The Chairman was describing, accurately, the month's central fact. The Committee left its target range at 3.50–3.75% for a fifth consecutive meeting, and in the same breath he noted that yields had risen enough between meetings to rank "around the top decile" of inter-meeting moves in twenty years. Policy did not tighten in July. Financial conditions did, and they did it on the long end, where the borrowing that funds artificial intelligence actually gets priced.

Fourteen billion-dollar venture rounds. The previous record for a single month was lower, and July set it while the listed version of the same trade was losing a fifth of its value. Private marks are stale by construction, but a record is a record, and no fact in this issue argues harder against our thesis.

Forty-two days, no policy change. Between the June and July meetings the Federal Reserve did nothing, and the ten-year rose 31 basis points while the thirty-year rose 36. If you were waiting for a central bank to signal the turn, you missed it, because the turn was priced by the people who buy the bonds.

The VIX at 15.99. A fear gauge sitting near the year's lows on the last day of a month that erased $2.2trn tells you the market did not read July as a systemic event. It read it as a bill arriving at one address.

Two developments were worth highlighting, the Chairman said, and he put the bond market first. "Nominal and real yields are materially higher across the Treasury curve," he told the room on 29 July. "In fact, some of the increases in market interest rates between FOMC meetings are among the most significant in the last two decades, ranking around the top decile or so."

Then he asked the question the whole month turns on. If the Committee did not change its policy rate, what happened? His answer was that attention had shifted to the data itself, helped by the fact that he had stripped forward guidance out of the statement. "Market participants are learning to play the ball, not the referee," he said, "and market prices will continue to respond in the direction and magnitude they see fit." A few minutes later he put it more plainly: the Committee had not done much in forty-two days, and "the markets have done quite a bit."

That is the cover thesis in the Chairman's own words, and our claim is about where the bill landed. A tightening delivered through the long end does not fall evenly across a market. It falls hardest on assets whose value sits in cash flows a decade out, and hardest of all on assets that must raise money today to build the thing that produces those cash flows. In 2026 that description fits one complex almost exclusively.

So the semiconductor index lost about 21% in July, its worst month since October 2008, and roughly $2.2trn of market value went with it, while the S&P 500 as a whole fell 0.13%. Consensus called this the AI bubble deflating. We think that reads the tape backwards. Nothing in the demand data broke in July. Alphabet raised its 2026 capital-spending guidance to a range of $195bn to $205bn, reported Google Cloud revenue up 82% year on year to $24.8bn, and disclosed a backlog that had grown $50bn in a single quarter to $514bn. No hyperscaler cut a dollar of planned spending. What deteriorated was the market's willingness to fund the spending: order books for hyperscaler bonds thinned from roughly five times covered in February to below two times in July.

Read that way, July was not a referendum on whether artificial intelligence works. It was a repricing of who can afford to build it, and the market answered by selling everything that borrows and buying the versions of the same trade that do not.

Ordered latest-first. Sources in the register (p.50); each row is dated inside the coverage window.

The war came first and did the least. US strikes on 8 July and the exchanges of 13 and 14 July put Brent up more than 20% on the month, but oil was already relenting by the 27th and the semiconductor complex kept falling anyway. Energy set the inflation backdrop; it did not set the direction of the equity.

The hinge is 22 July. Alphabet delivered a quarter that was, on any demand measure, excellent, and raised its capital spending by roughly $15bn at the midpoint. The stock fell about 5%. That is the whole thesis in one session: the market had stopped rewarding scale in AI and started pricing the funding of it.

By the last week the sort was complete. Korean memory, the most capex-levered expression of the trade anywhere, tripped consecutive circuit breakers. Indian IT services, the least capex-levered, was finishing its best month in six years.

The first channel needs no explanation to anyone who has priced a security. Push the long rate up 36 basis points and every cash flow twenty years out is worth less today. It applies to all long-duration assets equally, which is precisely why it cannot be the whole story.

The second channel is narrower and it is where July's damage concentrated. For most of the last decade the hyperscalers funded capital spending out of operating cash flow, which made their building programmes indifferent to the bond market. That is no longer true. The Bank for International Settlements records the shift plainly: Alphabet, Amazon, Meta, Microsoft and Oracle have moved from funding capex with operating cash flow to issuing debt, with the five largest set to spend more than a trillion dollars on AI-related capex across 2025 and 2026. Morgan Stanley's tally of AI-related global debt issuance stood at $236bn at the end of May, four times the level a year earlier, on a path to roughly $570bn by the end of 2026.

A borrower of that size is no longer a price-taker in the bond market. It is a competitor to the sovereign, bidding for the same pool of investors who buy thirty-year duration, in a month when that pool was already absorbing heavy government supply across the United States, Germany, the United Kingdom and Japan at the same time. When two large issuers want the same buyer, the buyer sets the price.

