Samsung and SK Hynix Are Upgrading Their China Memory Plants: What Reuters Reports

Reuters reports that Samsung Electronics and SK Hynix are upgrading their Chinese memory factories to make more advanced NAND flash, driven by huge AI demand and a fast-growing Chinese rival. The details come from unnamed people familiar with the plans: Samsung's Xi'an plant is slated to move to 286-layer NAND and SK Hynix's Dalian site to 240-plus layers. Both companies publicly deny adding capacity, and the key open question is what the US government has actually licensed. This post separates the reported numbers from the confirmed facts, and explains why it matters for the AI hardware supply chain. Reuters' text was carried by several outlets, which we used because the original sits behind a wall for automated readers.

Quick reference: what is reported and what is not

What exactly did Reuters report?

According to the report as carried by outlets such as the TBS News copy, Samsung plans to convert its Xi'an plant to 286-layer NAND. One source put the Xi'an project's monthly output at 5,000 wafers in December 2026 and 40,000 wafers by December 2027. Xi'an currently produces chips of roughly 200 layers and accounts for about 40% of Samsung's global NAND output, a large share that explains why the site matters.

SK Hynix is described as looking to move Dalian from 192-layer to 240-plus-layer products. One source said the company may seek US permission to do research and development on chips with more than 300 layers at Dalian.

The motivation, per the report, is twofold: meet the surge in AI-driven demand and fend off YMTC, the state-backed Chinese maker that has been gaining ground. Omdia data cited in the story show YMTC at 38.5% of China NAND in the second quarter, ahead of Samsung at 33.5% and SK Hynix at 17%.

Timeline: how the China memory story got here

• 2023: Samsung and SK Hynix received open-ended approval to ship US chipmaking tools into their Chinese fabs without individual licences, as the validated end user arrangement.
• 2025: Washington said it would end that status and would not approve licences for upgrades or expansion at existing Chinese plants. Existing tools could keep running.
• Early 2026: The regime moved toward annual approvals for equipment shipments, which eased some supply concerns according to TrendForce.
• March 30, 2026: Samsung reportedly completed its Xi'an conversion from 128-layer to 236-layer lines (TrendForce, citing Korean press).
• July 2026: SK Hynix was reported to be restarting its delayed Dalian Plant 2 expansion in the second half of the year.
• October 21, 2026: Reuters reports the further step to 286 and 240-plus layers, and the companies deny expanding capacity.

What are people asking about this?

Is a layer upgrade the same as an expansion? Not legally, and not physically. More layers raise the bits per wafer without adding wafer starts, which is why the companies can say capacity is unchanged while output in bits still rises. Regulators will judge which side of the line a given tool installation falls on.

Why would a Korean company invest in China at all? Because the plants already exist and hold a large share of output. Xi'an alone is about 40% of Samsung's NAND, so letting it fall behind YMTC would cost market share in the very segment AI storage buyers are bidding for.

Is this about HBM? No. HBM is stacked DRAM made elsewhere; this report concerns NAND flash, the storage side of the same AI supply squeeze.

What do Samsung and SK Hynix say?

Both companies push back on the "expansion" framing. Samsung told Reuters it is not expanding production capacity or constructing new manufacturing facilities. SK Hynix said it is not seeking capacity expansion at Dalian. That wording matters because of the US rules below: a technology upgrade (more layers per wafer, same equipment footprint) is a different thing from adding wafer starts, although in practice a higher-layer node can raise the bits produced per wafer.

Note what the companies did not say: neither is quoted as denying the layer-count targets or the upgrade itself. The anonymous wafer numbers remain unconfirmed.

The export-control puzzle

The US previously treated the Chinese fabs of Samsung and SK Hynix as validated end users, which allowed US-made tools to flow in without individual licences. In 2025, Washington moved to end that arrangement; Blocks and Files reported at the time that the two would lose the ability to bring in equipment for capacity expansion or technology upgrades from the end of that year, while existing tools could keep running. The regime then shifted to annual approvals for equipment shipments, as TrendForce described in July when it reported SK Hynix reviving a delayed Dalian Plant 2 build.

