--- date: 2026-09-01 subject: "4Paradigm buys Ascend 950 | Galactic Energy's first liquid rocket flies | US-China Model Cost Gap" --- **4Paradigm commits to Huawei Ascend chips**, planning to spend more than 1 billion yuan ($150 million) on 950-series accelerators, which puts a paying commercial customer behind Huawei's domestic substitute for restricted Nvidia parts. **Galactic Energy orbited its first liquid-fuel rocket**, the Zhishenxing-1, which lifted off from Jiuquan today and moves the company into the medium-to-large class needed for frequent commercial flights. **Chinese models closed to within three points of Claude Opus 5** on the Artificial Analysis Intelligence Index, where Moonshot AI's Kimi K3 and Zhipu's GLM-5.3 both scored 60, and GLM-5.3 gets there at under a third of Opus 5's cost per task. # Top Stories - **Enterprise AI developer 4Paradigm puts more than 1 billion yuan into Huawei Ascend chips** — 4Paradigm plans to spend more than 1 billion yuan ($150 million) on Huawei Ascend 950-series accelerators, people familiar with the matter told AI Frontline. The Ascend series is Huawei's AI accelerator line and the domestic substitute for the Nvidia parts that U.S. export controls keep out of China. Founder and chief executive Dai Wenyuan traveled to Shenzhen before the disclosure to meet Huawei executive director Yang Chaobin and discuss how the compute partnership would run. The Hong Kong-listed developer will use the chips to run large models in finance, manufacturing and wider enterprise work. Customer token calls, the unit of text a model processes, rose about 620% year on year. The Ascend 950PR drew more positive customer-test feedback than the earlier 910C, Reuters reported in March, as better software compatibility lowered the cost of moving code off Nvidia's stack. [Sina Tech](https://finance.sina.com.cn/tech/it/2026-09-01/doc-iniqhwmx8233255.shtml) [state-affiliated] - **Galactic Energy puts its first reusable liquid-fuel rocket into orbit** — Galactic Energy's Zhishenxing-1 rocket, also called Pallas-1, reached orbit today after lifting off from the company's own pad at Jiuquan, fired for the first time. The flight moves the company from solid-fuel launchers into medium-to-large liquid vehicles, the class that decides whether a launch firm can fly often and cheaply enough for commercial payloads. Its first stage clusters seven CQ-50 engines burning liquid oxygen and kerosene, and can still complete the mission if one of them quits. To slow itself down for landing the engine can throttle down to a lower threshold. No recovery was attempted, and the company says the path there runs in stages toward a full vertical landing of the first stage. Two Chinese rockets have been recovered already: the Long March 10B, caught by a net on a sea platform July 10, and LandSpace's Zhuque-3, which landed on legs Aug. 19. [Sina Tech](https://finance.sina.com.cn/jjxw/2026-09-01/doc-iniqiaua3667941.shtml) [state-affiliated] - **Chinese models close on U.S. capability scores while U.S. vendors keep the spending** — Moonshot AI's Kimi K3, last covered [August 7th](https://chinatechdaily.org/archive/china-daily/2026-08-07/) for escaping its test sandbox, now comes within three points of Claude Opus 5 on the Artificial Analysis Intelligence Index, a composite capability score. Zhipu's GLM-5.3 ranks alongside it, per Bank of America's August monitor of Chinese AI models. Chinese models are taking a growing share of tokens processed while U.S. vendors keep most of the spending. Anthropic ranked second in usage on the Vercel AI platform while capturing 64.8% of spending there. DeepSeek V4 Flash led OpenRouter, a marketplace routing developer requests across models, on tokens processed through Aug. 17, ahead of Tencent's Hy3 and Xiaomi's MiMo-V2.5. GLM-5.3 runs at under a third of Claude Opus 5's cost per intelligence task. Bank of America expects the four largest Chinese cloud platforms to nearly double combined capital spending in 2026. [Wallstreetcn](https://wallstreetcn.com/articles/3780788) - **Chinese liquid-cooling makers report 99% leak-test pass rates on AI server parts** — China's leading liquid-cooling manufacturers hold helium leak-test pass rates of 99% or better on the welded assemblies that carry coolant through AI servers, against the roughly 70% TMTPost says European and U.S. plants may manage. TMTPost