You could watch it happen in the order books rather than in the spreads. Coverage on hyperscaler bond issues, the ratio of orders to bonds offered, fell from roughly five times in February to below two times in July on Apollo's series. Those two endpoints are not measured the same way, and the difference matters: February is a peak-demand reading (Oracle drew $129bn on $25bn) while July is a final book (Amazon closed at $41bn on $25bn). Amazon's July deal peaked at $62bn on $25bn, or 2.5 times, before banks trimmed the spread and orders fell back to $41bn. On a like-for-like peak basis demand roughly halved rather than collapsing by two-thirds. Spreads barely moved, because a deal that clears is a deal that cleared. What changed was how thin the bid had become, and a two-times book is a warning that the next deal prices wider or shrinks.

Oracle became the market's chosen instrument for expressing this. Having raised $43bn of debt and $5bn of equity in its 2026 fiscal year, with roughly $40bn of combined debt and equity planned for fiscal 2027, its five-year credit default swap reached a record 203 basis points on 24 July, having more than quadrupled since mid-2025 and passed its 2008 peak and became the standard proxy for AI funding risk. The equity market took its cue from the credit market's price of one name.

In 1994 the Federal Reserve took the funds rate from 3% to 6% inside twelve months, and the thirty-year Treasury yield rose from about 6.35% to a peak of 8.16% in November, ending the year at 7.89%. A Fortune feature that October called it the great bond massacre and put the worldwide decline in bond values near $1.5trn. Orange County, which had levered a municipal portfolio into that curve, filed for bankruptcy that December. What the episode did not do was end the expansion or the bull market. The S&P 500 finished 1994 roughly flat on price (−1.5%), then compounded for five years.

In 2000 the damage looked similar at the index level at first and was completely different underneath. The Nasdaq Composite reached a closing high of 5,048.62 on 10 March 2000 and eventually lost about 78% into October 2002, because telecommunications capex had been built against demand that never arrived. Fibre went dark. Utilisation collapsed. The revenue line, not the discount rate, was the thing that broke.

July 2026 has 1994's shape and none of 2000's evidence. The shock arrived through the long end rather than the income statement, and it arrived while the demand series were accelerating: Google Cloud revenue grew 82% year on year, and the backlog rose $50bn in a single quarter to $514bn. That is the opposite of dark fibre. It is a customer queue getting longer while the cost of serving it goes up.

One disanalogy cuts against us, and it is the important one. In 1994 the shock had an author. Alan Greenspan was raising rates, so the market could price when he would stop, and the moment he did, in early 1995, the repair began. July 2026 had no author. The Federal Reserve was stationary and the curve moved on its own, which means there is no reaction function to front-run and no obvious signal that the tightening is finished. A market shock without a central bank behind it can persist longer than a policy cycle, because nothing has to relent.

Line the tape up by what each asset needs from the capital market and July stops looking chaotic. Assets that generate cash and hold contracted revenue did well: commodities up 12%, value up 3.8%, American real estate up 2.6%, the FTSE All-Share at an all-time high on energy and financial weightings. Assets that consume capital did badly, in proportion to how much they consume. The equal-weighted S&P rose 1.0% while the capitalisation-weighted index fell, which is the arithmetic signature of damage concentrated in the largest, most capital-hungry names.

Sterling, euro and yen investors had a materially different month from dollar investors, because the dollar fell 1.31%. The MSCI World ex-US index rose 2.1% in dollars against the S&P's small decline, and the rotation out of America was as much a currency event as an equity one.

Beth Hammack of Cleveland, Neel Kashkari of Minneapolis and Lorie Logan of Dallas each voted on 29 July to raise the federal funds target by a quarter point. Three dissents in one direction is the first such split since September 2016. The company is small but Warsh is not alone in it: Arthur Burns met three dissenting votes at his first policy decision in February 1970, and Paul Volcker four at his second in September 1979. Kevin Warsh declined to argue with it. Asked by the Financial Times to characterise their case, he said he would not give their best arguments, and added: "So you're right, I asked for a good family fight, and I got one. That's the purpose, that's the design feature."

Behind the vote sat a committee that has stopped pretending the last five years did not happen. "There is no soft inflation target, there is no soft implicit target," Warsh told the room. "There is only a target, and it is 2 percent." He had opened by noting that five-plus years of above-target inflation "cannot be cured in nine weeks, or by a single month of modest price decreases," which was a direct reference to June's CPI print of −0.4% on the month.

The rest of the world's anchor banks held with him. The European Central Bank left the deposit rate at 2.25% on 23 July after June's hike, the Bank of England held at 3.75% on a 6–3 vote, and the Bank of Japan stayed at 1.00%. The tightening that did occur came from the periphery: Korea raised 25 basis points to 2.75% and New Zealand raised 25 basis points to 2.50%, its first increase in three years. Indonesia surprised by holding at 5.75%.

What makes this a policy story rather than a procedural one is the second half of Warsh's opening. He described the surge in high-tech capital spending as "remarkable," put four-quarter growth in AI-related equipment and software at nearly 20%, and then asked whether the resulting rise in memory and logic chip prices indicates "a broader inflationary dynamic." A central bank that has begun treating the AI build-out as a source of price pressure has a reason to keep rates high that has nothing to do with the labour market, and every reason not to relieve the very funding cost this issue is about.

