Sandisk is up 764% in six months. That is not a typo, and it is the single cleanest number in a quarter full of them.
The S&P 500 and Nasdaq just booked their best quarter since 2020, up roughly 15% and 21%. And in the same window, the Magnificent Seven — the seven stocks that supposedly power those indices — shed more than $2.3 trillion in market value, falling about 10% in June alone.
Two facts that should not coexist. The melt-up and the selloff happened at the same time, inside the same tape.
That contradiction is the whole story. Call this an "AI rally" and you miss it entirely.
The quarter has now delivered a partial verdict. It is more surgical than anyone rooting for a clean answer wanted.
Start with the gauge that actually ran. The Philadelphia Semiconductor Index soared 81% in Q2 — its best quarter ever — against a Nasdaq 100 up 25% and an S&P 500 up 14%. The SOX is up 94% on the year, on track for its best year since the dot-com peak in 1999.
Now hold that next to the poster child. Nvidia, the largest company in the world and the name synonymous with the AI trade, is up just 4.5% year to date. That makes it the weakest stock in the semiconductor index.
Read those two lines together and the "AI rally" label collapses. The index that defines AI hardware had its best quarter in history while the stock that defines AI had one of its worst relative years.
The money did not leave AI. It migrated within it.
The money did not leave AI. It migrated within it — from the marquee names to the suppliers actually booking revenue off the build-out right now.
Where it went is the tell. The S&P 500 leaderboard for the first half is stuffed with memory and storage names — the companies booking real revenue off the build-out right now. Micron jumped 301% in six months and crossed a $1 trillion market cap. Sandisk, again, is up 764%. Western Digital, Seagate, and Intel — up 257% on a foundry turnaround — round out the top five.
"We're seeing investors follow the bottlenecks in semis, which at the moment is good for memory and good for Intel's resurgence as a foundry," said Sean Sun of Thornburg Investment Management.
Bottlenecks are cash. Memory was scarce this quarter, so memory suppliers got paid, and the market paid up for them in turn.
Meanwhile the high-beta AI generals stalled. Nvidia at 4.5%, Broadcom at 7.6%. "Nvidia and Broadcom are running into those bottlenecks, so they're not the high-beta names the way they used to be," Sun said. "Right now investors want more torque to the strongest themes."
That is the first half of the split: the market rewarded suppliers collecting cash today over the marquee names whose next leg depends on demand it cannot yet book.
The second half of the split is harsher, and it lands on the check-writers.
The Magnificent Seven lost more than $2.3 trillion in June, per the Financial Times — Nvidia, Meta, Apple, Microsoft, Alphabet, Amazon, and Tesla all down about 10%, the group's worst month in more than a year. And the SOX reportedly rose 93% over the first half in the same stretch, per the Financial Times.
Startup Fortune put the mechanism plainly: the market "has started separating the companies collecting cash today from the companies writing the biggest checks and promising the payoff later."
The checks are enormous. Microsoft, Alphabet, Amazon, and Meta are expected to spend roughly $725 billion on capital expenditure in 2026, up from $410 billion last year. Microsoft alone put its 2026 plan at $190 billion, with CFO Amy Hood pointing to rising memory and component costs.
Sit with that loop for a second. The same memory scarcity minting 300% gains for Micron is the line item inflating hyperscaler capex — and pressuring the hyperscalers' stocks. The suppliers' windfall is the buyers' cost.
Nothing about the capex is abstract. Memory, electrical gear, data centers, power, cooling — all paid for before the AI revenue shows up. Amazon and Microsoft can absorb that longer than anyone. But the market is not being unreasonable when it asks when the cash comes back.
That is the referendum the quarter half-answered. Investors will pay for AI hardware demand today. They are done extending free credit to the AI story.
CJ Muse of Cantor Fitzgerald framed the whole six months in one sentence: "The story of the past six months is the market going all-in on AI infrastructure, but now people are asking if this is sustainable and if we should be worried."
His specific worry is the load-bearing one. "The biggest concern is about whether hyperscalers will sustain and grow their investments beyond 2026." That single clause — beyond 2026 — is the fault line under an 81% quarter.
The SaaS side of the ledger got the same treatment, quietly. The quarter's software wrap was titled, bluntly, "SaaS premium valuations under pressure." The premium multiple that older enterprise software wore for a decade is compressing while capital crowds into the parts of AI that print revenue now.
Startup Fortune's read on private markets follows the same logic: non-AI software isn't being marked down for overspending on GPUs, it's being ignored because so much capital is still crowded around AI. Different mechanism, same repricing.
So the melt-up was never a broad vote of confidence. It was a sorting.
Chips that book revenue today: rewarded, to the tune of the best quarter in the index's history. Hyperscalers writing the checks: down $2 trillion-plus and asked for proof. SaaS premium: compressed. Nvidia, caught between being the story and being the supplier, went nowhere.
Which sets up the uncomfortable part. The suppliers are getting paid because the check-writers keep writing checks. The SOX at 81% is, at bottom, a leveraged bet that $725 billion of hyperscaler capex keeps flowing and keeps growing.
The warning signs are already blinking. The semiconductor index plunged 7.9% last week — its worst week since April 2025 — then swung from down 3.2% to up 3.8% in a single Monday session. Apple has been forced to raise prices to absorb memory costs. OpenAI is reportedly weighing a delay to its IPO, which Bloomberg flagged as a red flag for a major chip spender.
The sector isn't cheap on any read except Nvidia's. The SOX trades near 26 times forward earnings against a 10-year average of 19; ARM sits above 140 times, Intel near 100. That's the definition of priced for perfection with less room for error.
Here is the chain the quarter built and did not resolve. Chip revenue depends on hyperscaler capex. Hyperscaler capex depends on AI revenue arriving before the payback date moves out again.
If that payback keeps sliding — if the answer stays on the conference stage instead of showing up in financial statements — the check-writers cut. And the suppliers riding that demand, the ones this quarter crowned, are the next thing repriced.
The best quarter since 2020 didn't answer whether AI pays off. It just decided who gets paid first, and started billing everyone else for the wait.