whose inflation — core PCE projected at 3.4% for 2026 — is the reason it must.

That makes the build endogenous to its own cost of capital in a way the July note did not model. Each increment of capex supports the activity that supports the rate that raises the hurdle on the next increment. The loop has a natural governor — the plateau the payback clock already identifies as the payback condition — and it has a sign: a build that slows lowers the rate that reprices it. This is a mechanism, not a forecast; the note flags it because the pre-hike framing ("rates are a headwind to AI") has the arrow pointing one way, and the statement points it both ways.

Tripwires (in-note; none is a publishing hold). The 28 October FOMC (Polymarket prices a further 25-basis-point increase at 54.5%, another 2026 increase at 83.5%) · the ten-year at 5.25% and the thirty-year at 5.50% · the four's Q3 guides · CoreWeave's Q3 print and any facility priced after it · Oracle's next issuance against its $7.2bn fiscal-2027 and $10.1bn fiscal-2028 maturities · the Bank of Japan, which raised to 1.25% on 18 September on a 7–2 vote, so that both legs of the yen-funded carry are now rising · investment-grade and high-yield spreads, which have not moved (0.78% and 2.70% on 17 September).

The three answers rest on the following.

Observed (primary sources, dated). The FOMC statement and Summary of Economic Projections of 16 September 2026; Treasury yields from FRED through 17 September and the 17:05 ET quote feed on 16 September; FactSet Earnings Insight of 4 and 18 September 2026 for the forward multiple; every coupon, principal, maturity and covenant figure from the issuers' own 10-K, 10-Q, 6-K and 8-K filings listed in §7, with accession numbers; CoreWeave's 2026 capex and operating-income guides from its company-published call transcript; IREN's fiscal-2027 capex guide and Nebius's 2026 capex guide from two independent transcriptions each of their August calls; market capitalisations and daily closes as of 18 September.

Inferred (computed from the observed, with the code archived). The pass-through ratio in §1 (earnings yield change over risk-free change); the derived WACC and re-run hurdle in §2 (wacc_rerun.py); the per-50bp table; the realized rate betas and window returns in §3 (rate_beta.py, ordinary least squares of daily log returns on daily yield changes, 34 observations, t-statistics shown); the funder-tier ratios (funder_tier.py).

Assumed (stress inputs, not claims). An equity beta of 1.10 and an equity risk premium of 4.5% for the big four's cost of equity — at 4.0% the derived WACC is 8.9%, at 5.0% it is 9.9%, and the hurdle sensitivity of $3.55bn per 50 basis points is unchanged by either; a 10% debt weight; the July note's 60/40 asset split and three-year short-life clock, defended there; 2026 capex of $190bn / $185bn / $200bn / $135bn for Microsoft / Alphabet / Amazon / Meta (company guides, tracker basis), $37bn for CoreWeave and $22.5bn for Nebius (2026 guide midpoints), and $27.5bn for IREN — which is its fiscal-2027 guide midpoint (July 2026 to June 2027), the only forward-year figure in the table, labelled as such; Microsoft's fiscal-2026 and Oracle's fiscal-2026 operating income from their 10-Ks, Amazon's and Meta's first-half 2026 operating income annualised for the ratio table. The regression window opens on 31 July because it is the first session after the four's second-quarter results and the start of the summer's move in the ten-year; a window from June would add sessions in which the four's earnings, not the rate, drove the tape. The ten-year is used as the ruler for every name, including borrowers whose loans float on SOFR, because the equity claim on a long-duration asset is priced off the long rate; the SOFR sensitivity is reported separately where the issuer discloses it. The Gordon terminal-multiple arithmetic in §1 is an illustration of a fixed equity premium, not a forecast, and the note makes no claim about where any multiple will go. The note cannot separate, within one quarter, an equity premium that fell from a repricing that is still owed. Realized rate betas over 34 sessions carry R² between 0.08 and 0.28 for the individual builders; the tiering is robust to the ruler (TLT, LQD, SPY all print where they should) but individual coefficients are estimates with the t-statistics shown, and the note applies one significance threshold to every row.

Federal Reserve Board, FOMC statement and Implementation Note, 16 September 2026; Summary of Economic Projections, 16 September 2026 (fomcprojtabl20260916.pdf). FRED series DGS2, DGS5, DGS10, DGS30, T10Y2Y, DFEDTARU, BAMLC0A0CM, BAMLH0A0HYM2, T5YIFR, DEXJPUS, pulled 18 September. FactSet Research Systems, Earnings Insight, 4 September and 18 September 2026. Microsoft Form 10-K for the fiscal year ended 30 June 2026 (accession 0001193125-26-323660). Alphabet Form 10-Q for the quarter ended 30 June 2026 (0001652044-26-000071). Amazon Form 10-Q for the quarter ended 30 June 2026 (0001018724-26-000026). Meta Platforms Form 10-Q for the quarter ended 30 June 2026 (0001628280-26-050705). Oracle Form 10-K for the fiscal year ended 31 May 2026 (0001193125-26-277521) and Form 10-Q for the quarter ended 31 August 2026 (0001193125-26-389274). CoreWeave Form 10-Q for the quarter ended 30 June 2026 (0001769628-26-000366), Note 10, and company-published corrected transcript of the second-quarter 2026 earnings call, 11 August 2026. Nebius Group Forms 6-K of 20 March and 24 August 2026 (0001104659-26-032735; 0001104659-26-100347), Letter to Shareholders, Q2 2026, and two independent transcriptions of the 12 August 2026 call (The Motley Fool; Investing.com). IREN Limited Form 10-K for the fiscal year ended 30 June 2026 (0001878848-26-000052), Note 23, FY26 results release of 27 August 2026 (Form 8-K, 0001878848-26-000051), and two independent transcriptions of the 27 August 2026 call (The Motley Fool; Investing.com). Bank of Japan statement of 18 September 2026 as reported by CNBC and Bloomberg. Polymarket contract prices, 18 September 2026. Market capitalisations and daily closes, 30 July to 18 September 2026. Analysis code and every extracted filing table are archived in the note's data/ directory; the July hurdle engine is re-run unchanged with the cost of capital as a parameter. The evidence-coverage gate for this note passed on 18 September 2026 at 60 of 60 files.