MSCI China rose 9.0% in dollars in July, Hong Kong-listed Chinese equities gained roughly 13%, and the internet and software funds that hold the country's AI complex did better still, with the KraneShares CSI China Internet ETF up 16.4% and the iShares China Large-Cap ETF up 15.5%. In the same weeks MSCI Korea fell 17.9% in dollars and the Kospi triggered a market-wide circuit breaker in consecutive sessions on 28 and 29 July, a first for that benchmark (the KOSDAQ had done so in 2008 and 2020).

Two things drove the divergence and they are connected. The first was product. Moonshot AI released its Kimi K3 large language model, and Chinese cloud and software equities rallied on the demand it implied for domestic inference capacity. The second was supply. Reuters reported that China had begun mass-producing deep ultraviolet lithography machines, the tools that make advanced chips, and ASML fell more than 8% on 27 July. The word carries less than it sounds: the maker is targeting roughly five tools in 2026 and twenty in 2027, against ASML's roughly 130 immersion systems a year, and the machines are immersion DUV still dependent on Japanese components. The report drove 28 July; SK Hynix, whose record quarter still missed consensus, drove the 29th. For Korea, whose index is effectively a leveraged position on selling memory into someone else's build-out, that combination is an existential re-rating rather than a valuation one.

Set against this issue's thesis, China's month is the clean confirmation. The Chinese AI trade is models, inference and software sold into a domestic market, funded largely by corporate cash flow and state direction rather than by dollar bond issuance. It carries the option on AI without carrying the funding risk, and in a month when the world repriced funding risk it was bid.

The macro underneath was poor, which makes the equity move more telling. China's second-quarter GDP grew 4.3% year on year, down from 5.0%, the weakest since the fourth quarter of 2022 and the first time the government has missed its growth target since the pandemic. Investors bought Chinese technology in July despite the economy, not because of it.

Every risk dashboard on every desk we know carries high-yield option-adjusted spread as its credit tripwire. In July that tripwire moved nine basis points, from 275 to 284, and told its owners that nothing was happening. Meanwhile the largest borrowers in the corporate bond market watched their order books halve twice over. Both facts are true, and only one of them was useful.

Here is the argument, pressed as hard as we can press it: in a primary market dominated by a handful of enormous, high-quality issuers, the spread is a lagging indicator by construction. A deal gets marketed, the syndicate reads demand, and if the book is thin the deal is cut in size or pulled rather than priced at a level that would print a wider spread. The clearing price stays respectable because the clearing quantity absorbed the stress. Coverage falling from roughly five times in February to under two times in July is the same information the spread would have given you weeks later, and it was available in real time to anyone reading new-issue commentary.

This is not a subtlety. It is the difference between owning the AI complex through the last week of July and not. The equity market worked it out and sold the borrowers. The credit market's headline number said all clear, and the credit market's actual plumbing said the marginal buyer of thirty-year corporate paper had gone quiet.

Two facts settle the argument better than the aggregate does, and one of them needs a caveat first. Amazon brought $25bn in July and closed with a final book of $41bn, 1.6 times covered. The headline five-times figure from February is a peak-demand reading rather than a final book, so the two are not like for like; measured consistently on peak demand, Amazon's July deal drew $62bn on $25bn, or 2.5 times, against Oracle's 5.2 times in February. Demand roughly halved. That is a smaller claim than the raw endpoints imply and we would rather print it than the flattering version. The second fact is the control. Across the same span, on Fortune's reporting of the same underlying data, the coverage ratio for the broad investment-grade market slipped by only about half a turn. Hyperscaler demand did not fall because bond demand fell; it fell because this borrower, at this size, exhausted its buyer.

The obvious objection is that a two-times book is still a book, and it is a fair one. Deals cleared. Nobody failed. But coverage is a gradient, not a switch, and the direction of travel across five months is the signal. When your marginal buyer is thinning while your issuance calendar is growing, you find out where the bid actually is on the day you cannot afford to shrink the deal.

Put the coverage ratio on the dashboard next to the spread. It is published, it is free, and in July it was six weeks early.

The Bank for International Settlements put the transition in a single sentence in July, the most consequential sentence written about this cycle: Alphabet, Amazon, Meta, Microsoft and Oracle have shifted from funding capital expenditure with operating cash flows to issuing debt. The same work records the five largest hyperscalers spending more than a trillion dollars on AI-related capex across 2025 and 2026, with commitments outpacing both earnings and free cash flow.

The 2026 numbers alone are worth restating slowly. Alphabet guided to $195bn–$205bn, having raised the range from $180bn–$190bn a quarter earlier. Amazon sits near $200bn, Microsoft near $175bn for the calendar year, and Meta at $130bn–$145bn. Four companies, more than $700bn, in twelve months. For scale, that is a private capital-formation programme larger than the annual gross fixed investment of most G20 economies.

What is being bought is not speculative in the way the word bubble implies. Alphabet's cloud revenue grew 82% year on year to $24.8bn in the June quarter and its backlog rose $50bn in three months to $514bn, which is a contracted queue rather than a hope. Kevin Warsh, from the other side of the table, described four-quarter growth in AI-related equipment and software at nearly 20% and called the surge remarkable. Demand is present, measurable and accelerating.

