Chip Brief · September 2026

China Is Building Every Layer of the Chip Stack at Once. Korea's Retail Investors Are Buying the American One.

CXMT, Guangdong and Huawei each announced a domestic answer. Korean retail money headed for US chip stocks.

By vizmaya · September 2026
11.95nm
CXMT's claimed half-pitch on its fifth-generation DRAM platform
The company announced it at the World Manufacturing Convention in Hefei. It gets there with quadruple exposure. Samsung and SK Hynix run 10nm-class processes. CXMT says it is now close to the most advanced mass-production memory.
CXMT says its DRAM process has caught up

The company unveiled two 24-gigabit LPDDR5X products, both already in mass production. Some of its high-end chips are already in flagship Chinese phones.

Every number in this story comes from CXMT itself. Nobody outside the company has verified it yet.

50%
More gross dies per wafer than the fourth generation, according to CXMT
Gross dies are counted before defects. Quadruple patterning adds process steps. So cost per good bit depends on yield, and CXMT did not disclose yield.
Ask for yield, not die counts

Memory buyers should ask for yield and reliability data instead of die counts, and check sanctions exposure before qualifying anything.

Tool vendors should track who supplies the extra etch, deposition and metrology steps that quadruple patterning needs.

Guangdong is building the SiC line and the EDA tool at the same time

A silicon carbide line in Nansha, Guangzhou is running near full capacity under the province's "Qiangxin" (Strong Chip) initiative.

Separately, a Shenzhen software firm says it has rebuilt an EDA tool from scratch. Two layers of the stack, one province, one push.

¥10 billion
Approximate value of the Nansha SiC project
It targets 240,000 six-inch wafers and 240,000 eight-inch wafers a year.
100 engineers, one schematic

The Shenzhen firm says its tool lets 100 engineers work on the same schematic at once. It also claims product launch cycles are cut by more than 40%.

Read the claim closely. It covers schematic design only. Place-and-route and signoff aren't mentioned.

The SiC line is the nearer threat

Actionable insight: EDA vendors should ask customers which stages of the flow they have moved to domestic tools before assuming any market loss.

The SiC line is the nearer-term problem. Suppliers of SiC devices and substrates should model price pressure in 8-inch wafers as this capacity ramps, especially in EV and power-supply segments.

Huawei pulls its Ascend roadmap forward and bets on scale

At Connect in Shanghai, Huawei said the Ascend 960DT will launch as early as Q1 2027. That is nine months ahead of plan. The 960PR follows in Q3.

Huawei's AI chips still trail Nvidia's top-end offerings, and the company is not pretending otherwise. It is changing the question.

1 million
The most chips Huawei says its UnifiedBus interconnect can tie into one system
Rather than match Nvidia chip for chip, Huawei links lots of processors together. It also sells ready-made clusters, like the Atlas 960 SuperPoD: more than 4,000 accelerators connected by optical fiber.
When the network is the product

When single-chip performance lags, the network and the power budget become the product.

Optics, networking and cooling suppliers should expect heavy demand from cluster-scale designs like this, and should work out who serves it inside China. Anyone forecasting Nvidia's China revenue should pressure-test any assumption that it holds steady.

India, Taiwan and South Korea

The money side of the story. While China announced hardware, Korean retail investors were busy voting with their brokerage apps.

$1.1 billion
Net US share purchases by South Korean investors in the third week of September, about 1.5 trillion won
They cited easing worries about an AI spending slowdown and growing hopes for a memory recovery. The Kospi fell while the Nasdaq 100 rose, and a stronger dollar against the won added to the pull.
SOXL, Nvidia, Micron and TSMC

Most of the money went into SOXL, a 3x leveraged semiconductor ETF, and into Nvidia, Micron and TSMC.

That concentration matters. A leveraged fund moves fast in both directions.

Don't mistake 3x flows for conviction

Capital is leaving Korea's home market for US names, and the leveraged ETF concentration means swings can hit hard both ways.

IR and finance teams at chip companies should expect sharper retail-driven sentiment moves, and should have answers ready on memory pricing.

Analysts shouldn't read 3x ETF inflows as fundamental conviction. Cross-check them against contract DRAM pricing and hyperscaler capex, and watch the won-dollar rate for Korean exporters.

Market movers

Largest daily moves among the tracked AI-infrastructure stocks, each on its home exchange.

| Company | Ticker | Close | Day | | --- | --- | ---: | ---: | | Super Micro Computer | SMCI | 40.35 USD | +9.50% | | Arm Holdings | ARM | 264.9 USD | +8.57% | | Intel | INTC | 108.8 USD | +7.67% | | CoreWeave | CRWV | 79.88 USD | -4.16% |

Methodology & sources

https://www.vizmaya.fyi/ai-data-centers

Company claims are reported as stated and have not been independently verified. This analysis is not investment advice.

vizmaya · September 2026