Infrastructure

A 14 Percent Crash Will Not Cut Your Memory Bill

Samsung fell 13.4 percent and SK Hynix 14.7 percent on a Chinese lithography tool. That tool makes commodity DRAM, not HBM. What it changes in your memory contract, and when.

InfrastructureBy Servola Tech Desk2026-07-286 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

A 14 Percent Crash Will Not Cut Your Memory Bill

Key takeaways

Samsung fell 13.4 percent and SK Hynix 14.7 percent in a single session on news that China has started building its own DUV lithography tools, and that tool reaches commodity DRAM but not the high-bandwidth memory those two companies actually earn on.
Do not read a supplier share-price crash as a price cut: the first Chinese tools are five units this year and about twenty in 2027, which lands as commodity DRAM supply around 2028 and changes nothing in a 2027 procurement budget.
China is roughly two generations behind on HBM, with ChangXin at HBM2 in mass production and HBM3E targeted for 2027, while Samsung began shipping HBM4 in February 2026 and SK Hynix holds about two thirds of Nvidia's HBM4 allocation.
The sell-off was not an overreaction, because conventional server DRAM has risen to roughly 1.3 dollars per gigabit, close to HBM3E levels, so most of the current Korean profit sits in the half of memory a Chinese second source can eventually copy.
If you are being offered a five-year fixed-price hardware contract with memory bundled in, split the term: take the long commitment on HBM-side capacity and keep DDR5 and LPDDR5X on a shorter term or a price-review clause, because that is the segment where a second source is coming.

Seoul halted trading, then halted it again

The Korea Exchange stopped trading in both of its markets on Tuesday, the third simultaneous circuit breaker of the year. When the KOSPI reopened it closed at 6,023.66, down 732.09 points or 10.84 percent, after touching 5,992.91 intraday. It was the second-largest single-day decline in the index's history. The KOSDAQ closed at 705.85, down 7.72 percent, and fell below 700 during the session for the first time since April 2025.

The damage was concentrated with unusual precision. Samsung Electronics closed 13.39 percent lower at 220,000 won, its worst single day in close to two decades. SK Hynix fell 14.65 percent to 1,555,000 won. In Tokyo, Kioxia dropped 18.33 percent. The broader indices moved far less: the Nikkei 225 fell 4 percent to 62,364.92, Taiwan's Taiex fell 4.7 percent with TSMC down 3 percent, and the Hang Seng actually closed up 0.4 percent. Shanghai lost 1.2 percent.

This was a memory sell-off wearing a market sell-off's clothes. The equity analyst Jing Jie Yu described the market as likely spooked by the progress of China's chipmaking equipment capabilities and by what that would mean for the competitive position of the incumbent chip and chip-equipment leaders. The specific trigger was a Shanghai state-owned enterprise beginning production of immersion deep ultraviolet lithography tools, a category Netherlands-based ASML has effectively owned, with a stated plan of five units this year and roughly twenty in 2027 for ChangXin Memory, SMIC and Hua Hong.

The tool is a DRAM tool, and HBM is not a DRAM problem

The machine that moved the market cannot build the product the market was pricing. Immersion DUV is the workhorse of conventional memory and mature logic. High-bandwidth memory, the stacked product that sits beside an AI accelerator, is gated somewhere else entirely. Its constraint is partly the next lithography step, where analysts covering China's memory build-out judge that EUV becomes unavoidable for cutting-edge DRAM and HBM, and EUV in China remains at prototype stage while DUV is assessed to have entered commercialisation. Chinese DUV is real; Chinese EUV is not yet a product.

The other constraint is not lithography at all. It is packaging. The December 2024 United States export controls named the equipment used for through-silicon vias, etching and the related stacking steps, which is precisely the toolchain that turns DRAM dies into an HBM stack. A country can hold a credible DUV line and still be unable to assemble competitive HBM, because the two problems are solved with different machines.

The generational gap shows in the roadmaps. ChangXin has HBM2 in mass production, moved on HBM3 during the first half of 2026 and targets HBM3E in 2027. Samsung began delivering HBM4 in February 2026, and SK Hynix is understood to hold close to two thirds of Nvidia's HBM4 allocation for the Vera Rubin platform. Nvidia is asking for data rates above 10 gigabits per second against a JEDEC standard of 8. On that specific product, on that specific specification, the gap is two generations wide and the tools that would close it are the ones under control.

None of this makes the Chinese programme unserious. It makes it a commodity-DRAM programme for now, which is a different thing from an HBM programme, and the difference is the whole argument.