frames that gap as what slows North American cloud giants trying to stand up an alternative supply chain at home, where the binding metric is first-pass weld yield. A helium leak test evacuates a finished assembly, charges it with helium and checks whether gas escapes through a seam. Shanxi Taigang Stainless Steel says U.S. and Chinese specifications for the 316 stainless used in coolant manifolds differ by two percentage points of nickel. The company has issued its own grade to settle the difference. A coolant-distribution-unit assembly line in Qingdao is booked through the end of December, and a Foshan component maker says orders have nearly doubled year on year. [TMTPost](https://www.tmtpost.com/8124529.html) - **China's new customer-service standard tells companies to make the handoff to a human easy** — China's first national standard on handoffs between AI and human customer service took effect today. The standard is recommended, so it sets a benchmark companies and regulators can point to rather than a rule carrying penalties of its own. It calls for a convenient entry point to a person and for customer data to carry across so nobody repeats themselves, and it sets where an automated agent should stop rather than limiting where one may be deployed. AI may explain rules, collect materials and generate a ticket, but final confirmation of price, refunds and other binding terms should come from a human agent or a formal process. After-sales service accounted for 26.79% of complaints in the first half of 2026, the China Consumers Association reported, naming false AI promises and difficulty reaching a human agent.[Global Times](https://china.huanqiu.com/article/4T27CvhItvC) [Caibo](https://caibocn.com/newsinfo/11218401.html) [state media] # Policy & Regulation - **Commerce ministry orders faster standards work on smart cars and AI devices** — The Ministry of Commerce (MOFCOM), which oversees trade and domestic consumption policy, and six other departments called yesterday for faster application of standards covering intelligent connected vehicles, smart home appliances and AI terminals. The push targets the basic communication protocols and data interfaces smart terminals use, which decide whether a device from one vendor works with another's. Smart home interoperability standards are singled out for accelerated drafting. The document targets roughly 60 trillion yuan ($8.9 trillion) in total retail sales of consumer goods by 2030, with 10 trillion yuan ($1.5 trillion) markets in green, smart and health consumption. Financial institutions are encouraged to increase support for consumption of home appliances, digital products and smart home products. [Sina Tech](https://finance.sina.com.cn/tech/roll/2026-08-31/doc-iniqfear7329508.shtml) [state-affiliated] # AI & Foundation Models - **DeepSeek publishes open weights for its first image-reading V4 model** — DeepSeek published the weights for DeepSeek-V4-Flash-Vision-Exp on Hugging Face yesterday, following the Aug. 21 image-input API launch this newsletter covered on [August 24th](https://chinatechdaily.org/archive/china-daily/2026-08-24/). That moves the model from a paid interface to files anyone can run, so the vision capability works on hardware DeepSeek does not control. The model runs to roughly 305 billion parameters, the values adjusted during training, and pairs the V4-Flash architecture with a vision module that reads on-screen text and analyzes charts. On DeepSeek's own benchmark tables the model beat Anthropic's Claude Opus 4.8 on two of four multimodal tests and trailed it on the other two. DeepSeek's interface documentation says larger images are downscaled toward a total pixel count near 800 by 800 before they enter the model, a limit on tasks that need fine text or texture detail. [CnBeta](https://www.cnbeta.com.tw/articles/tech/1575922.htm) - **MiniMax puts hosted tiers of its faster H3 Max video model online** — MiniMax switched on 768p and 480p tiers of H3 Max across its open platform and its design tool, both handling text-to-video and image-to-video generation. MiniMax's API documentation lists it as a fast generation version that drops the 2K output and the multimodal reference inputs the open-source H3 supports. The underlying H3 has been downloaded more than 24 million times since MiniMax open-sourced it on Aug. 3, and its text and vision encoder is