The four largest buyers of artificial-intelligence infrastructure spent roughly $165 billion between them in the second quarter. Over the same stretch, the exchange-traded fund that tracks semiconductor manufacturers fell 21 per cent, and the fund tracking memory makers fell 32 per cent. Both figures are accurate, and holding them in the same frame is the work facing anyone reading the market this week.
The instinct is to treat one of the two as noise. It is not. July was the month the equity market stopped treating record AI capital spending as an automatic argument for owning the companies that supply it, and started asking a narrower question: who actually keeps the money at the end of the chain. That question has a different answer for a hyperscaler than for a chipmaker, and the July tape is what it looks like when investors begin to price the difference.
What the Chip Selloff Actually Says About Demand
A 21 per cent monthly drawdown in a major sector fund is not a rounding error. Nor is a 32 per cent fall in memory names, which is the sharper of the two moves and the more informative one. Memory is the part of the semiconductor complex most exposed to inventory cycles: it is sold in volume, priced on spot markets, and historically prone to violent swings when buyers pause to work through what they already hold.
What the move does not say is that AI demand has stopped. The S&P 500 finished July essentially flat, which is not the shape of an index digesting a collapse in its largest growth theme. A flat index alongside a 21 per cent sector fall is a rotation, and rotations are about relative claims on future profit rather than the existence of the profit itself.
The distinction matters commercially. A chipmaker’s earnings depend on unit shipments and pricing power in a market where its customers are among the largest and most sophisticated procurement operations in the world, and where several of those customers are designing their own silicon. A hyperscaler’s earnings depend on what it can charge for the services the chips make possible. Those are different businesses with different margins, and July was a month in which the market treated them differently.
The Spending Is Accelerating, Not Slowing
Set against the selloff, the capital numbers are unambiguous. The roughly $165 billion that Amazon, Alphabet, Microsoft and Meta committed in the second quarter is 87 per cent above the same quarter a year earlier, and 393 per cent above the same quarter three years ago. That is not a plateau. It is a curve still bending upwards at a point where most industrial investment cycles have already flattened.
Research group Epoch AI, which tracks the series quarter by quarter, puts the average growth rate at 72 per cent a year since the second quarter of 2023, with a 90 per cent confidence interval of 66 to 78 per cent. Its combined figure for the group was $36.8 billion in the second quarter of 2023. Two and a half years later it had grown roughly fourfold. If the trend holds, the group’s projection for full-year spending runs to around $770 billion.
Numbers of that size stop being a technology story and become a macroeconomic one. Second-quarter real GDP growth came in at 1.5 per cent, and non-residential fixed investment contributed 1.15 percentage points of it. On those figures, business investment, a category now heavily weighted towards data centres, power and the equipment inside them, accounted for the majority of the quarter’s growth. An economy expanding at 1.5 per cent with three-quarters of that expansion sitting in one investment theme is a more concentrated economy than the headline rate suggests.
Why Record Margins Are Not the Reassurance They Look Like
The earnings backdrop is genuinely strong. S&P 500 second-quarter earnings are tracking growth of 47 per cent year on year, the fastest since the second quarter of 2021, and net profit margins have reached 16.7 per cent, the highest on record. On any conventional reading, that is a corporate sector in excellent health.
The complication is where the growth is concentrated. A 47 per cent aggregate growth rate driven substantially by earnings per share gains at a handful of very large technology companies tells you less about the median listed business than the number implies. It also raises a question about durability that record margins tend to obscure: margins at an all-time high are, by construction, further from their long-run average than at any previous point, and the capital being deployed to sustain them is rising faster than the revenue it currently supports.
None of that is a prediction. It is an observation about what has to remain true for present valuations to be justified: namely that the spending converts into revenue at rates that have not yet been demonstrated at this scale. Investors who sold semiconductors in July were not necessarily disputing the demand. Some were disputing the price they were being asked to pay for a fixed share of it.
The Fed Spent July Arguing About a Rise, Not a Cut
Running underneath the equity rotation is a monetary backdrop that has changed character. On 29 July the Federal Open Market Committee voted 9-3 to hold the federal funds rate in a range of 3.5 to 3.75 per cent. The three dissents all came from regional presidents: Beth Hammack of Cleveland, Neel Kashkari of Minneapolis and Lorie Logan of Dallas. All three wanted to raise rates. The post-meeting statement recorded that they “preferred to raise the target range for the federal funds rate by ¼ percentage point at this meeting”.
That is the first time since September 2016 that three policymakers have dissented with a unified view of which direction rates should move. “We’re reading this as a Committee with vocal hawks,” said Ian Lyngen, head of US rates at BMO Capital Markets. Kay Haigh, global head and chief investment officer of fixed income and liquidity solutions at Goldman Sachs Asset Management, put it more directly: “The Fed appears to be running out of patience with above-target inflation, despite recent data coming in cold.”
The inflation record explains the impatience. Core PCE ran at 3.3 per cent year on year in June, the 64th consecutive month above the Fed’s 2 per cent target. Officials favouring tighter policy have pointed to tariffs and to higher energy costs tied to the conflict in the Middle East. The committee’s own June projections had already pencilled in one quarter-point increase by the end of the year. For a business planning its 2027 financing, the working assumption should be that the next move in US policy rates is more likely up than down.
Bonds Are the Part Nobody Is Watching
The equity rotation has drawn the attention, but the bond market is where the more consequential repricing has happened. The 30-year Treasury yield ended July at 5.27 per cent, its highest since July 2007. The US bond market has now been in drawdown for six years, the longest such stretch on record, having begun when the same yield sat below 1 per cent.
That is the number that reaches the real economy fastest. Long yields set the discount rate against which every capital project is judged, price commercial mortgages and corporate refinancing, and determine the cost of the government’s own borrowing. Federal debt has risen by $3.6 trillion over thirteen months and is approaching $40 trillion. A company weighing a data centre, a warehouse or an acquisition is doing that arithmetic against a long rate that has moved a long way.
It also explains the year’s least-discussed equity statistic: value stocks are ahead of growth by more than 20 percentage points so far in 2026, on pace for their largest outperformance on record. That is what a sustained rise in the discount rate does to assets whose value sits furthest in the future.
What to Watch From Here
The labour market is the swing factor, and it is sending mixed signals. Jobless claims are at their lowest since January 2024, which argues against imminent weakness, but the Fed’s own statement noted that job growth has “kept pace with the workforce and the unemployment rate has changed little” even as the labour force has contracted. A steady unemployment rate produced by a shrinking denominator is a different thing from a strong one. The current expansion is 74 months old.
For business readers the practical question is not whether to hold semiconductors. It is what the July split implies for planning. Three things follow from it. Capital is still flowing into AI infrastructure at a rate no other investment theme is matching, so supply-chain and property demand tied to it is unlikely to soften in the near term. Financing has become materially more expensive at the long end, and the central bank is debating whether to make it more expensive still. And the equity market has begun distinguishing between companies that spend on the theme and companies that sell into it, a distinction that has already cost the second group a fifth of its value in a single month.
Chairman Kevin Warsh, in his first meeting in the chair, has made clear he intends to say less about where rates are going and more about the conditions that would move them. On the evidence of July, the conditions are the ones worth watching.
This article is reporting on market and corporate developments and does not constitute investment advice.
About The Author
More Stories
The Construction PMI Jumped to 44.7 and That Is Still Contraction
Funding Circle Nearly Quadrupled Its Profit Lending to Small Firms
Warm Weather and Promotions Lifted Retail Sales 1% in June