The problem is arithmetic rather than belief. A programme this size, growing this fast, against cash flows that grow more slowly, closes the gap by issuing securities. That converts a technology decision into a capital-markets decision, and it hands a veto to a constituency the sector has never had to please: the people who buy long-dated corporate paper. In July that constituency exercised the veto, and Alphabet's own experience on 22 July is the cleanest illustration available. It raised its spending, reported excellent demand, and its shares fell about 5%. The market had stopped reading capex as a signal of confidence and started reading it as a funding requirement.

The order matters because stress does not remove a funding requirement, it relocates it. A programme that cannot place enough public bonds at an acceptable price does not shrink by $200bn. It moves down the stack, into equity it would rather not issue and into structures that do not appear on the balance sheet at all.

Sell the spare capacity. Meta was reported in July to be planning a cloud business to sell excess AI compute to external customers. Read as strategy that is a growth initiative. Read as finance it is a company converting a depreciating fixed asset into contracted revenue, which is what you do when the marginal dollar of debt has become expensive.

Pre-fund and stop. Oracle raised $43bn of debt and $5bn of equity in fiscal 2026, plans roughly $40bn of combined debt and equity in fiscal 2027, and said in June that it expects no additional debt issuance in calendar 2026. That is a borrower who has already been to the well and has decided not to test it again this year.

Phurichai Rungcharoenkitkul's paper, Working Paper 1367, calls the build-out one of "among the largest technology-driven investment booms in U.S. history" and then does the thing most commentary avoids, which is to put a number on the excess. Using balance-sheet and deal data in a calibrated model, it finds over-investment of around 1.5 times the efficient level, rising toward three times in the scenario where demand for AI services turns out to be less elastic than the builders assume.

The mechanism is a race. When several well-capitalised firms each believe the winner takes a durable share of a very large market, each has a private incentive to build ahead of demand, and the sum of individually rational schedules exceeds what the market needs. Add debt financing and circular equity ties between suppliers and customers and the overshoot gets larger before any participant is forced to notice.

Two features make this land harder than the usual bubble commentary. It comes from the institution whose job is telling central banks what to worry about, and it arrived on 14 July, in the middle of the month, ahead of the worst part of the semiconductor drawdown. The bond desks read it, and the timing shows in the tape.

Now the objection, which we hold seriously. An over-investment estimate is a statement about a counterfactual efficient level that nobody can observe, and it is highly sensitive to the assumed elasticity of demand, which is exactly the parameter in dispute. A 1.5-times overshoot against an unobservable benchmark is not the same class of claim as a measured leverage ratio. Every large infrastructure build in history has looked over-built at the point of construction and merely early in retrospect, and railways, fibre and cloud each carried the same critique.

The paper's real contribution is not the multiple. It is the observation that the financing structure now amplifies the overshoot, because debt-funded racing does not self-correct the way equity-funded racing does. That part we accept without reservation.

Start with the simplest test. If the cost and availability of capital had genuinely turned against the AI build-out in July, capital should have been harder to get in July. It was not. Global venture funding hit $65bn for the month, double the level a year earlier, and fourteen companies raised billion-dollar rounds, the most ever recorded in a single month. Nvidia put a reported $5bn into Safe Superintelligence. SambaNova closed a first tranche of $1bn at an $11bn valuation. Joulent took $1.75bn for energy infrastructure. This is not what a funding squeeze looks like.

Now the listed market. High-yield spreads widened nine basis points on the month and investment-grade by about two. The VIX finished July at 15.99, near the year's lows. Every large deal that came to market cleared. Not one issuer failed, defaulted or postponed publicly. The desk will point at Amazon's 1.6-times book and at a median issuer spread of 40 basis points against 30 last year, and those are real. They are also small, annual rather than monthly, and entirely consistent with a large borrower paying a modest premium for size. A discount for supply is what any market charges when one name brings a lot of paper, and it is not a squeeze.

And then the fact that should trouble this desk most: SOXX took a record $6.9bn of inflows during the index's worst month since 2008. Investors did not flee the crowded corner. They added to it, in size, while it fell. That is the behaviour of a market with abundant capital and a positioning problem, not one with a capital problem.

The simpler explanation for July needs no bond market at all. The semiconductor index had more than doubled in the first half of 2026 and set a record on 22 June. It was the most crowded, most levered, most momentum-driven position on earth. Crowded positions mean-revert violently and they do not require a macro cause to do it. Almost half of July's sessions moved 4% or more in one direction or the other, and every session had a 2% range. That is the signature of forced position management, not of a considered repricing of long-run funding costs.

Which leaves this column with an uncomfortable conclusion about its own magazine. We observed a rate move and a chip selloff in the same four weeks and built a causal chain between them, because a causal chain is more publishable than a crowded trade unwinding on its own schedule. The rate move was real. The selloff was real. The arrow between them is the part we assumed.

