Software had its best August since 1990. Semiconductors lost $1.1 trillion in a fortnight. Utilities fell 4.8% in the middle of the largest electricity demand boom in seventy years. Gold rose 10.5%. We think they are one story: in August the price of long-dated money rose, and assets were re-ranked by how much of it they need.
Four stories that were reported separately, and the common thread underneath them.
August produced four headlines that appeared in four different sections of the financial press. Software staged the sharpest sector rotation in years and posted its strongest August in data going back to 1990. Semiconductors shed roughly $1.1 trillion of market value in the back half of the month. Utilities fell 4.8%, in the middle of what the sector itself describes as the largest demand inflection in seventy years. Gold had its best month since January, while the thirty-year Treasury yield reached 5.33%, a level last seen in 2007.
Each was explained on its own terms. Software rose because the fear that artificial intelligence would eat the software industry proved wrong. Semiconductors fell because the trade was crowded. Utilities fell on a California wildfire liability bill. Gold rose on debasement. Each explanation has merit. Taken separately, they miss what the four have in common.
Sort the month by duration — by how far into the future an asset's cash flows sit, and how much borrowed money is required to get there — and the ranking is close to ordered. Assets with no future cash flows at all led: bitcoin, silver, gold. Then short-duration equity, where the cash arrives now: software, which sells subscriptions off an asset-light balance sheet. Then the index. Then the capital-intensive layer that must borrow for decades to build what it sells: semiconductors, the physical AI complex, utilities. At the bottom, the benchmark long-duration asset, the thirty-year government bond.
This note argues that August was not a risk-on month, an AI-payoff month, or a rotation month in the ordinary sense. It was a month in which the price of long-dated money rose, and assets were repriced by how much of it they need. Section four sets out the clearest single piece of evidence we found.
August reversed the direction of 2026 without denting the magnitude of it.
Second-quarter earnings growth for the S&P 500, with 97% of the index reported.
The index's year-to-date total return. Earnings are rising far faster than prices.
The 30-year Treasury yield in August, a level last seen in the summer of 2007.
US public debt. The most recent trillion took roughly five months to add.
The S&P 500 delivered 52% earnings growth in the second quarter, the fastest since 2021, with 86% of companies beating on earnings and 77% on revenue. The index is up 13.1% on the year. The two figures cover different windows, but the direction is clear: earnings are growing considerably faster than prices.
When earnings grow faster than prices, the valuation multiple compresses, even in a rising market. Record earnings alongside record highs are usually read as confirming each other; here the multiple is moving the other way. That is consistent with a higher discount rate, and the discount rate is anchored at the long end of the Treasury curve, which is where the pressure was in August.
For the record, the index closed August at 7,686.14 after a 0.33% fall on the final session, with the Nasdaq Composite at 26,370.89 and the Dow at 53,185.90. The record close came on 13 August just above 7,800, and the index spent the remaining three weeks within about 2% of it — slouching sideways on thin volume rather than correcting.
The second point the scoreboard makes is about regime. August's leadership — the Magnificent 10 up 7.5% against equal weight at 1.9% — reads like the old order restored. Zoom out and it is a counter-trend month. Russell 1000 Value has returned 23.1% this year against 4.1% for Growth. Equal weight leads cap weight by 2.4 points. The Russell 2000 is up more than 19%. One good month for the mega-caps did not repair a nineteen-point factor gap; it dented it.
A double round trip in rate expectations, a record high on the thirteenth, and a mid-month hinge that reversed the leadership.
Two features of that sequence stand out. The first is the round trip in rate expectations: above 80% in late July, roughly 30% in the second week of August, 60.4% by the twenty-eighth. Anyone positioned for a directional view on the Fed was whipsawed twice inside thirty-one days, which is part of why commentary on the month reads as contradictory.
The second is the hinge around 17 to 19 August. Before it, leadership was broad and the physical AI trade was working. After it, semiconductors, electrical equipment, power infrastructure, networking gear and construction rolled over together, while software kept climbing. They rolled over on precisely the days the long end broke to a nineteen-year high. The next two sections set out what that cost, and why we attribute it to the bond market.
The AI trade rotated from the people who sell the shovels to the people who use them.
The broad S&P software index in August — its best August in data going back thirty-six years.
Semiconductor stocks in a Yahoo Finance basket that fell in the back half of the month.
Market capitalisation lost by that semiconductor group in a fortnight.
Atlassian, which nearly doubled. Okta rose 28% and CrowdStrike 20.5% on results.
The conventional account of this is a story about artificial intelligence finally paying off. For most of 2026 the bear case on software was that AI would cannibalise it: why pay for a seat licence when a model can do the work? Software and services fell roughly 20% in the first half of the year on that argument. Then second-quarter results arrived and showed the opposite — SaaS companies adding AI-driven revenue rather than losing business to it. Salesforce rose 22% on its print. Cybersecurity names beat and raised on agentic demand. The fear unwound, and a heavily shorted, deeply derated group re-rated violently.
Those explanations are reasonable, but they do not account for the timing. Software had been recovering since the spring; the sharp divergence began around 17 August, as the long end moved to a nineteen-year high. Nor was it confined to semiconductors. Everything physical rolled over together: chips, electrical equipment, power infrastructure, networking gear, construction. Those industries have almost nothing in common at the level of fundamentals. What they share is exposure to capital spending that is financed over long periods.
Software shares none of that. It has minimal capital intensity, negative working capital, subscription revenue that arrives monthly, and no need to borrow for thirty years to deliver next quarter's product. In discounted-cash-flow terms it is a short-duration asset that happens to be classified as growth. When the discount rate on long-dated money rises sharply, that is precisely what you want to own — and precisely what the market bought.
We think this reframing matters for what comes next. If August's rotation was an AI-payoff story, it should continue as long as software earnings keep beating. If it was a duration story, it continues only as long as the long end stays under pressure, and it reverses quickly if the thirty-year rallies. Those two readings imply very different portfolios. The evidence in the next section favours the second.
