NextFin News - SK Hynix, Samsung Electronics and Kioxia are all facing the same earnings question: is the AI memory boom a short-lived pricing spike, or the beginning of a longer regime that can survive the next round of supply growth? Samsung has already told investors its second-quarter 2026 operating profit was approximately 89.4 trillion won on sales of 171 trillion won, while Kioxia and Sandisk said on July 3 that they had started production of 10th-generation 3D flash memory at Kitakami in Japan. The market is not just waiting for three earnings prints. It is trying to decide whether the current lift in HBM, DRAM and NAND pricing is cyclical enough to fade or structural enough to reset earnings power across the memory chain.
That distinction matters because memory has always been cyclical, but the current cycle is being pulled by AI infrastructure rather than by smartphones or PCs. HBM is now a core input for AI accelerators. DRAM and NAND are increasingly tied to server buildouts, not just consumer replacement demand. That changes the demand mix and gives suppliers more leverage than they had in earlier upswings. The question for this earnings season is not whether prices have improved. Samsung’s own guidance already answered that. The question is whether those prices can hold once customers, competitors and capex plans react.
Samsung’s July 7 guidance is the cleanest hard anchor. The company said consolidated second-quarter sales were approximately 171 trillion won and operating profit was approximately 89.4 trillion won, with guidance centered in a range of 89.3 trillion won to 89.5 trillion won. That was a record quarter on the profit line and a sign that memory pricing had already improved enough to offset weakness elsewhere in the business. Samsung sits at the intersection of memory supply, foundry spending and device demand, so its numbers usually act as a read-through for the rest of the sector. When Samsung posts this kind of profit, the industry pricing environment is usually strong enough to matter even after a high bar is already set.
Kioxia’s July 3 product announcement points in the same direction. The company and Sandisk said they had begun production of 10th-generation 3D flash memory products at Fab2 in Kitakami, Japan. The detail matters because it shows the industry is still adding capacity in a controlled way rather than through a flood of undisciplined supply. For investors, Kioxia is a test case for whether AI storage demand can absorb new NAND bits without forcing prices lower. If it can, the recovery in storage pricing may prove broader than a single quarter. If it cannot, the market will quickly revert to treating the move as a classic memory upcycle.
SK Hynix is the cleanest test of the HBM thesis. HBM sits next to AI accelerators and moves far more data than conventional DRAM, which makes SK Hynix the best lens on whether AI infrastructure spending is still translating into pricing power. If its earnings show continued tightness in HBM and strong utilization, the market will read that as evidence that AI demand is still outrunning supply. If not, the sector could shift fast from scarcity premium to margin normalization.
Market Reaction Is Now Part Of The Story
The market is no longer treating these companies as pure cyclical names, but it has not fully rerated them as permanent growth assets either. That in-between status is what makes the setup unstable. Samsung shares fell 7% after the company’s record quarter, even though sales and operating profit both beat a very high bar. The message was simple: a great number can still fail to move the stock if investors had already priced most of it in.
That first-order reaction matters, but the second-order effect is more important. If investors think the memory upcycle is already mature, they will stop focusing on the current quarter and start focusing on future bit supply, capex and whether HBM demand can keep absorbing new capacity. That changes earnings from a backward-looking confirmation into a forward-looking stress test. Stocks stop trading only on whether the quarter was good. They start trading on whether the next wave of supply can still be sold at a premium.
This is why the earnings season can move more than one name. Samsung, SK Hynix and Kioxia sit at different points in the same chain, but the transmission mechanism is shared: tighter AI demand raises HBM and NAND prices, which lifts revenue per bit, which expands margins, which then invites more capex, which later pressures pricing. The mechanism is simple. The timing is not. Earnings are where investors can see whether the industry is still on the favorable part of the curve or has already passed the inflection point.
There is also a cross-asset read-through. If the companies show that AI demand remains strong and pricing discipline holds, the signal does not stop at semiconductor equities. It supports the broader case that AI infrastructure spending is still early enough to justify elevated capex by cloud and chip customers. If guidance softens, the market may start to question whether AI spending is rotating from acceleration into digestion. That would hit suppliers first, but it would also feed back into the valuation of the wider AI trade.
Is This A Cyclical Upswing Or A Structural Reset?