On 1 July the Nifty IT index printed a fifty-two-week low at 25,699.10. Over the following four weeks it rose 16.8%, recovering more than 21% from that low, and closed out its strongest month in six years. The broader market went along quietly: the Nifty 50 gained 2.2% to 24,383.60 and the Sensex 2.1% to 78,094.64, the first back-to-back monthly gains of 2026.

In Issue 02 this desk recommended the opposite. Our fourth trade was long India's power, grid and packaging assets and underweight IT services, on the argument that artificial intelligence was doing structural rather than cyclical damage to the business of selling software labour. We grade that on page 42. The reason it failed is the reason this issue exists, so it belongs here first.

The market in July was not asking which businesses AI helps. It was asking who has to raise money to participate. On that test Indian IT services scores better than almost any asset in the world. It sells the application and integration layer, funds itself from operating cash flow, carries negligible net debt, holds no fabs, signs no multi-year power contracts, and issues no thirty-year paper. When the cost of long-dated capital rose sharply, capital rotated toward the version of the AI trade that does not consume any.

The flows corroborate it. Foreign portfolio investors turned net buyers of Indian equities for the first time in four months, and total FPI inflows including debt reached ₹40,031 crore, the strongest month of 2026. Indian coverage attributed the IT rally to earnings and valuation, both of which mattered. What that coverage largely missed is that the money arriving was rotating out of Korean memory and Taiwanese foundry, and it was doing so on a funding argument rather than a growth one.

The results themselves were solid rather than spectacular. Tata Consultancy Services reported a total contract value of $9.5bn for the quarter and disclosed annualised AI revenue of $2.6bn, a figure that matters mainly because it is now large enough to report separately. The print landed after the close on 9 July; the sector fell about 1% that session and rallied 3.5% on the 10th. Infosys, LTIMindtree, Wipro and Oracle Financial Services followed, and the sector put together four consecutive up sessions in the second half of the month.

The reaction was disproportionate to the news, and that is the informative part. A sector trading at a fifty-two-week low on the first of the month had priced a structural de-rating, which is to say it had priced our June thesis. Confirmation that revenue was stable and that AI was showing up as a revenue line rather than only as a threat was sufficient to force a repricing of that assumption.

There is a real analytical question underneath, and the July numbers do not settle it. AI genuinely does compress the effort required to deliver a unit of software work, and in a business that has historically billed for effort that is a threat to pricing. What TCS's disclosure shows is that the same technology is currently generating enough new mandates, in migration, integration and model deployment, to more than offset the compression. Whether that holds once the migration wave is complete is unknown, and the current order book cannot tell you.

What can be said now is narrower and more useful: the compression argument is a multi-year thesis and the market was pricing it as a this-quarter fact. July corrected the timing error rather than the argument.

The rupee closed July at 95.35 to the dollar, having ended June near 94.6, and it gained 15 paise on the final session on the back of the inflows. Over the month it lost ground. That is a genuinely awkward combination, because the dollar index itself fell 1.31% in July: sterling, the euro and the yen all had a good month against the dollar, and the rupee did not.

Oil explains most of it. Brent rose from about $73 at the end of June to roughly $92 at the end of July, and WTI gained 20.6%. India imports the overwhelming majority of the crude it consumes, so a move of that size lands directly on the current account within weeks. June's issue argued that a falling oil bill was buying the Reserve Bank room to stay on hold while the rest of the world tightened. July took that room back.

The equity flow numbers, meanwhile, do not reconcile, and we are printing the disagreement rather than choosing. Several outlets reported foreign portfolio investors buying ₹20,200 crore of Indian equities in July; others reported ₹15,157 crore, from what appears to be the same depository data. The gap is most likely a treatment difference between primary-market and secondary-market purchases, but we cannot establish that from the public releases and we are not going to assert it. Total inflows including debt were ₹40,031 crore, the best month of the year. For the calendar year to date, foreign investors remain net sellers of roughly ₹2.54 lakh crore.

So July was a good month for Indian assets and a mediocre one for the Indian currency, and both facts trace to the same source. The capital came because the AI rotation favoured India. The currency lagged because the energy shock that drove the rotation is also India's largest import.

The benchmark ten-year rose to 6.85%, which by the standards of the month is a modest move. The American thirty-year added 36 basis points and the Bund 35; India's 17 was less than half of that, and it arrived without a policy meeting, since the Reserve Bank's committee did not sit in July.

Three things did the work. Crude was the largest: Brent moving from about $73 at the end of June to roughly $92 at the end of July feeds directly into India's inflation and current-account arithmetic, and the bond market prices that before the data confirms it. (Source [27] frames the same move from $77, an early-July level rather than the June close; we use the June close throughout for consistency with the Tape and the Scorecard.) Banking-system liquidity tightened, which is the mechanical constraint on how much paper the domestic bid can absorb. And on 31 July the index provider deferred India's entry into the Bloomberg Global Aggregate Index for a second time, removing an expected source of passive foreign demand. India's FAR bonds have been in Bloomberg's emerging-market local-currency index since 2025; the Global Aggregate applies a higher operational bar, and the stated reason was that the reforms need to be more firmly established in day-to-day market practice.

