The AI Trade Meets the Cost of Money This Week

Two Dates on This Week's Calendar

Most weeks do not deserve a preview. This one does, because two unrelated events land within forty eight hours of each other and between them they set the terms for the rest of the year.

On Wednesday, Nvidia reports quarterly results. On Friday, Kevin Warsh delivers his first Jackson Hole keynote as chair of the Federal Reserve, at a symposium that runs from Thursday through Saturday.

One is a company. The other is a central bank. They have nothing to do with each other, and they are going to be read as a single story, because the market's two open questions right now are whether the artificial intelligence buildout is worth what is being spent on it and whether interest rates are going up rather than down. Wednesday speaks to the first. Friday speaks to the second.

Where the Market Is Standing

The setup matters, because the same news lands differently depending on what is already priced.

Earlier this month the S and P 500 closed at 7,798.99, its twenty seventh record close of 2026, and the index is up about 13.4 percent on the year. The Dow crossed 54,000. By any conventional read, this is a market in good health.

The detail underneath that is more interesting. The Nasdaq is roughly 2 percent below the record it set back in early June. An index heavily weighted toward large technology companies has spent the summer failing to make new highs while the broader market kept making them.

That divergence is the whole reason Wednesday matters. Money has been rotating out of the most aggressively valued technology names and into everything else, and it has been doing so since roughly the start of the summer. The market has been quietly asking whether the AI trade has run ahead of itself, and it has not received an answer.

Wednesday: One Company Reporting for an Entire Industry

Nvidia designs the chips that most large artificial intelligence systems are trained and run on. Its results have become the single best available read on whether the industry's spending plans are real.

This is unusual and slightly unhealthy. Nvidia sells to a small number of enormous customers, so its revenue is a direct measurement of what those customers are actually buying, reported before most of them say anything themselves. A supplier has become the industry's disclosure mechanism.

The interesting question on Wednesday is not whether the quarter was good. The interesting question is what the company says about the quarters after it, because a supplier's forward guidance is a statement about its customers' commitments.

The Number That Actually Matters

Here is the figure to hold in mind, and it is not Nvidia's.

The four largest cloud computing companies, Amazon, Microsoft, Alphabet and Meta, have guided to combined capital expenditures of roughly 725 billion dollars for 2026. That is up something like 77 percent from about 410 billion dollars the year before, and last year was itself a record. Amazon alone is pointing at roughly 200 billion dollars. Microsoft is tracking near 190 billion. Alphabet has guided to a range of 175 to 185 billion, and Meta to 115 to 135 billion.

In the June quarter the four of them spent about 166 billion dollars between them, a figure up sharply from a year earlier and up materially from the March quarter alone. Alphabet by itself reported around 44 billion dollars of capital expenditure in the quarter, bringing its trailing twelve month spending to roughly 132 billion.

Numbers at that scale need an anchor. Capital expenditure means money spent on physical assets that last for years, which here means data centers, land, power infrastructure and the chips inside. It is not an operating cost that flexes with demand. It is a commitment made years before the revenue it is meant to produce.

It is also increasingly a commitment that runs into physical limits rather than financial ones. A data center needs electricity, transmission capacity, cooling and a site, and none of those can be ordered at the speed capital can be raised. Grid connections are queued years in advance in many regions, and turbine and transformer lead times are measured in years rather than months. That constraint cuts both ways for a supplier like Nvidia. It can slow the pace at which chips get installed and paid for, and it also means the spending already committed is unusually hard to cancel, because much of it is contracted well ahead of delivery.

The Loop That Makes Investors Uneasy

There is a structural feature of this arrangement that deserves attention, and it is the source of most serious skepticism about the trade.

A large share of Nvidia's revenue comes from a handful of cloud companies. Those same cloud companies are, in various arrangements, also investors in or suppliers to the artificial intelligence firms that rent the resulting capacity. Money moves in something closer to a circle than a line, and revenue growth that depends on a small number of counterparties who are themselves spending ahead of their own revenue is more fragile than the same growth spread across thousands of independent customers.

This is not an accusation of anything improper. It is a description of concentration. The relevant question is what happens to the supplier when one large customer decides to slow down, and the answer is that it shows up immediately and visibly.

There is a related pressure. Each of those large customers is developing its own chips, partly to reduce dependence on a single supplier and partly to lower cost. Every one of them is simultaneously Nvidia's biggest customer and its most credible future competitor.

The Accounting Question Underneath

The subtler issue is how this spending flows through the income statement, because the timing creates a gap between cash and reported profit.