Of the twelve strongest prior August rallies in software since 1990, the median path held roughly flat to mid-September and then weakened into month-end. We do not trade seasonality, and twelve observations is not a sample to bet on. But it is worth knowing that software enters September with the momentum and the crowding, while the physical AI trade enters it already defensive.
Two groups of companies facing the same electricity demand moved differently on the days long yields rose.
Utilities were the worst sector in the S&P 500 in August, down 4.8%. On its face that is surprising. American electric utilities are facing the largest demand inflection in about seventy years, driven by exactly the data-centre build-out that made semiconductors the story of the decade. The fundamental backdrop was supportive.
What happened instead offers unusually clear evidence for this note's argument, because the sector's composition allows a like-for-like comparison.
On 10 August and again on 21 August, the long end backed up and regulated utilities fell as a bloc — AEP, Southern, Duke and NextEra all lower, uniformly, with no single-name catalyst. On the same days, merchant and AI-power-exposed names finished flat. Constellation closed unchanged on 21 August while AEP fell 3.79%.
Both groups sell electricity, and both face the same load growth on the same grid. The key difference is that a regulated utility earns a return set by regulators, which makes its equity behave much like a bond. Utility investment-grade debt carries the longest duration of any credit sector, roughly twelve to fourteen years, and the equity behaves the same way. A merchant power producer selling into AI demand is not a bond; its cash flows are variable and near-term.
When long rates rise, the bond-like group falls and the other stays flat, with no company-specific news, that points to a discount-rate effect rather than a fundamental one. Nothing in August suggested the AI power-demand thesis had weakened; for the regulated group, the rate move mattered more.
One complication: two idiosyncratic drags did land on the sector late in the month, a California wildfire-liability bill that hit PG&E and Edison, and a fall in GE Vernova. Those are real and they are not duration. They do not explain the uniform, no-news selling on 10 and 21 August, which is the part that carries the argument.
Regulated and merchant utilities face the same demand. On the days long yields rose, only the group that trades like a bond fell.
Section 04 · The natural experimentA nine-basis-point rally that reversed within a session, before a more lasting move once the Treasury signalled it would do more.
The context is two auctions the week before. On 12 August a $42bn sale of ten-year notes cleared at 4.683%, the highest auction yield since 2007. On 13 August a $25bn thirty-year sale cleared at 5.216% — the highest since 2001. Weak employment data and benign inflation prints, which would ordinarily rally the long end, did not.
Then on 19 August the Treasury said it would at least double the size of its liquidity-support buyback operations in the 10- to 30-year sector. Thirty-year yields fell as much as nine basis points to 5.19%. By the following session they had climbed seven basis points back to 5.26%, close to where they were before the announcement.
The scale is worth holding in mind. The Treasury had already repurchased $92.1bn of Treasuries in the second quarter and $18.7bn in the third, with a further $61bn planned through early November before this announcement enlarged it.
The mechanics explain why the first move faded. A buyback is not quantitative easing. The Treasury is not creating money to retire debt; it is repurchasing long-dated bonds and funding those purchases by issuing shorter paper. The total quantity of government liabilities is unchanged. What changes is the maturity profile — less duration outstanding, more bills.
And the Treasury is already funding overwhelmingly short. Bills accounted for 84.9% of gross marketable Treasury issuance in the twelve months to July. The buyback is not a departure from the funding strategy; it is an intensification of one already running at close to its limit.
If the problem in the long end were a liquidity problem, that swap would help, and the relief would persist. It did not persist at the first attempt — though, as the next section sets out, it did at the second. What the round trip tells you is that the constraint is not the plumbing. It is the sheer stock of duration that private balance sheets are being asked to absorb, at a moment when the Federal Reserve is no longer a large buyer and the federal deficit is running near two trillion dollars. Shortening the maturity of the debt does not reduce that stock. It postpones it, and it raises the amount that must be rolled at whatever rate prevails later.
Two further details are worth keeping. The announcement landed the day before a sixteen-billion-dollar auction of new twenty-year bonds, which reads less like strategy than like nerves. And the operation is framed as liquidity support, which is a description the Treasury has an incentive to keep using: the moment it is understood as yield management, the independence question that Warsh spent his Jackson Hole speech defending becomes considerably harder to answer.
The July federal deficit — the largest monthly shortfall since March 2021.
The deficit through July already exceeds the entire prior fiscal year's $1.775tn.
July tariff receipts. A record positive $31.35bn was collected in October 2025.
Year-on-year change in July receipts and outlays respectively.
It is worth being precise about why the fiscal position deteriorated in August rather than simply noting that it did, because the mechanism is unusual and it is reversible.
Tariff revenue did not merely fall. It went negative — a refund of $8.5bn in July following $25.56bn of refunds in June, against a record $31.35bn collected in October 2025. Receipts that had been treated as a structural offset to the deficit turned into an outflow. Combined with outlays rising 21.7% year on year in July, of which national defence rose 19.9%, the annualised deficit reached 6.1% of GDP, up from 5.7% a month earlier.
That produced a July deficit of $432bn, a record for the month, and a fiscal-year-to-date deficit of $1.799tn with two months still to run — already larger than the $1.775tn recorded for the whole of the previous year.
The ratio in the chart above is the one we would put in front of anyone who thinks the long end is mispriced. At 98.4%, mandatory spending and debt service together consume essentially the entire federal revenue base. Every discretionary dollar — defence, infrastructure, research, the machinery of government — is financed at the margin by issuance. That is the structural bid for long-dated capital described in section seven, arriving from the sovereign rather than the corporate side, and it is why we read the thirty-year yield as a supply variable.
The first attempt at reassurance did not hold; yields fell more durably only once the Treasury signalled it was prepared to do more. That suggests the marginal buyer of long-dated Treasuries is responding to the scale of official support rather than to statements. We are not adding duration on the strength of announcements of this kind.
What happened after the first attempt, and why the dollar and gold moved with it.
The first attempt did not hold, but that is not the whole sequence.