That last one deserves attention because it is a pure flow event with no economic content. India spent two years reforming its bond market to attract exactly this kind of index-tracking money, and the Fully Accessible Route has been delivering: on Bank of Baroda's estimates, net inflows of $2.3bn in June followed by $864mn in the first three weeks of July. Other series put the July run-rate differently, and we have not been able to reach the primary tables to adjudicate. A delayed inclusion does not undo those reforms, and it does postpone a large tranche of price-insensitive buying that the market had already discounted into the curve.

For this desk the page is uncomfortable, because Issue 02's third trade was long Fully Accessible Route government bonds at around 6.7% on a thesis that included a falling oil bill. The oil bill rose 20% instead and the yield went to 6.85%, which means the position lost money on the rate leg. We grade it on page 42.

The two cases are separable by a test that runs over the next two quarters. A squeeze gives back most of its gain once the selling pressure that created it is exhausted, and it does so without any change in the fundamental series. A re-rating holds its multiple and shows up as sustained foreign buying and a growing disclosed AI revenue line. Watch those two series, not the price.

Fourteen companies raised rounds of $1bn or more in July on Crunchbase's count, the highest in a single month on that tracker, and total global venture funding reached $65bn, double the level of a year earlier. Other trackers count materially fewer billion-dollar rounds; we use Crunchbase throughout and label it. Nvidia's investment in Safe Superintelligence led the table, reported by Bloomberg at about $5bn (Nvidia confirmed an investment but not the amount, and no round size, lead or valuation was disclosed), followed by roughly $2.8bn for the Chinese video-generation company Kling AI and $1.8bn for the European defence group Helsing.

Look at where the money went and the private market was buying the same things the public market was selling. Joulent took $1.75bn for energy infrastructure. SambaNova closed the first tranche of a $1bn Series F at an $11bn valuation, led by General Atlantic, which is an AI chip company raising a billion dollars in the month AI chip equities fell 21%. Together AI raised $800mn in a Series C led by Aramco Ventures, with Vista Equity Partners and Nvidia participating, putting sovereign energy money and a chip supplier on the same cap table.

Two readings are available and we cannot choose between them on July's evidence. The first is that private capital is slower and stickier than public capital, that these rounds were negotiated weeks or months before they were announced, and that the marks will follow the listed complex down with the usual lag. The second is that private investors are correct, that the public drawdown was a positioning event, and that the people closest to the technology kept funding it because the demand they can see has not changed.

What we will not do is pick the reading that flatters this issue's cover. A funding squeeze that coincides with the largest month of venture funding on record is a funding squeeze with a serious hole in it.

A conventional sector framework had a bad month. It put Korean memory, American semiconductor equipment, Chinese internet software and Indian IT services in the same bucket, and that bucket contained the month's worst performer at −17.9% and its best at +16.8%, nearly 35 points of spread inside one sector label. Any risk system built on sector exposure was blind to the largest single driver of July's returns.

The framework that worked sorts holdings by their relationship to the primary market. At one end sit businesses that fund growth entirely from operating cash flow and carry no material capital programme. At the other sit businesses whose stated plans require them to raise substantial new money within twelve months. In between sit the suppliers, who carry their customers' funding risk without holding their customers' contracts, which is the worst position to occupy and the one Korea occupied.

We are running the book on that axis. Overweight the self-funded application layer and contracted real assets, underweight the externally funded build-out and its supplier tier, and hold the energy exposure that the same rate environment is repricing upward. The hedge is credit, because credit has not yet moved and equity has.

One caution about the whole construction. It works while the long end stays elevated and the primary market stays selective. Should the thirty-year fall back toward 4.90%, where it started July, the sorting variable loses its power and the trade reverses quickly, because everything we are underweight is the highest-beta expression of falling rates.

The three paths are separated by two observables that publish on different clocks. Order-book coverage on the next jumbo investment-grade technology deal arrives within days of it being marketed and distinguishes absorption from a break. The August and September payroll prints arrive on a fixed calendar and distinguish absorption from reversal.

Note what is absent from that list. Neither an inflation print nor a Federal Reserve decision separates these paths, because July demonstrated that the policy rate is no longer the binding variable. A committee that holds while the curve moves 36 basis points has already told you where to look.

A band rather than a point estimate is deliberate: we are claiming a range we can be graded against, not a precision we do not have. Over enough issues the test is calibration rather than accuracy. Of the events we assign 20–30%, roughly a quarter should happen. If far more or far fewer do, the desk is miscalibrated and the December issue will say so in a single table.

Three of these are deliberately uncomfortable. O5 would hurt the book badly and we still put it at 35–45%, because positioning exhaustion is a real force and our Against the Grain column makes that case at full strength on page 26.

The whole of this issue rests on the long end staying where July left it. Take that away and every conclusion inverts, and the case for taking it away is stronger than the cover admits.