The answer is both, but not on the same horizon. The near-term move is cyclical. Memory prices always respond to supply tightness, inventory swings and capex timing, and those drivers will eventually mean-revert. The long-term shift is structural. AI has permanently changed the composition of demand by making HBM and high-performance storage core parts of compute architecture rather than optional upgrades. That does not remove cycles. It changes their center of gravity.
The old cycle was driven mainly by consumer devices and general-purpose computing. The new cycle is anchored more heavily in data-center buildouts and AI inference, which are less tied to handset replacement cadence and more tied to model deployment and server density. That matters because structural change can look like a cyclical boom right up until the market tests it. A demand shock can be real and temporary at the same time. The question is whether AI has simply created a better cyclical upswing or a new baseline for memory demand.
The evidence for the structural case is stronger than in a normal upcycle, but it is not yet conclusive. AI server demand is pulling on HBM, DRAM and NAND at the same time. That creates a broader demand stack than a single-product bump. It also means that price increases in one segment can reinforce demand for adjacent segments as customers redesign systems around storage and memory bottlenecks. That is a structural feature. But supply remains highly responsive, and memory remains one of the most supply-sensitive businesses in semiconductors. If Samsung, SK Hynix and Kioxia all accelerate output at once, margins can still compress even if end demand remains healthy.
Samsung said on July 7 that its second-quarter consolidated sales were approximately 171 trillion won and operating profit were approximately 89.4 trillion won.
That guidance matters because it shows how much pricing power is already embedded in current results. It is not just that volumes are rising. It is that prices, mix and utilization are improving together. When memory businesses get all three at once, earnings usually rise much faster than revenue. The problem is that the same outcome attracts new supply. The thing that makes the quarter look exceptional can also set up the next downturn if the industry overbuilds.
The strongest counter-thesis is that the market is overreading a classic semiconductor cycle. AI demand is real, but it can still slow if cloud customers pause capex, if server deployments digest a burst of orders, or if inventories normalize faster than expected. A few quarters of strong HBM and NAND pricing do not prove a permanent regime change if suppliers and customers adjust at the same time. In that view, the market is not missing structural change; it is underestimating how quickly the shortage can fade once supply catches up.
The clearest falsifying signal for the structural case would be a sustained break in pricing and utilization across more than one product line. If HBM lead times shorten materially, if DRAM and NAND pricing flatten or fall after the next quarter, and if suppliers start talking about inventory normalization rather than constrained output, the market will have to admit that the AI boom is still mostly cyclical. If prices stay firm even as output rises, the structural case gets stronger.
What The Earnings Tell Us About The Next Leg
The near-term winners are the suppliers with the most direct AI exposure and the tightest product mix. SK Hynix has the clearest leverage to HBM demand. Samsung has the broader mix, which gives it scale but also more moving parts. Kioxia sits closer to the NAND recovery and storage side of the AI stack, where the question is whether data-heavy inference and enterprise storage can keep absorbing capacity as fast as it is added. If the reports confirm strong pricing across all three, investors will have a stronger case for saying the AI memory trade still has room to run.
The medium-term risk is the same one that has always governed memory: supply response. Higher profits invite capex, and capex eventually feeds more supply. The difference now is that the demand base is broader and more technically demanding. AI systems consume more memory per server than prior generations, so the industry may be able to absorb a larger increase in output before pricing breaks. But that only delays the cycle; it does not remove it. The current phase looks more like an extended upcycle than a permanent disappearance of volatility.
The long-term implication is more interesting. If AI infrastructure keeps requiring greater memory intensity per dollar of compute, the suppliers stop looking like pure commodity manufacturers and start behaving more like strategic infrastructure providers. That does not make earnings linear. It raises the floor for demand while leaving the ceiling free to swing. In practice, that can support higher average profitability than in previous cycles, but not immunity from downturns.
Base case: earnings confirm strong pricing and healthy demand, and the sector keeps trading as a favored AI beneficiary, though any guide that misses a high bar will still trigger sharp stock-specific moves. Upside case: management commentary shows AI demand still outrunning new supply, which would extend the upcycle. Downside case: companies emphasize inventory build, capex ramping or softer orders, shifting the narrative back toward a standard memory correction.
As of July 24, 2026, the real test is not whether earnings are strong. It is whether the companies can show that AI demand is still outrunning the industry’s attempt to catch up. If they can, the current move is bigger than a cycle. If they cannot, the market is likely just paying up for the peak.
Memory has entered a new demand era. The old supply discipline problem is still waiting at the end of the road.
Explore more exclusive insights at nextfin.ai.