The fuller sequence has three stages. Yields fell on the announcement. They round-tripped within a session, as the previous section describes. Then, over the following days, they fell again and further — the ten-year to roughly 4.63% and the thirty-year to about 5.18% — as reports circulated that the Treasury was considering not only larger buybacks but funding them from the Treasury General Account, the government's own deposit at the Federal Reserve. The operational detail was specific: raising the cap on each buyback operation in the ten-to-twenty and twenty-to-thirty-year sectors from $2bn to at least $4bn, across 9 September to 4 November.
So the intervention did have an effect, once the market understood the Treasury was prepared to do more. The more useful question is what holding yields down costs.
The dollar fell after the buyback announcement and gold rose. Those two facts are usually read as separate market colour. Read alongside the fall in yields, they suggest the intervention had a cost.
A central bank buying bonds — quantitative easing, or the 2011 Operation Twist that swapped short Treasuries for long ones and is the closest historical mirror of what the Treasury is now doing — has always had a stated economic objective: fight a recession, fight a deflation. The stated objective was never the government's own funding cost. That is the distinction, and it is the one the market appears to be pricing.
Lower long yields plus a weaker currency is, in combination, monetary easing. It is roughly what a rate cut does. If you obtain it from the Treasury rather than the Fed, at a moment when inflation has run above target for sixty-five consecutive months, some investors may conclude that debt-servicing costs have become a larger constraint on policy. Gold is one way to express that view, which may help explain why it rose in the same month as equities.
There is a respectable technocratic case for each of these operations: buybacks do support liquidity in off-the-run issues, maturity management is an ordinary Treasury function, and the maturity profile was also shortened under the previous administration. Many observers read the same facts as prudent debt management. Our point is narrower: whatever the intent, the market has begun to price long-dated yields partly as a function of fiscal and debt-management policy, not only of economic conditions. Once that repricing starts, it shows up in the currency and in the metal before it shows up in the yield.
Holding long yields down had a visible cost: a weaker dollar and a higher gold price.
Section 06 · The price of controlThere is a serious counter-reading of Jackson Hole worth holding alongside ours. Because 84.9% of gross marketable Treasury issuance in the twelve months to July was bills, the government's interest bill now reprices almost immediately with the policy rate. That gives the Federal Reserve a powerful fiscal reason not to raise rates — and gives the market a reason to suspect that any decision not to raise is about the debt rather than the data. On this reading Warsh's hawkishness is less a Volcker moment than an attempt to correct a dovish July press conference, and the credibility question stays open. We do not know which reading is right. We note that the second one is consistent with the dollar and gold behaviour described above.
The mismatch that matters here is less about leverage than about time.
Much of the AI build-out is being financed with debt that runs far longer than the economic life of a large part of what it buys.
Graphics processors lose value quickly. The Bank for International Settlements and the rating agencies put the decay at roughly 30 to 35% a year, which implies an economic life closer to three years than to ten. Hyperscalers typically depreciate servers and networking equipment over five to six years, and data-centre buildings over twenty-five to forty. US accounting standards permit, but do not require, component-level depreciation, and filings commonly group servers and network assets in a single class, which makes it difficult to tell how a mid-life component replacement would be treated.
On the other side are the liabilities. Amazon's July financing raised $25bn across eight tranches reaching to 2066. Alphabet issued a sterling century bond maturing a hundred years out, joining a register of issuers that otherwise consists of Oxford University, the Wellcome Trust, EDF and the government of Mexico. The five hyperscalers have issued roughly $219–225bn of bonds this year, more than double the whole of 2025.
Part of what long-dated money is buying therefore has a three-to-six-year life. The consequence is not default risk — highly rated companies do not fail because their funding is long. It is recurring demand for new capital. Equipment bought with the proceeds of a forty-year bond will need replacing several times before that bond matures, and each replacement needs funding while the original borrowing is still outstanding and being serviced.
Across a sector whose capital spending is projected to approach $800bn this year and to exceed $1 trillion a year from 2027, that becomes a persistent source of demand for long-dated capital, arriving in the same market in which the Treasury is funding a deficit of close to two trillion dollars. That is why we treat the AI financing story and the term-premium story as connected, and August as a month in which they met.
The risk in this mismatch is not default. It is a recurring call on long-dated capital.
Section 07 · The financingHolder-level data for these bonds is not published. The structure of demand for long-dated credit is, and it explains how this paper is placed.
Holder registers for corporate bonds are not published, so we cannot say who owns the hyperscalers' long-dated paper. What can be said is who typically buys credit at these maturities. A forty- or hundred-year bond suits relatively few investors, chiefly those with liabilities of similar length: life insurers, pension funds and endowments. Commentary around Alphabet's century bond and this year's long-dated technology issuance has pointed to the same group. We treat that as a reasonable working assumption rather than an established fact.
That buyer base is large and growing. US retail annuity sales reached $464.1bn in 2025, the fourth consecutive record year, with ten straight quarters above $100bn, and each contract is a long-dated promise that needs a long-dated asset behind it. Historically much of that demand was met by government, utility and infrastructure debt. Issuance in the private placement market, where insurers have traditionally been the largest buyers, was reported at about $81bn through May — the strongest start in data going back to 2016 — with technology borrowers prominent.
The channels for this capital are also widening. On 10 August Nvidia announced memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish financing platforms intended to mobilise more than $500bn of third-party capital for AI infrastructure over time. Chief executive Jensen Huang indicated that Nvidia might support up to a quarter of an individual financing.
Two of those institutions, BlackRock and Goldman Sachs, also manage large pools of retirement and pension assets, so over time some of this financing may reach diversified portfolios as well as specialist credit funds. That is how a large new category of investment-grade credit is ordinarily distributed. It does mean the characteristics of the underlying assets matter to a broad range of investors.
