![Weekly Dose of China Tech [9.21.2026]](https://substackcdn.com/image/fetch/$s_!vSX-!,w_1200,h_675,c_fill,f_jpg,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F10102d49-35d5-424b-83a0-4d4b2f2b539e_1600x840.png)
Dear Reader,
This week we dive into China’s tech acceleration for self-reliance, trading near-term profit margins and market stability for total control over its technological stack. From frontier AI labs and million-processor architectures to foldable supply chains and commercial EV exports, domestic giants are moving aggressively to bypass Western constraints.
Let’s take a look.
Weekly News... #
(I) Z.ai’s ZCode Security Risks and Backlash
Z.ai is facing a privacy backlash over its AI coding tool ZCode.
On September 18, a developer discovered a 313MB encrypted snapshot on his computer while investigating ZCode’s unusually large local storage usage. Records suggested that the tool had packaged around 345MB of a commercial project and repeatedly tried to upload it after the upload failed.
Other developers then found that ZCode could upload workspace data to Alibaba Cloud when users were logged in. The snapshots could include code, project history, logs, caches and other files. Some users also said uploads continued even after they disabled privacy and model-optimization settings.
Z.ai apologized and said the uploads were triggered by ZCode’s codebase indexing feature, which supports functions such as session recovery and Repo Wiki. The company said the feature was initially enabled by default and that the issue has now been fixed.
Z.ai also said it will open-source ZCode, invite third-party audits and compensate users with a weekly quota reset.
The incident highlights a growing problem for AI coding tools: when they can access an entire codebase, users need to know exactly what leaves their machines and why.
(II) Huawei Will Launch Ascend 960DT and 960PR Next Year
Huawei says Ascend 960DT will launch in Q1 2027, three quarters early; 960PR follows in Q3, one quarter early. Ascend 970 and 980 are planned for 2028 and 2029.
Over 1,000 Ascend supernodes are deployed, with 950 systems entering commercial use at scale.
Atlas 960 SuperPoD will support up to 15,488 Ascend cards across 220 racks. Huawei projects 3× the training and 4× the inference performance of Atlas 950 SuperPoD.
(III) ByteDance’s First-half Net Profit Fell Due to More AI Expenditure
ByteDance’s first-half net profit fell by a single-digit percentage to $20B as AI spending increased, The Information reports. Revenue rose about 30% to $120B, helped by TikTok’s international advertising and e-commerce growth.
The company is expanding AI cloud services and developing inference chips. It also reportedly secured a record $30B loan that could support AI investment, while looking to Seedance to drive more AI and cloud revenue.
(IV) BYD Plans to Sell Truck to Europe
BYD plans to launch its first heavy-duty truck in Europe next year and eventually manufacture locally, executive vice president Stella Li said at IAA in Hanover.
“For the long term, we will produce everything we sell in Europe here locally,” she said.
BYD also plans financing, solar-powered charging and maintenance support for fleets. The move comes as European truckmakers push for tariffs on Chinese electric trucks.
(V) Z.ai User Growth and Model Recursive Self-improvement
Z.ai claimed its coding users grew by over 1M in two weeks, while Coding Plan sales rose over 15-fold after resuming in July. The company told investors, GLM-5’s February launch drove a tenfold demand surge, exhausting spare compute within a week and forcing a halt to Coding Plan sales. It has now completed a $5B financing, mostly debt, following July’s $4B raise.
The company says its GLM-5.3-powered Infra Agent has designed, debugged, and optimized the production inference infrastructure for GLM-5.3-Flash across 100,000+ Chinese chips. In under two weeks, throughput reached 3× the initial baseline, with hardware efficiency and per-token costs reportedly comparable to mainstream Nvidia GPUs. The team calls it recursive self-improvement.
(VI) CXMT’s 5th Generation G5 Memory Platform Enters Mass Production
CXMT announces its fifth-generation G5 memory platform has entered mass production, reaching 11.95nm through quadruple exposure and producing at least 50% more dies per wafer than its predecessor.
It also unveiled two mass-produced 24Gb LPDDR5X products, each offering 50% more capacity than comparable previous-generation chips.
CXMT’s global DRAM revenue share reached 9.5% in Q2, up from 7.6% in Q1, according to TrendForce.
(VII) Huawei Unveils Peerium Architecture to Power Million-Processor Clusters
Huawei says its new Peerium architecture can scale to a million processors using nested parallelism, unified memory addressing and peer interconnects.