Start with the economy the yields are supposedly pricing. American second-quarter GDP came in at 1.5%, down from 2.1%. June payrolls added 57,000 against expectations near 110,000, roughly half. Headline CPI fell 0.4% on the month to 3.5% year on year, with core easing to 2.6%. Eurozone inflation cooled to 2.8% and British inflation to 2.6%. That is a picture of decelerating growth and decelerating prices, and it is a strange foundation for the sharpest inter-meeting rise in long-term yields in two decades.

Which suggests the move was technical. A month with a war-driven oil spike, heavy summer sovereign supply, thin August-adjacent liquidity, and an unusually large corporate issuance calendar can produce a large term-premium move that has nothing to say about the next two years. Every one of those conditions is temporary. September brings the seasonal return of duration buyers, and if inflation keeps falling while growth keeps slowing, the front end starts pricing cuts and pulls the whole curve with it.

There is also a political economy point we take seriously. A Federal Reserve chairman who has removed forward guidance and welcomed dissent has, by design, less control over the curve than his predecessors. That cuts both directions. The same machinery that produced a top-decile tightening with no policy change can produce a top-decile easing with no policy change, and it can do it on a single soft payroll print. Our entire book is a bet on a variable that has just demonstrated it can move violently without anyone deciding it should.

Finally, the sequencing. If cuts arrive within two quarters, the assets most punished in July are the ones with the most to gain, and they are precisely the assets we are underweight. The semiconductor index would not need a fundamental change to rally 25%, only a discount rate that goes back where it came from.

Your entire funding argument rests on order-book coverage falling from five times to under two. That is a soft number nobody audits. Why should I underwrite anything on it?

You shouldn't underwrite on it alone, and we would not defend the precision of either endpoint. What we would defend is the direction and its corroboration. Three independent things moved the same way in July: coverage thinned, Oracle's five-year credit default swap went to a record 203bp in late July, and equity sold the companies with the largest external funding requirements while buying the ones with none. Any one of those is weak evidence. Together they describe a market repricing access to capital.

Credit spreads widened nine basis points. If funding were really the problem, credit would have moved. Isn't the simplest reading that you are wrong?

It is the strongest objection and we have given it a full page of its own on page 26 rather than burying it here. Our honest position is that we cannot resolve it. The reconciliation we find least implausible is that primary-market stress shows up in quantity before it shows up in price, because a thin book gets managed by cutting deal size rather than by paying up. That is a hypothesis, not a finding, and if credit is still calm in October with coverage restored, we were wrong.

You were short Indian IT a month ago and you are long it now, after a 16.8% move. Why is that not performance-chasing dressed as analysis?

Because the reason changed, not just the price. In June we were short on a claim about AI destroying the services billing model, which is a multi-year argument we still hold and which July did nothing to settle. We are long now on a different and narrower claim: that in a high-long-rate regime the market pays a premium for AI exposure that carries no funding requirement. Those two views can both be true, and page 29 says plainly that we are not netting them. If you have a ten-year horizon the compression argument still applies to you.

What single piece of evidence would make you abandon the cover thesis?

A jumbo technology bond deal covered four times or better in the September window, priced at a normal concession, while the thirty-year sits above 5.20%. That would show the funding channel is open at the current level of rates, and the entire mechanism we have described would be dead. We have written that into the trades as an invalidation on page 39 rather than leaving ourselves room to reinterpret it.

If you are right, is there anything to do that is not simply "own less of the thing that fell"?

Yes, and it is the more interesting half. The sort is available in credit and in private markets, where it has not happened yet. Public equity repriced funding structure inside four weeks. Private marks have not moved at all, and a record fourteen billion-dollar rounds were written into the drawdown. If the public market is right, private AI infrastructure valuations carry a stale mark, and the sensible action for an allocator is at the pacing and secondaries level rather than in listed equity.

Last one. Does any of this matter if the Fed cuts in September?

No. It matters much less, and we have put that at 15–25% for a hike and given the reversal path 25–35% on page 40. A cut would take the long end down, collapse the funding premium, and make this issue a description of a month rather than a guide to a regime. We would rather tell you that than pretend the thesis is robust to its own main risk.

Three lines to read. Financial Conditions fell 3.14%, the family's clearest early warning and a tightening signal recorded before July's curve move. Capital Realisation fell 3.72%, the sharpest monthly drop, as exits got harder against a rising discount rate. Capital Formation kept climbing to 159.4, which is the build-out still committing while the cost of funding it rose.

Verdict: read the financing sections, treat the headline multiple as directional.

The argument. The AI build-out ranks among the largest technology-driven investment booms in American history, and a calibrated model of competitive racing puts investment at roughly 1.5 times the efficient level, rising toward three times where demand proves less elastic. The mechanism is that each well-capitalised firm rationally builds ahead of demand to secure a winner-takes-most position, and the sum exceeds what the market needs.

The flaw. The multiple is measured against an unobservable efficient level and is highly sensitive to the assumed demand elasticity, which is precisely the parameter in dispute. It is a modelling result, not a measurement, and the paper is more careful about this than its coverage was.

The takeaway a desk can use. The durable contribution is the observation that debt financing changes the character of the overshoot, because debt-funded racing does not self-correct the way equity-funded racing does. That reframes AI capex from a valuation question into a credit question, which is what July's tape then confirmed.

