When Record Profits Can't Stop the Selloff

Samsung Electronics posted an operating profit of 89.5 trillion won for the April-to-June quarter of 2026 — roughly $62 billion, a 19-fold increase on the year before. Revenue hit a record 171.5 trillion won. By almost any reading, it was a blowout. The stock fell 8% on the preliminary report in early July and lost more again after the full results in late July. The broader semiconductor complex tumbled alongside it. Within weeks, the Philadelphia Semiconductor Index, which had surged 86% earlier in 2026, had slid 29% from its peak, erasing more than $1 trillion in chip-sector market value and entering bear-market territory.

The market was not rejecting bad results. It was rejecting a valuation that had been built on the assumption that nothing could go wrong.

The bar had been raised into the stratosphere

Memory chipmakers are no longer boring cyclicals. The supply bottleneck for high-bandwidth memory (HBM), the specialised memory used alongside Nvidia's AI accelerators, turned Samsung, SK Hynix and Micron from steady earners into the most momentum-driven names in the semiconductor industry. SanDisk shares had climbed more than 570% year-to-date by the time the sell-off began; Micron was up more than 220%. At the peak, the three memory giants collectively topped a $1 trillion market capitalisation. The run-up was not modest, and valuations were stretched well beyond their historical ranges.

Then came a quarter in which Samsung's operating profit exceeded that of both Nvidia and Apple, and guidance pointed to a projected jump of 1,800%. Deutsche Bank analysts noted the results were "only" 6% ahead of estimates. The word "only" tells the whole story. After the kind of advance these stocks had enjoyed, the bar was no longer set at "good." It was set at "unreasonably perfect."

A similar pattern emerged with SK Hynix, which fell more than 10% in South Korea despite record quarterly revenue of 60.5 trillion won, and with Chinese tech names broadly. Alibaba, which had climbed 105% at one point on AI optimism, had retreated more than 50% from its 2025 peak by June, weighed down by domestic economic weakness, an AI model-theft allegation from Anthropic, and concerns that 2 trillion yuan of government data-centre spending could crowd out private cloud providers. These were separate problems — one a valuation squeeze in the U.S. and Korean semiconductor trade, the other a combination of geopolitics and structural Chinese headwinds. But both shared the same mechanism: expectations had outrun the evidence, and the correction was a matter of arithmetic.

The economics the market was actually questioning

The sell-off was not merely about stretched multiples. It exposed a genuine tension in the economics of AI infrastructure spending that investors had been willing to overlook during the advance.

Memory now accounts for roughly 35% of AI infrastructure capital expenditure, according to Bank of America. As demand for HBM surged, prices followed — which is precisely what turned memory into such a profitable business for Samsung and Micron. But the rising cost of memory has effectively become a tax on the very customers buying it. Apple and Microsoft, among others, have had to raise consumer-product prices to offset higher build costs. The question that began to trouble investors is whether demand destruction will follow in price-sensitive segments like mobile and automotive. A cyclical boom in memory prices is only as durable as the willingness of data-centre operators to absorb it.

Then there is the spending on the supply side. Samsung and SK Hynix announced a combined $554 billion investment for a new chipmaking hub in South Korea. Samsung alone is planning a second fab in Taylor, Texas, with production starting by 2030. On the demand side, the picture is impressive but not straightforward. Major U.S. cloud providers ended the second quarter with a combined revenue backlog of $2.3 trillion, according to Bank of America. Yet conversion rates are slow: Oracle estimates converting only 12% of its backlog to revenue in the coming year; Microsoft expects 30%. The gap between committed spending and realised revenue is large and growing.

On the competition front, Chinese memory chipmaker ChangXin Memory Technologies had a strong market debut, and state-owned Chinese entities began mass-producing domestic lithography machines, reducing reliance on the established supply chain. Whether this creates real overhang is debatable — Chinese firms are still years behind in advanced HBM production and face U.S. export controls on critical equipment. But in a market that had been pricing in a Samsung-SK Hynix-Micron duopoly, any crack in the assumption invites a re-examination of the entire model.

Micron's earnings report in mid-June briefly halted the decline. The company beat consensus by a wide margin — revenue of $41.5 billion against estimates of $35.9 billion, and gross margins of 84.9% — and announced 16 five-year take-or-pay contracts worth $22 billion in guaranteed commitments. The deals included pricing floors set above the company's historical best cycle peaks, effectively locking in margins through 2030. It was a powerful demonstration that AI demand was real and that the memory trade had moved from spot pricing to long-term contracting, adding stability to what had been a fiercely cyclical business. The rally did not last. The contracts proved the demand, but they did not answer the question of whether end customers could sustain the cost.

The recovery and the test ahead

By mid-August, the semiconductor complex had staged a substantial recovery. The Philadelphia Semiconductor Index was up roughly 20% from its July 29 bear-market low, heading toward a new bull market. The sell-off, which had seemed to embody an AI-capitulation thesis, was being reclassified as a healthy correction after a parabolic advance. There is a case for that reading. None of the underlying demand drivers had broken: HBM supply still cannot meet demand, hyperscaler backlogs remain enormous, and Samsung has secured long-term supply contracts with Amazon, Google, Meta, Oracle and Microsoft.

The decisive test arrives on 26 August, when Nvidia reports its second-quarter fiscal 2027 results. Nvidia, which holds an estimated 80-90% share of the AI accelerator market and has an order book worth $1 trillion for 2026 and 2027, is the market's primary bellwether. Consensus revenue for the quarter ending July 26 is approximately $92 billion, roughly 97% above the year-earlier period, with non-GAAP gross margins guided toward 75%. Nvidia has significantly underperformed the broader semiconductor index in 2026, rising just 21% against the PHLX index's 63%. The stock was trading around $217 on the week before the report, at a forward multiple of approximately 25 times earnings — slightly below the Nasdaq-100's 26x.

These are still strong numbers. But the question Nvidia now faces is not whether demand is real. It is whether the guidance it provides for the quarters ahead can justify a $5.2 trillion market capitalisation in an environment where the cost of its own ecosystem — memory, power, cooling, construction — is rising as fast as the revenue. The market that rejected Samsung's $62 billion quarterly profit is not interested in good results. It is interested in results that clear a bar that has been set by a 570% stock advance.

What the episode reveals

The deeper lesson from this episode is not that AI demand is peaking or that a bubble is bursting. It is that momentum markets have a way of converting a structural thesis into a perfection thesis. When a stock has doubled or tripled on a narrative, the evidence required to maintain the price is no longer that the narrative is true. It is that the narrative is unfolding faster than the most optimistic version of itself.

Samsung's record profit was real. SK Hynix's revenue was real. Micron's five-year contracts are real. What was not real was the market's assumption that these results would arrive without any question about the unit economics of the customers paying for them, the sustainability of capital-expenditure levels, or the competitive trajectory of Chinese producers. The sell-off was the market's way of re-introducing those questions into the pricing.

Whether the recovery reflects a rational repricing or a premature return to exuberance will be determined not by the next earnings report but by the pattern that follows. If hyperscalers continue to convert backlogs into revenue, if memory prices stabilise rather than spiral, and if capex discipline emerges, the advance was likely a correction, not a capitulation. If costs rise faster than the services AI generates, the tension will resurface — and next time, it will not require a 19-fold profit increase to trigger it.