built on the pretrained weights of Alibaba's Qwen3-VL-32B. PingWest puts the new 768p tier at under three seconds to render a five-second clip and describes as meeting live-streaming latency requirements. [PingWest](https://www.pingwest.com/w/316962) [MiniMax](https://www.minimax.io/news/minimax-h3-open-source) [MiniMax Platform](https://platform.minimaxi.com/docs/api-reference/video-generation-v2-create) - **AI developer Zhipu says annualized revenue reached $1.6 billion after doubling API prices** — Zhipu board secretary Xiao Lei told an earnings call yesterday that annual recurring revenue reached $1.6 billion at the end of August. Demand absorbed the repricing, with average interface prices up about 101% since the start of the year while token call volume grew more than fortyfold. Xiao described the figure as a management-reported order-revenue measure rather than recognized revenue. The Coding Plan subscription, effectively paused for new sales until compute supply improved in July, reopened at a higher price and its call volume kept climbing. Xiao said Zhipu has begun an initial distribution partnership with Amazon Web Services, with talks under way with two other large Western cloud providers. Chief scientist Tang Jie said the next-generation base model is already in training. [Wallstreetcn](https://wallstreetcn.com/articles/3780798) - **AI produced drama airs in Hunan TV primetime under a new review pathway** — A 30-episode series described as China's first fully AI produced long-form drama is airing in Hunan Satellite TV's primetime slot and simultaneously on Mango TV. It is the first title to run under the broadcast regulator's parallel produce-review-broadcast pathway, so state review is being fitted to AI generated content rather than used to hold it back. That pathway comes from a set of 21 measures the regulator issued. An analyst estimate puts the cost compression at roughly 92% per project. [Wallstreetcn](https://wallstreetcn.com/member/articles/3780704) # Chips & Semiconductors - **UniVista releases a domestic tool for 3D chip stacking** — Shanghai UniVista Industrial Software Group released an electronic design automation lineup, the software engineers use to lay out and verify a chip before a foundry builds it. The company says it filled gaps rather than patched existing tools, naming commercial-grade 3D-IC physical design and EDA agents as the two it closed. It bills the lineup as carrying China's first domestic physical design tool for 3D-IC work, meaning stacking multiple dies into a single package, alongside a domestic tool for convergence at advanced process nodes and an "Agentic EDA" system of software agents, also claimed as a domestic first. [Yicai](https://www.yicai.com/brief/103342568.html) [state-affiliated] - **BYD Semiconductor launches a 4D radar chip aimed at rival driving platforms** — BYD Semiconductor launched an automotive-grade 4D millimeter-wave radar chip built around its Xuanji A3 compute hub, where 4D radar adds height to the range, speed and bearing conventional car radar measures. The part can attach to third-party driver-assistance platforms as well as BYD's own, so it is aimed at the merchant market rather than captive use. The chip carries eight transmit and eight receive channels and connects over interfaces compatible with mainstream serializer-deserializer parts, the links that carry sensor data across a car. Co-tuning between the radar chip and the Xuanji A3 driving chip is complete. BYD Semiconductor plans to follow with a lower-channel-count version, chips for cabin and side-door sensing, and an ultra-wideband part. [Wallstreetcn](https://wallstreetcn.com/livenews/3158324) # Robotics & Autonomous Systems - **China drafts the first international test standard for legged robots** — China's Standardization Administration announced the release of an international standard for testing how legged robots move, drafted under Chinese lead. A test standard fixes what counts as performance, and this one scores movement on stairs, slopes and obstacles rather than on flat ground alone. It also folds perception, interaction and decision-making into the score, shifting evaluation toward whether a robot completes a task. The document forms part five of a series on service robot performance and test methods. [Global Times](https://china.huanqiu.com/article/4T24WV3yrsI) [state media]