One distinction matters for how asset life reaches a bondholder. Most hyperscaler issuance is senior unsecured: the claim is on the company and its total cash flows, not on particular equipment, so a short asset life affects bondholders mainly through the issuer's need to keep reinvesting. A growing number of asset-backed and project-finance structures work differently — CoreWeave's investment-grade financing secured on GPUs, reported at $8.5bn, is one example — and there the equipment is the collateral, so its rate of depreciation bears directly on recovery values.
We want to be careful here, because the catastrophist version of this argument is wrong. These are highly profitable, cash-generative, investment-grade companies. There is no imminent solvency question. The point is narrower and, we think, more durable: an asset-liability mismatch of this size does not resolve quietly. It resolves either through sustained refinancing demand that keeps long rates elevated, or through a repricing of the paper when the replacement cycle becomes visible in the accounts. Neither shows up in an equity screen.
Equity investors are debating how long AI equipment lasts. Credit spreads show how bond investors are pricing that uncertainty.
Hyperscaler bonds trading at a higher yield than they were issued at, by late July.
The typical yield increase since issue across that cohort.
Wide-spread 2026 new issues widened a month after pricing. The 2022–25 norm was 22–26bp of tightening.
There is an active debate about whether depreciation schedules for AI equipment are too long. Michael Burry has estimated that industry depreciation will be understated by $176bn across 2026 to 2028, with reported operating income at some firms running more than 20% above economic reality. Nvidia disputes that, pointing to customers who observe four-to-six-year useful lives in practice, and at least one accounting specialist has argued that current estimates are not unreasonable while noting the limits of how component replacement is disclosed. Goldman Sachs has identified chip useful life as the single variable with the largest bearing on cumulative capital requirements.
Credit pricing offers a separate read on the question.
Of 91 hyperscaler bonds issued in 2026 with comparable pricing data, 78 were trading at higher yields in late July than at issue, a median deterioration of about 22 basis points. Spreads widened across the curve: two-to-four-year paper from Amazon, Alphabet, Meta and Oracle moved from 30 to 40 basis points over Treasuries, five-to-seven-year from 50 to 60, and beyond twenty years from 108.5 to 118. New issues priced at more than twice the index spread widened around 10 basis points in their first month; in 2022–25, comparable deals typically tightened by 22 to 26 basis points.
Set that against the wider credit tape in August, in which high-yield spreads narrowed to a twelve-month low. The riskiest credit in the market rallied while the highest-quality technology credit cheapened. In a typical credit cycle those two move together. This looks more like a supply effect — a large new source of issuance entering an index — than a deterioration in fundamentals.
The supply mechanics are why we think this is structural rather than temporary. Hyperscaler issuance has gone from about 2% of all dollar investment-grade supply in 2022–24 to an expected 9% this year. Technology now accounts for roughly 10% of the Bloomberg Corporate Bond Index, up from 9% in 2024. Of this year's roughly $219bn, some $62bn equivalent was raised outside the dollar — in euros, sterling, Swiss francs, Canadian dollars and yen — which means the duration is being placed into every major bond market simultaneously rather than only into the American one.
And there is a great deal of room left. One estimate puts the additional dollar high-grade debt the hyperscalers could collectively issue before any of them approaches a 3% index weight — the threshold at which institutional concentration reviews typically begin — at about $1.7 trillion. That cuts both ways: it suggests fears of near-term indigestion may be overdone, and also that index limits are unlikely to cap supply soon.
One argument for the rise in long Treasury yields is crowding out. The hyperscalers are now selling large amounts of long-dated, high-quality paper to the same pool of duration buyers the Treasury relies on. More than half of this year's mega-cap technology bond financings have carried maturities of ten to fifty years, and estimates published in May put AI-related investment-grade supply for 2026 at about $300bn, equivalent to roughly $360bn of ten-year duration. Krishna Guha of Evercore ISI has described high-quality hyperscaler debt as a close competitor for sovereign debt, hedged against it in relative-value trading. The Economist observed in late August that private borrowing to fund the data-centre build-out “has also made capital costlier for everyone, including governments”.
The argument has limits. The Treasury's own long-dated issuance runs to trillions of dollars a year, against a few hundred billion from the hyperscalers, and the deficit, inflation, energy prices and Federal Reserve policy all bear on long yields at the same time. The contribution of any one source of supply is hard to isolate. We treat crowding out as a contributing factor rather than the explanation, and it is one reason we read the long end as a question of supply as much as of policy.
Useful-life assumptions differ across companies and are disclosed in varying detail, and reasonable estimates span a wide range. We do not try to settle the question company by company. Where accounting assumptions and credit spreads point in different directions, we give more weight to spreads, because they are set by investors with capital at risk.
The useful-life argument is being fought with accounting assumptions. An arm's-length market has been quoting the answer hourly all year.
Section nine described credit investors pricing equipment life cautiously. The strongest evidence on the other side comes from the rental market, and it is stronger than we expected.
There is an observable, arm's-length, hourly market price for the use of a graphics processor. It is not an accounting estimate or a sell-side model. It is what one party pays another to borrow a specific chip for an hour, and it is quoted continuously.
In mid-August the H100 — a 2022 part, two generations behind the leading edge — rented at roughly $2.77 an hour, up about 2% on the month and up 33% on a year earlier. The A100, launched in 2020, held around $1.65 and was up 17% year on year. The current-generation B200 rented at roughly $5.63, down 2% on the month and up only 7% on the year.
The oldest chip rose fastest in rental terms and the newest slowest, in the year the newest became generally available. Under the aggressive-depreciation thesis — that this equipment loses 30 to 35% of its value annually and is economically spent in about three years — a 2020-vintage accelerator should be sliding toward scrap pricing. Instead it costs more to rent than it did a year ago.
The contract market says the same. One-year committed rates rose about 40%, from $1.70 to $2.35 per GPU-hour, between October 2025 and March 2026 as capacity sold out. Research tracking the physical build reports that all capacity coming online through August and September was already booked. And the relationship between the two markets has inverted: spot pricing for leading-edge instances sat near 46% of comparable long-term reservation pricing in August, against a typical 25%, having risen 7.9% on the quarter. Short-dated compute is unusually expensive relative to contracted compute, which happens when there is nothing left to allocate.
