Sugon 8000 hits full load week one | Zhipu launches 1GW domestic build | Tencent Q4 NPO super nodes

Your morning brief on Chinese technology.

Bottom Line Up Front

Sugon disclosed a domestic 100,000-chip cluster at WAIC 2026 that averages over 150,000 compute jobs per day with peaks above 500,000, and unveiled Scale Fabric, a homegrown InfiniBand alternative built on 400-gigabit-per-second network cards and 880-gigabit-per-second switching chips. Zhipu began building a 1-gigawatt data center on Chinese designed accelerators and acquired XCore Sigma, a Chinese Academy of Sciences compiler spinoff that translates AI model code across different domestic accelerators, with vendors, site and completion date undisclosed. Tencent Cloud pledged large-scale domestic accelerator deployment to lower inference costs and said it will roll out near-package-optics super nodes in the fourth quarter of 2026, calling for unified NPO industry standards.

Sugon disclosed the Sugon 8000 supercluster, nicknamed Dengfeng, at the World Artificial Intelligence Conference (WAIC) 2026 in Shanghai. The system has averaged more than 150,000 compute jobs per day, with single-day peaks above 500,000. It was completed in July at the 2026 Intelligent Computing Application Conference of Hygon's "Photosynthesis" alliance (光合组织, Guanghe) and physically debuted at WAIC one week later. Sugon senior vice president Li Bin said the system could support 5% to 10% of China's current national token demand if devoted entirely to inference, the stage where a trained model answers user queries. Sugon also disclosed Scale Fabric, a homegrown combination of 400-gigabit-per-second network cards and 880-gigabit-per-second switching chips, as its alternative to InfiniBand, the dominant high-speed interconnect for large AI training clusters outside China.

Read at Leiphone ↗

Zhipu launches 1-gigawatt AI data center built entirely on Chinese chips, buys compiler vendor

Chinese AI startup Zhipu has begun building a 1-gigawatt AI data center that will run entirely on domestically designed accelerators, according to people close to the company. Zhipu confirmed the same day that it acquired XCore Sigma, a Chinese Academy of Sciences compiler and runtime spinoff focused on translating AI model code onto different Chinese accelerators, per Shanghai Securities News. Different domestic accelerators expose different low-level programming interfaces, and compiler software of the kind XCore Sigma makes determines how much useful compute a mixed-vendor fleet can deliver. The vendors, site and completion date were not disclosed.

Read at Wallstreetcn ↗

Tencent Cloud pledges large-scale domestic accelerator deployment, targets Q4 for optical super nodes

Tencent Cloud executives told a WAIC 2026 sub-forum in Shanghai that the cloud arm of internet giant Tencent will deploy Chinese designed AI accelerators at scale to lower inference costs. Tencent expects to roll out near-package-optics (NPO) super nodes in the fourth quarter of 2026 and called for unified NPO industry standards, per People's Finance. NPO puts the electro-optical conversion chip on the same board as the AI compute chip, eliminating the digital signal processor found in conventional pluggable optical modules. Tencent Cloud vice president and Tencent Network general manager Wang Yachen said optical interconnects have become the key to breaking bandwidth and power bottlenecks as domestic super node clusters scale up.

Read at ITHome ↗

AI & FOUNDATION MODELS

Tencent Hunyuan releases self-improving research agent Hyra-1.0

Tencent Hunyuan, Tencent's in-house foundation model division, released the first version of its AI research agent Hyra-1.0 on July 21 and posted open source results on GitHub. Tencent said Hyra performs recursive self-improvement, iterating on its own past outputs to produce progressively better solutions to research and engineering problems. Tencent reported that Hyra-1.0 set new records on three published benchmarks measuring AI agents doing AI research, including pushing NanoChat Autoresearch's validation bits-per-byte, a language model prediction-loss metric, down to 0.9015. Tencent also said Hyra refreshed best-known results on 29 of 55 open math problems and produced a PARP1 inhibitor candidate that outscored the marketed cancer drug olaparib on combined binding and drug-likeness scoring.

Read at Ifeng ↗

Alibaba releases third-generation Qwen image model built for commercial production workflows

Alibaba released Qwen-Image-3.0, the third generation image generation model in its Qwen foundation model family, on July 21. The model accepts prompts of up to 4,500 tokens and can generate, in a single pass, complex layouts including knowledge diagrams with mathematical formulas and nested user interface mockups. Qwen-Image-3.0 natively renders text in 12 languages and more than 20 fonts, which Alibaba positions for multilingual product posters and film and comic storyboards. API access is open for invited testing through Alibaba Cloud Bailian, Alibaba's cloud model serving platform, and Alibaba said Qwen Studio and consumer Qwen app rollouts are imminent.

Read at ITHome ↗

III

CHIPS & SEMICONDUCTORS

Sanctioned Chinese GPU maker Moore Threads links 256 GPUs into two-rack super node

Moore Threads demonstrated its MTT C256 super node system for the first time at the World Artificial Intelligence Conference (WAIC) 2026 in Shanghai, packing 256 GPUs into a single data-center-scale computing unit that fits inside two standard equipment racks. Moore Threads is on the U.S. Entity List, a Commerce Department restriction that cuts off access to U.S.-origin advanced chip technology. The company said the MTT C256 uses a single-layer network in which every one of the 256 GPUs connects directly to every other, breaking a roughly 64-GPU ceiling on prior single-layer designs. Moore Threads said the design compresses card-to-card communication latency to sub-microsecond level, or under one millionth of a second.

Read at ITHome ↗

TALENT & CAPITAL

China's AI talent war intensifies as top new-grad packages blow past 1 million yuan

Senior AI recruiter Nie Wei of the Risfond executive search firm told Chinese business outlet Gelonghui that AI talent demand at top Chinese tech firms has grown 80% to 120% a year for the past two years. A separate Hudson report cited by Gelonghui puts China's AI talent supply-to-demand ratio at 0.5, roughly two open positions per qualified candidate. Tsinghua University professor Liu Zhiyuan told Gelonghui that some human resources representatives have advised students at his natural language processing lab to "abandon their degree and join immediately." ByteDance's Top Seed campus hire program illustrates the escalation: it now offers more than 6 million yuan (about $830,000) a year for some core AI roles, up from about 1.5 million yuan in 2024 and 3 to 5 million yuan in 2025, per Sina Finance.

Read at Gelonghui ↗·Read at Sina Finance ↗