Why the sell-off was not an overreaction

The obvious conclusion from the paragraph above is that the market got it wrong, and that conclusion is itself wrong. It would be right if commodity DRAM were still the low-margin afterthought it was for most of the last decade. It is not. Conventional server DRAM has been repriced to roughly 1.3 dollars per gigabit, close to what Nvidia's flagship HBM3E commands. The shortage that has been pushing up the cost of ordinary devices has also pushed ordinary memory into premium pricing.

That changes what a Korean memory maker actually is. A large share of today's profit is being earned on the product a Chinese second source can eventually build, not on the product it cannot. TrendForce expects SK Hynix to hold around 50 percent of global HBM bit output in 2026, down from 59 percent in 2025, with Samsung rising from 20 to 28 percent. HBM is the defensible half. It is not the larger half.

So the market did something more coherent than panic. It marked down the earnings that a second source can reach and left the earnings that it cannot. A 14 percent move is what that arithmetic looks like when it is applied in one afternoon. The corroborating detail is SK Hynix's United States listing, which closed at 143 dollars against a 149 dollar IPO price, giving up the whole of a listing that is only months old.

What changes in your contract, and when

For anyone buying memory, in servers or in devices, the honest answer is that Tuesday changed the date, not the price. Five tools this year and about twenty in 2027 is a delivery schedule, and tools have to be installed, qualified and yielded before they produce sellable parts. Commodity DRAM relief from this source is a 2028 and 2029 question. Any 2027 budget built on the assumption that Chinese supply arrives first is built on a schedule nobody has published.

What it does change is contract term, and only in one segment. If a vendor is offering a multi-year fixed-price arrangement with memory bundled into the hardware line, the segments now have different risk. Commit long where the second source is not coming, which is HBM-side capacity, and stay short where it is, which is DDR5 and LPDDR5X. A five-year fixed price signed at the top of a shortage is a bad instrument in exactly the segment that is due a competitor.

The practical version is a clause, not a forecast. Ask for a price-review trigger on the commodity memory portion tied to a published index, keep the term on that line to two or three years, and leave the accelerator-attached memory on whatever term gets you allocation. You are not predicting when Chinese DRAM lands. You are declining to pay a shortage price for four years after it does.

And treat the share prices as what they are. A 14 percent move is a statement about a supplier's 2028 margin. Your unit price next year is set by wafer allocation, by contract, and by who is ahead of you in the queue. Those two things are related, but they are not the same number, and Tuesday only moved one of them.

Frequently asked questions

Does this mean memory prices are about to fall?

No. The Chinese tool programme is five units this year and roughly twenty in 2027, and those tools still have to be installed and qualified before they produce sellable parts. Any resulting commodity DRAM supply is a 2028 or 2029 event. Nothing in Tuesday's trading changes what you will pay in 2027.

Why did Samsung and SK Hynix fall so much further than TSMC or the wider market?

Because the news was specific to memory. TSMC fell 3 percent and the Hang Seng closed higher, while Samsung fell 13.39 percent and SK Hynix 14.65 percent. The Chinese tool reaches conventional DRAM, and conventional DRAM is currently carrying an unusually large share of Korean memory profit at roughly 1.3 dollars per gigabit.

Can China build HBM with the same tool?

Not on the current evidence. HBM is gated by the next lithography step, where EUV in China is still at prototype stage, and by the stacking toolchain for through-silicon vias and etching that the December 2024 United States export controls specifically named. ChangXin is at HBM2 in mass production with HBM3E targeted for 2027, against HBM4 shipping from Samsung since February 2026.

What should I actually do about my next hardware contract?

Split the term by memory type. Keep DDR5 and LPDDR5X on a shorter term or attach a price-review clause tied to a published index, because that is the segment where a second source is plausible later this decade. Accept longer commitments where they buy you allocation on accelerator-attached memory, where no second source is visible.

The market did not misread the tool. It read it precisely, and marked down the part of Korean memory that a second source can eventually reach. The mistake available to a buyer is different and more expensive: hearing a 14 percent fall and quietly assuming relief is coming, then signing a long fixed price on the one segment that is actually due a competitor.

DRAMHBMSamsungSK HynixChangXin MemoryDUV lithographyMemory procurementKOSPI

Anthropic Asked a Memory Maker for Its Own Chips

SK Group's chairman said on 25 July that Anthropic approached SK Hynix for supplies to build its own semiconductors. That adds a claimant to the queue you are in.

4 min read

Infrastructure

Your 2027 Memory Is Already Spoken For

SK hynix customers want 60 to 100 percent more AI memory in 2027. The chairman wants prices lower. That combination tells a European operator what to contract now.

5 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles