Article — "Memory Inventories Fall Below Ten Days as CXMT Takes the Commodity Base": Read the two numbers together and they don't fit the story the market tells about them.
SK Hynix's DRAM revenue rose 214% year over year in the second quarter of 2026. In the same window, its DRAM share fell from 39% to 26%.
That is not a company losing. That is a company winning so hard in one lane that it walked out of another — and the vacated lane is now worth arguing about.
That is not a company losing. That is a company winning so hard in one lane that it walked out of another — and the vacated lane is now worth arguing about.
Last week this column tracked CXMT's gap-narrowing against the Korean incumbents as a trajectory. This week the trajectory has a floor and a ceiling attached: Samsung and SK Hynix inventories below ten days, an explicit exhaustion warning for next year, and CXMT sitting at exactly 10% of global DRAM revenue two years ahead of forecast.
Start with the arithmetic nobody discounted properly.
KB Securities' Kim Dong-won puts it in one sentence: one HBM4 wafer occupies production capacity equivalent to three commodity DRAM wafers. Every wafer redirected to stacked memory for AI accelerators erases three wafers of ordinary bit supply.
That ratio, not a demand rebound, dropped Samsung and SK Hynix inventories below ten days as of the third quarter. KB projects DRAM and NAND bit demand to outrun supply by more than ten percentage points next year, with the "actual exhaustion of sellable memory supply" a live possibility in 2027.
Read that as accounting, not weather. The incumbents chose the mix. Server DRAM contract prices rose a cumulative 64% in the second half of 2025 and are forecast to climb roughly 270% across the year; HBM could add another 70 to 140% on top. Facing that spread, redirecting the best lines toward HBM is the rational trade.
But the same choice is a decision to underserve the commodity floor. And a floor left undefended gets occupied.
Now look at Hefei, because it is the mirror image of the same event.
CXMT held below 1% of global DRAM revenue in 2023. It reached 4% by the second quarter of 2025, reportedly 8% by year-end, and exactly 10% in the second quarter of 2026 — roughly two points of share per quarter, three quarters running.
Counterpoint and UBS had penciled that 10% milestone for 2028. It arrived two years early. And with it, the combined share of Samsung, SK Hynix and Micron slipped to 87% — below 90% for the first time in more than a decade, a threshold that would have read near 94% in any recent quarter.
Here is the part Seoul should find harder to swallow than the share loss itself. CXMT runs entirely on deep-ultraviolet lithography — no EUV — using self-aligned double and quadruple patterning to reach feature sizes that would otherwise need ASML's frontier machines.
Every extra exposure pass adds steps and accumulates alignment error. The estimated penalty is a cost-per-bit roughly 30% above the incumbents.
In a normal pricing environment, that penalty makes commodity DRAM margin-negative. CXMT would bleed on every chip.
But mainstream DRAM prices are up more than 200% from early 2025. The supercycle is precisely large enough to swallow a 30% cost gap and leave profit behind.
So the incumbents' margin choice does two things at once. It drains their own inventory below ten days — and it lifts the price umbrella under which a structurally higher-cost entrant becomes, for the first time, profitable.
The financials confirm it wasn't marginal. CXMT's first-half 2026 revenue hit roughly $22.37 billion, up 874% year over year, reversing a prior-year loss and beating its own pre-listing guidance.
That cash is being poured back into the floor. CXMT plans to move from about 300,000 wafers per month toward 550,000 to 600,000 by 2028 — approaching SK Hynix's roughly 590,000.
Sit with that comparison. The entrant everyone dismissed as a legacy-node also-ran is building toward the wafer capacity of the number-two incumbent, financed by the very shortage the incumbents created by vacating the segment.
And its demand base is captive in a way the incumbents cannot contest. Huawei, Xiaomi, Oppo, Alibaba Cloud and Tencent Cloud all hold long-term supply agreements with CXMT; a 2023 Chinese cybersecurity ruling barred critical-infrastructure operators from key foreign chips. China is the world's largest DRAM consumer, and that base insulates CXMT from the open-market pricing volatility that would otherwise expose its cost disadvantage.
Which reframes the export-control debate entirely. The controls aim at EUV access — the one tool CXMT has already routed around with SAQP. The sanctions target a chokepoint the entrant reached 10% without ever touching.
The historical rhyme is uncomfortable. In the 1980s, Samsung and Hynix ran the state-backed-entrant playbook against Japan — captive domestic demand, subsidized capacity, a commodity focus in the segment incumbents were leaving. CXMT is running the same script, and the Western answer this time is procurement bans rather than antidumping law, a mechanism whose effectiveness the 10% number now openly questions.
Now the trap, which is a timing trap.
The shortage is real and probably durable — Micron reports HBM and NAND sold out through the end of 2026, and SK Hynix's chief has flagged a squeeze stretching toward 2030. But shortages normalize. When this one does, prices fall back toward the level where a 30% cost penalty matters again.
Except the incumbents will have spent the interim optimizing away from commodity capacity, not toward it. Re-entering a floor you deliberately abandoned is not cheap; the lines, the yield learning, and the customer relationships all decayed while you chased HBM margin.
And the Chinese base won't be there to re-take. It is locked to CXMT by contract and by cybersecurity fiat. That demand does not come back to Seoul when spot prices soften.
The one loose thread is the demonstration effect. Reports of Apple weighing Chinese DRAM sourcing remain rumor — but if they harden, they prove a Western buyer will treat Chinese DRAM as fungible. That is the display-industry ending Korean analysts fear out loud: a one-year "technology lead" that evaporated in three, ceding a commodity market that never came back.
Against all of this, one incumbent is hedging the actual risk rather than the headline one.
Micron is committing $250 billion across the United States — two Boise fabs pulled forward to mid-2027, more in New York — and it is the only one of the three building memory front-end capacity on US soil. Samsung's Taylor plant is logic foundry; SK Hynix's Indiana site is advanced packaging. Neither is a DRAM fab.
The tariff politics make the logic explicit. Commerce Secretary Howard Lutnick is threatening tariffs up to 100% on memory not made in America; Chief Business Officer Sumit Sadana says Micron's US-made memory "will command a price premium."
Strip the nationalism and Micron's move is a bet on geography as a moat. If the commodity floor is going to be captured by whoever owns the captive local demand and the lowest all-in cost to serve it, then re-anchoring supply to the buyer's jurisdiction is the one moat that survives normalization.
The Korean incumbents traded a durable moat — wafer-allocation discipline that printed pricing power for a decade — for a cyclical one. HBM margin is real, but it is a margin, not a floor.
They didn't lose the floor to a better chip. They vacated it for a better quarter, and the supercycle is writing the successor's checks.