Article — "AI Trade Rotates Into Sellers as JPMorgan Lifts Capex Past $5 Trillion": Article — "AI Trade Rotates Into Sellers as JPMorgan Reportedly Lifts Capex Past $5 Trillion": Here is the contradiction most readers are misfiling as noise: BlackRock projects another $5 to $8 trillion in AI-related capex through 2030, and JPMorgan also reached back and drew a chart of 1999.
Bullish and bearish in one breath. A record spending forecast stapled to a dot-com warning.
The instinct is to treat one of those as the real signal and the other as hedging. It isn't. Both are load-bearing, and the piece you're reading exists to explain why they belong in the same sentence.
Start with the tape, because the tape is where the tension went visible.
Thursday closed with a split personality. The Dow surged 1.1%—up 560 points to 52,865—for its fourth straight weekly win, its best streak since October 2024, per Newsbytes. The Nasdaq slid 0.87% on chip and tech losses.
Zoom out and the split gets louder. The Philadelphia Semiconductor Index rose 87% in 2026 and just posted its best quarter on record, per JPMorgan.
Against that, the Roundhill Magnificent Seven ETF fell 7% from its peak, per JPMorgan. Microsoft logged its worst monthly loss since 2000, per JPMorgan.
So the suppliers are ripping while the spenders bleed. That is the divergence JPMorgan flagged, and it's the same one that has 1999 on the strategists' minds—when communications-equipment makers rallied even as the heavy capital spenders funding that buildout rolled over, roughly a year before the crash.
Here's the reframe. This looks like a bubble warning. It reads better as a product-cycle signal.
Because the industry is quietly changing phase—from building AI to using it. From training to inference.
Those are not the same workload, and they do not reward the same companies. Training is the brute-force construction of models; inference is running them in production, where the metrics that matter are latency, throughput per watt, and total cost of ownership.
Goldman put the mechanic plainly in a May note: a faster shift to inference accelerates revenue "by converting a fixed capital base into usage." That sentence is the whole game.
The hyperscalers have spent hundreds of billions building the capital base. The market is now asking the only question that matters next: who turns it into usage, and when.
BlackRock's own research dresses this up as an asset-allocation survey. Asked where to deploy next, respondents now favor energy and infrastructure over the big-tech and semiconductor names that led the first leg. That's not sentiment—that's the build-out-to-utilization handoff showing up in a poll.
Which brings us to the mechanical reason the funders specifically get sold. Not vibes. Arithmetic.
There is a line every hyperscaler is approaching: the point where capital expenditure exceeds operating cash flow. Cross it, and the company is spending more than the business is generating—funding the build from the balance sheet, betting on returns that live in a future quarter.
Oracle is already past that line, per Seeking Alpha. Amazon is sitting right on it on a trailing-twelve-month basis. Alphabet is heading there soon.
Combined AI capex from Meta, Microsoft, Amazon and Alphabet is on track to hit $725 billion this year, per JPMorgan. That is the number the market both celebrates and fears—because every dollar of it is real, and every dollar of it currently sits on the cost side of a ledger whose revenue side hasn't caught up.
Now the $5 trillion forecast and the 1999 chart stop contradicting each other.
The spending is durable. JPMorgan is telling you it goes higher, for longer, through 2030. And that durability is precisely why the spenders get punished right now—because it guarantees the crossover, name by name, before the monetization arrives to offset it.
You don't sell the funders because you doubt the spend. You sell them because you believe it.
You don't sell the funders because you doubt the spend. You sell them because you believe it.
This is where Jim Cramer's framework is directionally right and quietly incomplete. "Wall Street's now rewarding tech companies with products in high demand and punishing their customers," he said—reward the sellers, punish the buyers.
True as far as it goes. The picks-and-shovels names are winning; the ones footing the bill are not.
But "buyers versus sellers" describes the symptom, not the fault line. The real division is who converts a fixed capital base into revenue—and the tape is paying for proof of that conversion, not promises of it.
Suppliers get paid on delivery. A memory maker with demand off the charts books the revenue this quarter. A hyperscaler pouring capital into a data center books the cost now and hopes the inference revenue shows up later.
That timing gap is the entire rotation. Not fear of AI failing—impatience with the lag between spend and return.
And here's the tell that this is psychology repricing faster than fundamentals: the same names snap back on narrative just as violently as they sold off. A fundamental verdict does not reverse in a single session. A positioning verdict does.
Which is why the loudest question—bubble or rotation?—is the wrong one to be answering right now.
It's premature. You cannot adjudicate it from a divergence chart, because the divergence is exactly what you'd expect whether this ends like 1999 or ends with the spending validated.
The one thing that would actually settle it hasn't happened yet: a hyperscaler showing inference monetization catching up to the spend. A quarter where the capital base visibly converts into usage revenue fast enough to close the gap.
Until a name past the crossover line—an Oracle, an Amazon—demonstrates that conversion, the tape has no reason to stop paying for proof and start paying for promises again.
So watch the crossover names, not the semis. The chip rally tells you the build-out is real. It always was.
The verdict lives on the other side of the ledger—in the first hyperscaler that turns $725 billion of installed capital into revenue the market can see. That's the print that ends the argument.
Everything before it is just the market front-running a phase change it can feel but can't yet confirm.