The two pieces of evidence look contradictory, but they price different things.
The useful-life debate conflates technical obsolescence with economic scarcity. Technical obsolescence asks how quickly a chip becomes worse than its replacement; on that measure the decay is fast, and Nvidia's annual product cadence reinforces it. Economic scarcity asks whether you can obtain any working accelerator at all; on that measure there is a shortage, and a shortage puts a floor under every vintage regardless of how far behind it sits.
Today scarcity dominates, so old silicon holds its rent. But scarcity is a function of the build-out being unfinished — which is precisely the condition the $800bn of capital spending this year and $1.3tn next year is designed to end. The moment supply catches demand, the rental rate stops being set by scarcity and reverts to the technical curve. That is when the depreciation schedule on the income statement starts to matter, and it is the same moment at which the forty-year bonds financing the equipment are still thirty-six years from maturity.
So the credit market's caution and the rental market's strength are not in conflict. The rental market is pricing the next twelve months, in which the shortage persists. The bond market is pricing the next forty years, in which it cannot. Both can be right, and the difference between the two horizons is where the investment question lies.
Watch the spread between spot and one-year contract pricing. It compressed in the wrong direction this year, with spot rising toward reservation rates. A sustained move back toward the historic 25% relationship would be the first hard evidence that supply is catching up — and the first datapoint that makes an aggressive depreciation schedule the right assumption rather than a contrarian one.
Nvidia beat on everything and guided higher. The semiconductor complex had already been falling for a week, and kept falling.
Up 106% year on year, beating the 41-analyst consensus by about 4.5%.
More than doubled. Non-hyperscaler sales rose 138% to $40.3bn.
Hyperscaler capital spending, up from more than $800bn this year.
Alongside an AWS agreement to purchase two million GPUs.
On 26 August Nvidia delivered a quarter that would have been hard to improve on. Revenue of $96.2bn, up 106%. Hyperscaler revenue more than doubled. The non-hyperscaler business — AI clouds, industrial and enterprise, the part that answers whether demand is broadening beyond six customers — grew 138% to $40.3bn. Backlog above $2tn. Capital spending guided to $1.3tn for 2027. The shares rose after hours.
And it changed almost nothing. Semiconductors had been selling off since the seventeenth, and finished the month up 1% against software's 16%. Fifty-eight of sixty-one names in the group fell over that stretch. A single session of relief on the twenty-seventh, in which Broadcom rose 4.5% and the VanEck semiconductor ETF gained 3%, did not repair it.
The disconnect argues against the most common explanation for the month. If semiconductors had fallen on AI demand doubts, the single most authoritative demand datapoint available — from the company with roughly 80% of the accelerator market, with $2tn of backlog — should have arrested the decline. It did not.
The reason, we think, is that the market was not expressing a view on demand. It was expressing a view on what it costs to finance the equipment that demand requires. Nvidia's guidance is relevant to that question. Combined capital spending by Microsoft, Alphabet, Amazon and Meta reached $166bn in the June quarter alone, up 87% year on year and 27% from the March quarter. Over the ten quarters since the start of 2024, combined hyperscaler capex has risen 272%. Guidance now points to more than $800bn this year and $1.3tn next.
As a demand signal, that is very strong. As a supply signal — much of that spending is financed with long-dated bonds that must find buyers — it points to a large and growing source of duration, arriving as the Treasury funds a deficit of close to two trillion dollars. The same guidance is supportive for AI revenue and, through the financing it implies, a source of pressure on the long end. In August the market appeared to weigh the second more heavily.
Nvidia guided gross margin to 74% for the current quarter, a small move from a very high level. We will follow margins, together with the mix of hyperscaler and non-hyperscaler demand, as the new third-party financing platforms scale.
Warsh delivered a hawkish surprise. Only half the curve reacted, and that half is the tell.
Kevin Warsh, chairman since May, used his first Jackson Hole address to correct a muddled July press conference at which he had declined to share either his economic views or his framework. He called recent inflation trends concerning, citing a 3.7% twelve-month change in the July PCE index, and said that while the summer's readings were better than expected they did not persuade him that underlying trends had meaningfully improved. Headline CPI and core PCE have now run above the 2% target for sixty-five consecutive months.
He described the economy as having strengthened, attributed slower hiring to a flattening labour supply rather than weakening demand, and made clear that short-term rates remain the Fed's instrument — explicitly setting aside both artificial intelligence and the balance sheet as considerations for near-term policy. Both had been arguments he previously advanced for easier policy, which makes the reversal a deliberate one.
Markets read it as hawkish, and violently so. Futures-implied odds of a quarter-point September increase went from about 35% the day before the speech to 58% immediately after — a near-doubling in a session, settling around 60% in the following week. Gold fell 3.1% on the day. The dollar firmed. Two-year Treasury yields moved sharply higher.
And then the curious part: the ten-year barely moved. The front of the curve, which the Fed sets, repriced hard. The part that discounts every long-dated cash flow in the world did not.
That asymmetry is the most useful information the speech produced. A hawkish surprise from a credible central bank normally lifts the entire curve, because expected short rates are an input to every point on it. Here the front end repriced and the long end shrugged. One interpretation, offered by several analysts, is that investors were reassured — a Fed willing to raise rates will contain inflation, so long-dated compensation need not rise. An alternative reading, which we think deserves weight, is that the market increasingly treats the policy rate and the term premium as separate problems with separate owners. The Fed sets the first. The Treasury, the deficit and the AI financing complex between them set the second.
Warsh himself supplied the line that frames it. "We should not indulge a regime in which market participants are looking primarily to the Fed for their next trade." In August they largely did not. They looked at the deficit, at the supply calendar and at the metal. That is consistent with the regime he described, with the corollary that central-bank communication has less influence on the part of the curve that matters most for discounting equities.
Adam Posen of the Peterson Institute pointed to the tactical consideration: having signalled this clearly, the Fed would face questions about what had changed if it held in September with inflation still elevated. That decision will be taken against continued public calls from the President for lower rates.