Reuters says both companies need annual US licences to install some US-made equipment, that Washington said in 2025 it did not intend to approve licences for upgrades or expansion at the existing Chinese factories, and that it could not establish how far US authorities have approved equipment licences for the reported work. That gap is the real story: either licences were granted quietly, the work uses tools that need no licence, or the plans are aspirational. Nobody on the record has settled which.

Export controls gate restricting chip equipment flows into Chinese memory fabs

Why AI demand reaches all the way to flash

Most AI-hardware coverage focuses on HBM, the stacked DRAM beside GPUs. But training and inference clusters also consume enormous flash: datasets, checkpoints, retrieval indexes and key-value caches spill to SSDs. As AI buyers absorb supply, memory makers prioritise advanced, high-margin parts, and older NAND lines in China become a lever to add higher-density output without new buildings.

That fits the picture in our earlier coverage. Michael Dell warned the memory shortage is a multi-year problem, Samsung reportedly cut phone-part orders as AI memory costs rose, and US chip producer prices jumped on data-center demand. On the HBM side, Samsung reported 80% HBM4 yield and a 3nm base-die story with TSMC shows how contested that segment is. The Chinese-fab upgrades are the NAND counterpart.

Wafer grid representing higher-layer NAND output per wafer at upgraded fabs

How big is the China footprint?

The Xi'an figure is the striking one: about 40% of Samsung's global NAND output comes from one Chinese site. SK Hynix's exposure is smaller but growing. TrendForce, citing Korean press, reported that Samsung finished converting Xi'an 128-layer lines to 236-layer on March 30 and was in volume production, and that SK Hynix planned 238-layer V8 capacity at Dalian Plant 2 at 30,000 to 50,000 wafers a month from the second half of 2026. Those are separate reports with different layer numbers from the Reuters figures (286 and 240-plus), so the layer counts in circulation are not yet consistent. Read them as a direction of travel, not a spec sheet.

SK Hynix's Dalian business was built on its acquisition of Intel's NAND unit, and TrendForce notes it is viewed as the fastest-expanding of the company's NAND bases. That helps explain why a U.S. permission request for more-than-300-layer R&D, if real, would be notable.

What does this mean for builders and buyers?

You will not see an immediate price change. The reported ramps run from December 2026 to December 2027, resting on anonymous sourcing, and licensing could still block them. But three practical takeaways follow:

1. Storage supply is part of the AI squeeze. Teams planning on-prem inference or large local datasets should expect SSD pricing and lead times to follow the same tight trend as DRAM; see our note on GPU rental prices at record highs.
2. Geopolitics is baked into memory supply. A large slice of Samsung's NAND sits behind US licensing decisions, so policy moves can change supply quickly. Our Micron and Taiwan supply-risk post covers a different but related fragility.
3. Equipment, not demand, is the bottleneck. ASML says AI demand exceeds EUV supply into 2027 to 2028; NAND upgrades lean on different tools but face the same shortage logic.

Claimed versus verified

Verified by multiple outlets carrying Reuters: the report exists, both companies deny capacity expansion, and annual US licences are required for some US-made tools. Reported by anonymous sources only: the 286-layer and 240-plus-layer targets, the Xi'an wafer schedule, the possible 300-plus-layer R&D request. Not established: the extent of US licence approvals. We could not retrieve the original Reuters page directly, so every number above is attributed to the report and not to company filings. If Samsung or SK Hynix publish capex details in their next earnings calls, that will be the cleanest check.

Related reading

• SK hynix Q3 2026: record 83.33 trillion won operating profit and a memory shortage seen through 2030

• Michael Dell: memory shortage is a four-year problem

• AI chip supply chain: TSMC record, Samsung phone cut, GlobalFoundries interposers

• Samsung hits 80% HBM4 yield

• Samsung and TSMC NVHBM base die: claimed vs verified

• US chip producer prices up 27% on AI data centers

• ASML Q3 2026: AI demand exceeds EUV supply

Details are accurate as of October 21, 2026 and rest on Reuters reporting as republished by other outlets; check company statements for updates.