Its Lingqu interconnect links CPUs, NPUs, memory, SSDs and networking hardware, with Nested BSP underpinning the parallel computing approach.
Atlas 950 is the first Peerium product. Huawei says a 256,000-card cluster is being deployed, while an Atlas 960 system using near-packaged optics is in testing.
(VIII) The Federal Register Website Used Alibaba’s Qwen for AI Search
The Federal Register website used Alibaba’s Qwen for AI search despite FBI allegations that Alibaba improperly copied Anthropic’s technology, Reuters reports.
The feature was disabled Wednesday as social media posts drew attention to it. Its launch date remains unclear.
Experts noted the site handles public information and that open-weight models can run locally. Using Qwen alone does not establish that government data went to Alibaba.
(IX) Chinese EV Manufacturers Cut Reliance on CATL
Chinese EV makers are reducing reliance on CATL through in-house battery systems and alternative cell suppliers.
XPeng CEO He Xiaopeng says the company is taking battery work in-house this year, while distinguishing it from cell manufacturing. XPeng has invested in three cell companies.
Li Auto is shifting some models toward Sunwoda and in-house batteries. Xiaomi’s new N70 and N90 use its Longjia battery system with CALB and Sunwoda cells.
(X) Orienspace Opens a Series C++ Round to Support Reusable Rockets
Orienspace has opened a Series C++ round to support its reusable Gravity-2 rocket, following Gravity-1’s successful September 16 launch.
Gravity-2 targets 21.5 tonnes to low Earth orbit and 15 tonnes to a 500km sun-synchronous orbit. Second-stage firing tests are due soon, followed by first-stage tests ahead of its maiden flight.
The company also plans a shareholding restructuring, then IPO preparation when appropriate, to expand rocket and engine production.
(XI) Huawei Steps Back From Aito Partnership with Seres
Under a new agreement, Seres will take control of Aito’s vehicle development, manufacturing, sales and after-sales services, while Huawei will return to a technology provider role. Aito will also operate its own brand-exclusive retail stores.
The change only applies to Aito. Huawei will continue to take the lead in development and marketing for its other HIMA (HIMA) brands, including Luxeed, Stelato and Maextro.
The timing is notable. Seres is facing weaker profits, while reports have claimed that Huawei previously took around 50,000 yuan from each Aito vehicle sold and supplied more than 30% of the components used in its cars.
Giving Seres more control could improve its margins, but it also changes the balance of Huawei’s automotive strategy. Aito was built around a deep Huawei-Seres partnership. Now the question is whether Aito can keep the same Huawei-driven product experience with Huawei taking a smaller role.
Weekly Articles... #
They Quit AI Labs to Warn About the Future. This DeepSeek Engineer Stayed.
While Western AI pioneers warn of runaway superintelligence, a different crisis is unfolding in China. Shengyu Liu, the engineer behind DeepSeek-V4.1-Flash, is wrestling with the heartbreak of watching AI master the craft he spent a lifetime perfecting. Liu examines this harsh reality: as engineers actively build their replacements, will AI belong to everyone, or will a handful of corporations trap humanity in a high-tech oligarchy?
China Is Told It’s Behind in AI. Now It’s Being Asked to Slow Down.
While U.S. tech leaders position themselves as the self-appointed guardians of civilization, Chinese regulators view AI not as a Frankenstein’s monster, but as a tool already reshaping real-world lives. Is the Western outcry a genuine effort to protect humanity, or a strategic play for corporate dominance? Explore who truly gets to write the rules and whether we are heading toward technological abundance or a cyberpunk corporate oligarchy.
Apple Joins the Foldable Wars. China Supplies the Weapons.
7 years after Samsung and Huawei started the foldable smartphone race, Apple is finally entering the fray. The late arrival iPhone Duo relies heavily on the Chinese supply chain Huawei spent years refining. But with soaring costs and manufacturing bottlenecks limiting qualified production to just a few hundred units a day, can Apple overcome this crisis to challenge Huawei’s dominance?
China’s AI Realism Is Not Conspiracy
When Silicon Valley call for global AI safety rules while lobbying for strict export bans, Western observers often dismiss Chinese skepticism as state-driven conspiracy. But what if Chinese researchers aren’t brainwashed by propaganda? What if they are pragmatic realists? This article will help you understand why China’s AI elite views Dario Amodei as the architect of technological containment.