The rally was led by the instruments with the most operating leverage, which tells you what kind of rally it was.
Spot gold, from $4,045 at the end of July to roughly $4,430.
Silver, narrowing the gold–silver ratio through the month.
The GDX miners ETF, from $74.10 to $97.94.
China's August gold purchase, above July's 640,000 ounces.
Gold has had a violent year. It set a record near $5,600 in late January, then lost more than 20% in its worst quarter since 2013 as the Iran conflict pushed energy prices and inflation expectations higher and forced markets to price a more hawkish Fed. It bottomed near $4,000 in early August and then put in its strongest month since that January peak.
The internal structure of the August move is more informative than the headline. Silver gained about 15% against gold's 10.5%, and the miners gained 32%. When the leveraged expressions lead by that margin, the buying is not primarily reserve-manager accumulation, which is patient and price-insensitive. It is financial demand arriving quickly.
Underneath that, the structural bid did continue. China added roughly 20 tonnes in August, its largest monthly purchase in nearly three years, lifting reported holdings to $350.08bn from $306.35bn — though most of that $43.7bn swing is the price, not the tonnage, and it is worth saying so. Central banks bought a quarterly record 288.9 tonnes in the second quarter, up 62% year on year, and they did it while the price was falling. Reserve managers were adding into the worst quarter since 2013. That is the floor this rally built from.
The debasement framing is the one that fits. The trigger was not a growth scare — equities rose in the same month. It was the combination of a $40tn debt stock crossed in August, a Treasury visibly uncomfortable with the long end, and a deficit near $2tn with no central-bank buyer of size. The Congressional Budget Office projects federal debt held by the public rising from 101% of GDP this year to 120% by 2036. Gold does not need inflation to work in that setting. It needs the market to doubt the arithmetic, which is a lower bar.
Central banks bought a quarterly record of gold in the quarter its price fell hardest since 2013 — behaviour more consistent with strategic reserve allocation than with trading.
Section 13 · The metalEnergy equities gained 7.4% in a month Brent fell 4.3%. That divergence is the investable part.
Energy has been the best-performing major asset class of 2026, and August extended it: the sector gained 7.4% after 12% in July. What makes the month interesting is that it did so while Brent fell 4.3% to $92.71 and WTI finished essentially unchanged at $86.34.
There is a specific mechanism, and it is not in the crude price at all. On 17 August the Nymex diesel crack spread — the margin between diesel futures and WTI crude — rose above $100 a barrel for the first time on record. Refiners earn that spread. An energy company that buys crude and sells distillate gets richer when the gap widens, whatever the underlying barrel does, and in August the gap did something it had never done before.
This is what the closure of the Strait of Hormuz actually does to the market. It does not simply raise the price of oil; it disrupts the movement of refined product, which is harder to reroute than crude and which is what economies actually consume. The strait has remained essentially closed to most shipping since the conflict began in February, and the refining bottleneck is where that shows up.
Sector and commodity normally move together. When they separate, the market is usually telling you something about margin rather than price. Oil is still roughly 24% above where it sat before the conflict began at the end of February, and the Strait of Hormuz has never fully reopened. Attacks on shipping in the second week of August pushed Brent back toward $90 and dampened hopes of a reliable transit corridor.
The macro consequence runs through inflation. Energy did a third of the work in the most recent headline CPI print, with gasoline up 3.9% on the month and the energy index up 2.1%, leaving energy 16.3% higher than a year earlier. That is what keeps headline inflation at 3.4% while core behaves considerably better, and it is why Warsh can describe underlying trends as unimproved without contradicting the softer prints.
For a portfolio, the implication is that energy is currently functioning as an inflation hedge that also generates cash, which is rare and which is why we continue to hold it. The risk is symmetric and worth naming: a durable Hormuz reopening would take a substantial premium out of the price, and energy equities have now rallied twice on a commodity that has stopped rising.
Five weeks after an intervention that cost two treasuries their credibility, the currency was back where it started.
We flagged the coordinated yen intervention in the July note as the most under-priced macro event of that month, with the longest half-life. The half-life turned out to be about five weeks.
By the end of August the yen had softened again, and TOPIX had gained 3.9% in large part because of it. This is the same machine Japanese equities have run on for four years: a weaker currency lifts translated exporter earnings, the index rises, and nothing structural changes. The Bank of Japan sits at 1%, the highest policy rate since 1995, having held in July against a dissent from Takata who wanted 1.25%. Board members have since been notably more forward-leaning, with Takata describing 2026 as a regime change and Governor Ueda stressing that conditions remain accommodative.
The lesson we draw is about intervention generally, and it rhymes with the Treasury buyback in section five. Interventions that do not change the underlying rate differential buy time, not levels. Japan spent a record sum in the spring and saw it undone within six weeks; the joint operation in July bought about five. Until the BOJ moves the policy rate enough to change the carry arithmetic, currency operations are a tax on the reserve account.
The bond market is where Japan is moving most. The ten-year JGB reached a thirty-year high of 2.97% in August and 3.04% in early September, with the thirty-year at 4.22%, against a policy rate still at 1%. Second-quarter GDP was soft in real terms — up 0.3% on the quarter against a 0.5% consensus — but the deflator rose 2.6% year on year against a 2.3% consensus, and nominal GDP has now run in a 3–5% range for eight consecutive quarters. Nominal growth running roughly 90 basis points above the ten-year yield is the classic configuration for yields to keep rising. Money markets moved to price around 28 basis points of tightening at the 18 September meeting and 44 by year-end.
Korea sits in a related position. The Bank of Korea raised to 2.75% in July, its first increase in more than three years, after the won touched a seventeen-year low of 1,561.5 in June. The currency has since recovered towards 1,485 and the governor has pointed to a large current-account surplus as grounds for further strength. Korea remains the cleanest high-beta expression of the semiconductor cycle in public markets, and it remains, as we wrote in July, a national benchmark that behaves like a leveraged bet on global memory demand.
