A $1.3 trillion tantrum: chips sell off while Samsung prints the best quarter in its history

On Tuesday, Samsung Electronics told investors to expect about 89.4 trillion won in operating profit for the second quarter, roughly $58 billion, and about nineteen times the 4.68 trillion won it booked in the same period last year, according to the company’s preliminary guidance. It is comfortably the most profitable quarter in Samsung’s history.

The stock fell more than 6 percent at one point in Seoul trading and stayed weak into the close.

The nominal reason was revenue. Samsung guided sales to about 171 trillion won against a consensus near 172.2 trillion, a miss of roughly 0.7 percent, while beating the consensus profit figure by close to $3 billion, per Wccftech’s breakdown of the print. The market took the record, found the one line that came in light, and sold.

That is not how a market behaves when the question is whether demand exists.

Five sessions, $1.3 trillion

Samsung’s bad Tuesday landed mid-avalanche. In five sessions through July 8, roughly $1.3 trillion in semiconductor market value was erased, by Peter Cohan’s accounting in Forbes: Intel down 21 percent, Micron down 22, AMD down 8, the Philadelphia Semiconductor Index off 10.8 percent.

The trigger came on July 2, and it was two words. Meta said it planned to launch a cloud business selling excess data center capacity, and as Axios’ Matt Phillips reported, “excess capacity” is the phrase the AI buildout is not supposed to contain. The SOX dropped 6.3 percent that day. Look at who fell hardest: KLA down 12 percent, Lam Research down 9.7, Applied Materials down 10, and Corning, which sells the fiber that stitches data centers together, down 13.6. Caterpillar, which sells the machines that dig the foundations, lost about 7 percent.

Not the companies selling AI compute. The companies building the buildings.

Behind all of it sits one number. Hyperscalers have raised AI capital spending 67 percent this year, to $650 billion, and the Forbes piece names the fear plainly: that spending at that level may be unsustainable. That figure is doing the work in every model on the street. Bull cases discount it forward; bear cases ask what happens to everyone downstream if it merely flattens. So when Meta hinted it might have built too much, the market repriced the $650 billion, not the appetite for chips. Demand is the best-documented fact in the industry right now; the same Forbes piece notes second-quarter semiconductor earnings are projected to grow 131 percent.

What Nvidia’s pop actually priced

The third data point of the week is the strangest one. The Information reported that Beijing plans to let its top AI firms, including Alibaba, ByteDance and DeepSeek, buy a limited number of Nvidia H200 chips, a story Bloomberg picked up on July 8. Nvidia rose 3.7 percent on the news, per Yahoo Finance’s July 9 roundup.

Then read the fine print, compiled by TrendForce on July 9: total approvals may be capped below 200,000 chips, less than half of what Chinese firms requested earlier this year, and Beijing would reportedly restrict the H200s to model training, steering inference work onto domestic processors, including Huawei’s. Washington’s export licenses, for their part, cap each approved customer at 75,000 chips.

So: a capped, conditional, training-only allocation, confirmed on the record by neither government, covering less than half of stated demand. Worth 3.7 percent of Nvidia.

Put the week’s three moves side by side and a pattern falls out. A record profit gets sold over a 0.7 percent revenue miss. A trillion-plus dollars evaporates on a hint that one buyer overbuilt. The sector’s bellwether rallies on a rumor about permission slips. None of these are demand signals. They are prices on the two things still genuinely uncertain: policy, in two capitals, and the payback period on $650 billion a year — a period that stretches with money expensive again, futures putting zero odds on a July Fed cut and roughly one-in-four on a hike.

The trouble with “mid-cycle reset”

Morgan Stanley’s label for all this, per Forbes, is a mid-cycle reset rather than a top. Wedbush’s Dan Ives says the AI trade is in the “3rd inning, 1 out.” The street consensus, in other words: healthy digestion, cycle intact.

“Expectations are up, and fundamentals are struggling to meet these high sky-high demands,” FBB Capital’s Mike Bailey told CNBC.

The tape half-supports the reset framing. A reset implies discrimination rather than exit, and there is real dispersion in the wreckage. Micron, down 22 percent, still sits 66 percent below the average analyst price target, per Forbes’ tally; Intel, down 21, trades 8 percent above its target. Sellers are sorting names, not dumping a sector.

But “mid-cycle” quietly assumes the cycle’s engine is fine, and the engine is the $650 billion — precisely what the market spent five sessions interrogating. The hardest-hit names were the longest-duration claims on that spending holding up: the equipment makers and the fiber-and-concrete end of the trade. That looks less like expectations resetting within a cycle and more like the market re-underwriting the cycle’s one load-bearing assumption. Our read, not Morgan Stanley’s.

We are not going to tell you which of them is right; calling tops in this tape is a reliable way to be famous briefly. What has clearly changed is the burden of proof. Two years ago a chipmaker had to show that AI demand was real. This week Samsung showed a $58 billion quarter built on memory prices that, per Wccftech, rose 90 percent in the first quarter and another 50 to 60 percent in the second, with the company reportedly pushing for up to 20 percent more in the third. Nineteen times last year’s profit. It bought them a 6 percent drawdown.

The next proofs are already on the calendar, via Forbes: TSMC reports July 16, Intel on July 23. Samsung breaks out its divisions on July 30. Three more chances to print a record and get sold for it.