Verdict: the official-sector baseline; read it for the concentration framing.

The argument. The committee's periodic assessment of vulnerabilities in the UK and global financial system, published inside the coverage month, with attention to concentration in equity markets and the valuations attached to AI-linked assets.

The flaw. Financial stability reports are written to avoid moving markets, so the language is necessarily hedged and the concrete thresholds a risk officer wants are largely absent. It tells you what the authorities are watching rather than what they would do.

The takeaway a desk can use. Read alongside the BIS paper it establishes that two separate official institutions independently flagged AI-linked concentration and leverage in the same month. When the official sector converges, the regulatory response follows within a year, and that is a positioning input rather than a headline.

The Regime Break (Issue 02) · the June 2026 shift in which anchor central banks retired the disinflation-era reaction function, moving hawkish in concert and refusing to ease when the triggering oil shock collapsed. July's 9–3 hold with three dissents to raise is the first out-of-sample confirmation.

Higher-and-Energy-Scarce (Issue 02) · the successor regime in which rates settle high and energy is a recurring source of inflation shocks, so central banks ease only on realised disinflation. July's oil round-trip tested it and it held.

The Reaction-Function Reset (Issue 02) · the market re-pricing the response rule rather than the forecast. July extends it: with forward guidance withdrawn, the market now prices the path itself, which is how a top-decile tightening occurred with no policy change.

The Power Bill (Issue 02) · the thesis that the binding constraint on the AI supercycle is firm electricity rather than silicon. Issue 03 amends it: in July the binding constraint was neither, it was the availability of long-dated capital. The Power Bill is the medium-term constraint, the Funding Verdict the immediate one.

The Bifurcation (Issue 02) · India's split of portfolio (fast) against strategic (patient) capital. July added an axis: domestic equity funding against foreign debt funding.

Capital Formation Index (CFI) · the India two-axis gauge, part of the eleven-index family reported on p.46. June 2026 vintage: 159.4, up 1.04% on the month.

The Divergence Trade (Issue 02) · long the outlier that holds against the field that tightens. Its June expression through Indian government bonds lost money in July and is graded on p.42; the framework stands, the entry point did not.

The Lexicon is a living standard, and its value is that definitions do not drift between issues. Where an amendment is made it is stated in the issue that makes it, dated, and carried forward in that form. Generic market terminology is not defined here.

The coverage window is 1–31 July 2026. Every datapoint, quotation and transaction printed above is dated inside that window or explicitly labelled as of 31 July, with two declared exceptions. The Gauges (p.46) run on the capital-index family's June 2026 vintage, because the family publishes with a one-month lag and that is the latest the engine returns; the vintage is stamped on the page. Looking Ahead (p.45) is forward-dated by construction.

Market levels are point-in-time reads at or near the close of 31 July 2026 and are not official month-end settlements. Where two sources gave different levels for the same instrument, the Source Register names the one printed and the discrepancy is noted on the page. The US 10-year is printed at 4.75% per source 1; source 2 prints 4.74%, a difference of snapshot timing rather than of fact.

● denotes a filed, official or primary-source figure. ◙ denotes a figure constructed by Gravitywell Research from sourced inputs, including the repricing decomposition on p.11 and the rate sensitivity on p.25. ◯ denotes an estimate or forecast, including the $570bn 2026 AI issuance run-rate and the Brent intraday peak derived from the 27 July close.

Four items are carried as unresolved rather than smoothed over. India's July foreign equity inflow is reported as both ₹20,200 crore and ₹15,157 crore and we print both. The Kospi's local July decline is reported between 23% and 33% and we print only MSCI Korea's −17.9% in dollars. Several standing Indicators (US 2-year, USD/JPY, EUR/USD, GBP/USD, silver, copper, EU TTF, ether) were not verified to a citable July close before this issue closed and are shown blank rather than estimated. And we have no mechanism that reconciles record inflows into a record drawdown, which is argued against us on p.26.

Capital Axis carries desk bylines only. No article in this publication is attributed to a named individual, and the masthead on p.03 lists desks rather than people. The Conversation (p.44) is a desk device written in both voices and says so on the page; no quotation on it is attributed to any real person.

Every quotation elsewhere in this issue is verbatim from the Federal Reserve's published transcript of the 29 July 2026 press conference. This publication does not use unnamed sources, composite quotations, or paraphrase rendered as quotation.

The cover illustration was generated by an image model against a brief written by this desk, then processed to the house palette and stock. It is credited as AI-generated on the cover itself. No photograph, and no depiction of a real person, appears anywhere in this issue.

The eleven indices and the Capital Cycle Clock on p.46 are computed by the Gravitywell engine at METHODOLOGY_VERSION v3.2 (2026-07), built to the OECD/JRC composite-indicator handbook and IOSCO benchmark principles. Every print carries a drop-one-pillar uncertainty band, a data-coverage ratio and a list of pillars whose contribution cap bound. Two caps bound in this vintage and are disclosed on the page. The printed values match the engine output exactly.