A 25% move in bitcoin funds is either noise or the cleanest read on the thesis. We think it is the second.
Bitcoin exchange-traded funds returned 25.4% in August, rebounding sharply from a weak July, with the associated volatility index edging up 3.2%, which Cboe read as a rally accompanied by two-way volatility rather than a steady climb.
We hold no position and are not making a case for the asset. We include it because of where it sits in the ranking that opened this note. If August was a month in which investors reduced their exposure to sovereign promises, then the asset with no sovereign attached to it at all should have led. It did, by a wide margin, in the same month gold had its best run since January and silver outran gold.
We treat it as an indicator rather than an allocation. A single month proves nothing on its own; the same instrument fell in July. What is harder to dismiss is the coherence of the ordering across seven very different asset classes in the same thirty-one days. Correlation across a debasement basket is a reasonable indicator that the market is pricing a currency question, and the debasement basket in August was unusually well behaved.
Five developments that would have led the month in a quieter year.
Société Générale expects the group's aggregate free cash flow to go negative and not to recover for at least two years. A sector that funded itself internally for a decade is becoming a structural net borrower, which is the mechanical reason the bond supply keeps arriving.
A published analysis of the footnotes in nine large technology companies' most recent filings — uncommenced leases and purchase commitments, mostly AI-related — arrived at about $3tn, against roughly $600bn of reported capital spending over the same period, and around three times what those companies owe under their outstanding leases and long-term borrowings combined. An estimate cited in June put the figure at $1.8tn, though it may not have been prepared on the same basis. New third-party financing structures, including the platforms Nvidia announced on 10 August, are part of the same broader shift toward funding outside corporate balance sheets.
Up from 9% in 2024. Index-tracking credit portfolios therefore hold a growing share of AI-related issuance by construction.
The domestic activity backdrop is strong, which is precisely why the hawkish case is credible. We do not read the long-end sell-off as a growth scare.
Unemployment ticked down to 4.1%, wages rose 3.2% against CPI of 3.4%, and layoffs and quits were both stable. Few hires, few fires, and real earnings slightly negative. It is an unusual configuration and it gives the Fed cover in either direction.
On a global basis growth returned 2.6% and value 2.6% — identical to a decimal place. The American software-versus-chips story was violent enough to dominate the narrative, but at the level of the world index August produced no factor winner at all.
Energy and technology led, but there were signs of broadening. An estimate that large is the strongest argument against the bearish reading of this note, and we hold it in view deliberately.
Twenty-one deals priced in August against single-digit monthly counts from January through June, with July and August together raising about 69% of the year's total issuance. Alongside it, foreign institutional investors bought roughly $2.5bn of Indian equity, the largest monthly inflow in twenty-two months and the highest of any emerging market, with the capital described as rotating out of Korea and Taiwan. That is the duration trade in this note expressed across Asia: money leaving the two markets that are leveraged bets on the semiconductor capital cycle, and arriving in the one that is not.
The first US strike on Iranian positions since late July, with Iranian state media reporting retaliatory attacks on US bases in Jordan. Crude barely moved on it, which tells you how much security premium is already in the price.
Five positions, and the evidence that would change each of them.
The response to the 19 August buyback announcement — a reversal, then a more lasting decline once the Treasury signalled more support — suggests official action can influence long yields, at a cost visible in the dollar and in gold. Shortening the maturity profile does not reduce the duration the private sector must absorb. We are not extending duration into official reassurance, and we would want a materially higher real yield before we did. What would change it: a credible reduction in the deficit path, or evidence that hyperscaler issuance is decelerating rather than compounding.
The equity is priced for the build-out succeeding. The investment-grade paper is priced closer to the risk that the asset life is shorter than the schedule assumes, which is the live question. Within credit we have no appetite for the forty-year and hundred-year tranches, where the gap between the life of the debt and the life of the compute it partly funds is widest. What would change it: componentised depreciation disclosure, which would let us test the useful-life assumption directly.
Second-quarter earnings grew 52% and the index is up 13.1% on the year. The windows differ, but earnings are growing considerably faster than prices, which is what a higher discount rate does to the multiple. We do not need an earnings miss to lose money from here; we need long rates to stay where they are. That argues for holding businesses whose cash flows arrive sooner. What would change it: a sustained fall in the 30-year below roughly 4.75%.
Two consecutive months of sector leadership on a commodity that stopped rising tells us the market has accepted the level. Energy currently provides an inflation hedge that also generates cash. What would change it: a sustained reopening of the Strait of Hormuz, which would take a substantial security premium out of the price.
The rotation began on the day the long end broke, not on an earnings date, and it took the whole physical complex down together — chips, power equipment, networking, construction — which have no common fundamental. Owning software here is a way of owning short duration. We are sizing it as such, which means we do not expect it to survive a long-end rally. What would change it: software continuing to outperform through a sustained fall in the thirty-year, which would mean the payoff reading was right and ours was wrong.
The rental market says old accelerators are scarce and getting dearer; the bond market says the paper funding them is getting cheaper to buy. Both can be true because they price different horizons. Practically, that makes us more comfortable with near-dated exposure to compute scarcity and less comfortable the further out the maturity goes — which is the same conclusion we reach on the credit, arrived at from the opposite direction. What would change it: the spot-to-reservation spread reverting toward its historic 25% relationship, which would mean the shortage is ending sooner than the bonds assume.
The August rally happened alongside an equity rally and soft inflation prints, which rules out the conventional explanations. Central banks bought a quarterly record into the worst price quarter since 2013. That is reserve behaviour, not trading behaviour. What would change it: evidence that official-sector accumulation is slowing at higher prices, which would leave the rally resting on financial demand alone.
Equities, metals and bitcoin rose while the long bond fell. We read August as a month about the price of long-dated money more than about risk appetite.