When a company spends billions on a data center, the cost is not deducted from profit in that year. It is capitalized and then depreciated, meaning expensed gradually across the asset's estimated useful life. If a company assumes a chip is useful for six years, it charges one sixth of the cost against profit each year. Assume three years and the annual charge doubles.

Nobody knows how long an AI accelerator stays economically useful, because the technology is improving quickly and the workloads keep changing. If the useful lives being assumed turn out to be too long, then reported profits across the industry are currently overstated, and the correction arrives later as larger depreciation charges or write downs.

This is not a hypothetical concern to the market. It is one of the reasons technology stocks stalled over the summer while the rest of the market advanced, and it is why capital spending guidance has started to be received as a risk rather than as a sign of confidence. One of the large cloud companies saw its shares sell off in late July on exactly this issue.

Friday: A New Chair's First Jackson Hole

The second event is different in kind. Jackson Hole is an annual symposium hosted by the Federal Reserve Bank of Kansas City, and its keynote has become the venue where chairs signal shifts in thinking that are too significant for a routine statement.

This one carries more weight than usual because Warsh took over as chair only earlier this year, and because the committee he leads is visibly divided. At the July 29 meeting the Federal Open Market Committee held the target range for the federal funds rate at 3.50 to 3.75 percent on a 9 to 3 vote. All three dissenters wanted a hike. That is the most dissents in favor of tighter policy since September 2016, and the most striking set of early dissents against a new chair since 1970.

Warsh has been unambiguous about the framing. In testimony to Congress on July 14 he said the members of the committee have no tolerance for persistently elevated inflation, and noted that underlying inflation over longer horizons is determined largely by monetary policy.

The next policy meeting is September 15 and 16, and futures markets have been pricing roughly a one in three chance of any move at all. Friday is the last scheduled opportunity for the chair to shape expectations before it.

What Would Count as News on Friday

The data since that July meeting has cut both ways, which is why the speech is genuinely uncertain rather than theatrically uncertain.

On the softer side, July inflation cooled, with headline at 3.4 percent and core at 2.5 percent, and payrolls fell by 23,000 with 103,000 taken out of the prior two months by revisions. On the firmer side, headline inflation is still well above the 2 percent target, energy prices remain elevated, and inflation has now run above target for more than five years.

What would move markets is not a signal about September specifically. It is anything that reveals how this chair thinks about the framework itself, particularly how much weight he puts on a weakening labor market against an inflation record that has been poor for half a decade. A chair who emphasizes the five year record is a chair whose committee is closer to a hike than to a cut.

The Complication Neither Event Solves

Sitting behind both is the bond market, which had a genuinely difficult week.

The thirty year Treasury yield reached 5.34 percent on Tuesday, the highest since 2007, with government bond markets in France, Germany and Japan hitting multi decade highs of their own. The Treasury responded midweek by doubling the size of its debt repurchase operations, and yields pulled back.

That matters for Wednesday more than it appears to. Higher long term yields reduce the present value of profits expected far in the future, and companies spending 725 billion dollars now for revenue expected later are the most exposed to that arithmetic of anything in the market. The AI trade and the long end of the bond market are connected, and the connection is a discount rate.

How to Watch the Week

The useful frame is that these two events answer different questions and the market will try to combine them anyway.

Wednesday is about whether the spending is real and continuing. Strong results with confident guidance says the buildout has further to run. Any hedging about the pace matters far more than the reported quarter, because the reported quarter is history and the guidance is a statement about several hundred billion dollars of committed spending.

Friday is about what money will cost while that spending gets paid for. A chair who sounds worried about inflation raises the discount rate applied to every one of those future profits, which is a problem for exactly the companies Wednesday is about.

The uncomfortable combination is a confident Wednesday followed by a hawkish Friday. That would confirm the spending is happening and simultaneously raise the bar the spending has to clear.

The Bottom Line

A chip company reports on Wednesday and a central banker speaks on Friday, and the two together will decide whether the summer's rotation out of technology was an early warning or a pause.

The number worth carrying into both is 725 billion dollars, the combined capital spending four companies have guided to for this year, against roughly 410 billion last year. That is the commitment the entire trade rests on. Nvidia's results are the clearest available evidence of whether it is being honored, and its guidance matters far more than its quarter.

What the Federal Reserve controls is the rate that spending gets judged against. With the thirty year Treasury yield touching levels last seen in 2007 and a committee that produced three dissents in favor of a hike last month, the cost of capital is moving in the wrong direction for anything whose payoff is years away. Both events are really about the same question, which is what a distant profit is worth today, and the answer has been getting smaller all month.