Chinas Memory Pivot Is a Direct Threat to Samsung and SK Hynix
US export restrictions are severing Chinas access to AI chips, and the collateral damage is landing squarely on Korean memory makers. Samsung and SK Hynix face a narrowing window to offset NAND declines with HBM growth.
The NAND Reversal That Wasnt Supposed to Happen
TrendForce moved its NAND price decline forecast forward by a full quarter in two days. What was once expected to bottom out in the second half of next year is now arriving in Q1 2026, and the downward slope steepens from there. The reasoning is blunt: Chinese cloud providers are ordering far less than North American ones are adding.
This is not a routine cycle dip. It is the first visible tremor from a policy decision that has been building for years. When Washington tightened its grip on how Chinese companies access foreign AI compute, it did not aim directly at Korean memory manufacturers. But the blast radius landed squarely on them.
On November 20, Samsung Electronics fell 5.9 percent and SK Hynix dropped 6.5 percent in Seoul trading. AMD, already a short-seller target for its China exposure, shed 7.7 percent in New York. The Philadelphia Semiconductor Index slid 2.4 percent overnight. The selling was indiscriminate — the market priced in a scenario where Korean memory firms lose a revenue stream they assumed would outlast the cycle.
Why Chinese Demand Matters More Than You Think
NAND flash powers storage drives inside AI servers. Every enterprise SSD in a data center carries a chunk of it. For years, the prevailing assumption among memory makers was simple: build capacity, chase volume, let price stabilize later. China absorbed a meaningful share of that volume. Not the majority, but enough that its withdrawal shifts the entire equilibrium.
Chinese cloud providers already cannot buy the latest GPUs. They have also stopped spending on the storage that would feed those machines. Distributors are liquidating NAND inventory at spot prices that have already fallen more than 20 percent. The gap between North American order growth and Chinese order contraction is, according to TrendForce, large enough to tip global pricing into negative territory starting next quarter.
Supply is not slowing down either. Capital expenditure across the NAND industry is projected to rise 26.7 percent to 29.9 billion dollars next year, and a growing share of that output will come from quad-level cell products that pack more bits per die. Total NAND supply is expected to expand 25 to 30 percent against weaker demand. This is textbook oversupply, and it compounds because the cycle was supposed to be recovering.
The Real Driver: Export Controls Hit Chinas Compute Pipeline
On November 16, the US Commerce Department required export licenses for advanced AI chips flowing through Indonesia, Malaysia, Singapore, Thailand, and Brazil. The rule took effect three days later. Its target was specific: Chinese companies renting datacenter space in Southeast Asia to run NVIDIA chips remotely. If your server sits in Jakarta but belongs to a Beijing-backed entity, you now need permission to operate it.
The results appeared immediately. Moonshot AI, the maker behind the Kimi language model, cancelled a 150 million dollar equipment lease at an Indonesian datacenter, citing the new regulation. In early November, Oracle similarly pulled out of a deal to provide roughly 100,000 advanced NVIDIA chips to Tencent under US government pressure.
An estimate from the Institute for Progress suggests that about three-quarters of all AI compute power Chinese firms access abroad is concentrated in Southeast Asian datacenters. When those venues stall or slow expansion, they do not just reduce GPU rentals. They delay the SSD and memory orders that come bundled with every rack.
Who Wins, Who Loses, and What Moves Next
North American cloud providers are still expanding. They are ordering enterprise SSDs for AI agent servers and general-purpose infrastructure through at least next year. Their demand is real, incremental, and currently outpacing everything else. That is why the DRAM spot picture remains milder than NAND — TrendForce projects mainstream DRAM average selling prices rising five to ten percent in Q1, and HBM4 climbing from $2.05 to between $3.00 and $5.50 per gigabit next year.
But Samsung and SK Hynix live on NAND as much as they live on DRAM. They hold 28 percent and 19 percent of global NAND shipment volume respectively. A sustained price decline does not merely compress margins; it undermines the recovery thesis that justified this year capital spend and investment posture.
AMD faces parallel risk. Hunterbrook Media flagged that AMD generated 22.4 percent of its revenue from China and Hong Kong last year, with Singapore alone accounting for another 12.4 percent. The firm argued that AMD has continued shipping newer chip firmware to blacklisted Chinese server maker IE Information Technology after sanctions were imposed, suggesting a compliance gap. AMD denied wrongdoing, but the stock does not care about denials. The market cares about exposure.
The Decoupling Question No One Is Asking Out Loud
The immediate narrative is about cyclical pricing and short-term demand swings. The longer one reads the data, the more the picture resembles something else entirely: a deliberate fragmentation of Chinas AI hardware supply chain away from Western-controlled components, with collateral damage to third-country suppliers along the way.
Korean fabs are caught in a position no one designed for them. They benefited from Chinese demand during the recent upcycle. They now lose to the same demand contracting under regulatory pressure. The strategy response from Seoul has been largely reactive — doubling down on HBM4 and server DRAM to offset NAND weakness, hoping the premium product mix delivers enough margin insulation.
That may work in the near term. It does not change the structural reality that any credible long play in the AI hardware era must account for a China that cannot reliably purchase Western-adjacent silicon. When a major economy builds its own pipelines around those restrictions, every factory upstream recalibrates for a smaller addressable market — or bets aggressively on a different geography.
Samsung and SK Hynix are making that bet on the Americas. Whether it pays off depends on how fast American hyperscalers add storage capacity, whether HBM orders carry enough weight, and whether Chinese demand stabilizes before the NAND oversupply digests itself. None of those variables are in Seoul’s hands. But the clock is running on every quarter they assumed would be safe.