Section 18 · What we thinkThe dates that decide whether August was a turn or an interruption.
| Date | Event and why it matters |
|---|---|
| 10–11 Sep | August CPI. The first full print since the Jackson Hole speech. A firm number makes the September meeting live; a soft one complicates the case for a hike. |
| 15–16 Sep | FOMC. Priced at roughly 60% for a 25bp increase after Jackson Hole. The press conference matters more than the decision, because the question is how the Fed describes its influence on long-term yields. |
| Mid-Sep | Bank of Japan. Holding at 1% since June, with one board member already dissenting for more. A move here would do more for the yen than either intervention did. |
| Ongoing | Hyperscaler issuance calendar. Watch the tranche structure, not the size. More forty-year-plus paper extends the mismatch described in section seven; a shift to shorter maturities would narrow it. |
| Quarterly | Third-quarter results and any change in depreciation schedules. A shortened useful-life assumption at a large issuer could prompt a reassessment across the sector. |
| Ongoing | Long-dated Treasury auctions, especially twenty- and thirty-year. Tail size and dealer take-down are now the cleanest read on whether the marginal buyer is still there. |
| Ongoing | Strait of Hormuz transit volumes. Still the single largest swing factor for headline inflation, and therefore for the rate path and for everything discounted by it. |
| Nov | US midterm elections. Relevant here mainly because of the suggestion that the Fed's hawkish impulses are being held until afterwards — a claim that is unfalsifiable now and testable in December. |
Almost every major asset rose in August. The main exception was long-dated government debt.
Raffles Square Capital · September 2026Sources. Compiled from contemporaneous reporting and published research including Reuters, Bloomberg, CNBC, the Financial Times, NPR, PBS, Axios, Fortune, Al Jazeera and Yahoo Finance, alongside primary and house material from J.P. Morgan Asset Management, Goldman Sachs, Vanguard, Cboe, the Chicago Booth Review, Morningstar, State Street, Nasdaq, FactSet, the World Gold Council, the Bank for International Settlements, the US Treasury, the Federal Reserve, the Congressional Budget Office, EY, Sprott Money, LSEG, Kiplinger and Nvidia's own newsroom, together with published sell-side and periodical research from Jefferies, CLSA, The Economist, Barron's and the Wall Street Journal, commentary from Evercore ISI as reported, Investing.com and The Footnotes Analyst, together with contemporaneous single-session market commentary for the intraday sector moves cited in sections three and four, and Bank of America research as reported by third parties for the GPU rental rates in section ten.
Confidence. Index and sector returns, the 30-year and 10-year Treasury levels, the Jackson Hole speech and the market reaction to it, the 19 August buyback announcement, the Nvidia financing announcement of 10 August and the hyperscaler spread data are multiply sourced. Monthly gold is reported here as +10.5% using a July close of $4,045 against roughly $4,430 on 31 August; some coverage cited figures between 10% and 14% depending on the window and whether futures or spot were used, and one analyst quotation described it as the strongest monthly gain this century, which we could not corroborate and have not repeated. Silver is given as +14.9% on the same basis, with other sources quoting 15% to 17%. Hyperscaler year-to-date issuance is reported between $219bn and $225bn across LSEG, S&P Global, BNP Paribas and J.P. Morgan depending on the cut-off date and whether non-dollar and related-entity issuance is included. GPU depreciation rates, useful-life assumptions and the Burry estimate are contested and are presented as such. The $464.1bn annuity figure is for calendar 2025. Second-quarter earnings growth of 52.0% is the FactSet blended figure with 97% of the index reported. Year-to-date index returns are quoted through month-end August; intra-month snapshots published on 21 August showed 12.1% and 15.5% on a price basis, and the two are not directly comparable. August software and semiconductor returns are quoted from the IGV and SOXX ETFs at 16% and 1%; a broader S&P software index gained roughly 14% and the PHLX semiconductor index roughly 2.2%, and LSEG put software and services at 13.5%, so the size of the gap is robust but its exact magnitude depends on the index chosen. The $1.1tn semiconductor value figure and the 58-of-61 breadth count are from a single published basket measured from 17 August. The single-session utility moves on 10 and 21 August are from market commentary rather than exchange data and have not been independently verified, though the pattern recurs on both dates. Nvidia's fiscal second quarter 2027 figures are from its own release of 26 August. GPU rental rates are spot quotations that vary materially by provider, region, interconnect and commitment, and the figures used here are one research desk's tracking as reported at two dates in mid-August; published H100 rates across the wider market in the same period ranged from under $2 to roughly $7 depending on the provider and contract type, so the year-on-year direction is the reliable signal rather than the level. The $1.7tn index-headroom estimate and the 46.3% spot-to-reservation figure are single-source sell-side estimates. US fiscal figures — the July deficit, the fiscal-year-to-date total, tariff receipts, the receipts and outlays growth rates and the 98.4% interest-and-entitlements ratio — are US Treasury data as compiled and presented in third-party research, and the 98.4% ratio uses a specific definition of entitlements given beneath that chart; other definitions would produce a different number. The record diesel crack spread, the 84.9% bill share, the Treasury buyback sizes, the August auction yields and the Japanese GDP and JGB figures are likewise taken from published research rather than retrieved from source databases. September hike odds are quoted at 35% before the Jackson Hole speech and 58% immediately after, per contemporaneous reporting; other published series showed roughly 56% rising to 60.4% over a slightly longer window, and the two are measured at different moments rather than in conflict. The $3tn off-balance-sheet estimate is one publication's reading of filing footnotes and is not a reported accounting figure. Holder registers for the hyperscaler bonds discussed in section eight are not public; the buyer base described there is the typical one for these maturities, not an observed one. The CoreWeave financing figure is as reported in secondary industry compilations. Charts are drawn from the figures cited; where a continuous series was unavailable, intermediate paths are illustrative and end-points are the reported values.
Disclaimer. This article is commentary and general information only. It is not investment advice, not a recommendation to buy or sell any security, and not an offer or solicitation in any jurisdiction. It does not take account of any reader's objectives, financial situation or needs. Past performance is not a guide to future returns. All figures are as at the date of publication and are subject to revision.
