1. WAIC 2026 Grand Opening with an Unprecedented Lineup of Domestic Computing Power
On July 17, the 2026 World Artificial Intelligence Conference (WAIC) opened in Shanghai with a record‑breaking exhibition area exceeding 100,000 square metres. For the first time, the conference featured an independent “Chip‑Compute Integration” pavilion, bringing together over 100 companies including Huawei, Kunlunxin, Moore Threads, and Tsingmicro, showcasing more than 200 cutting‑edge products, 67 of which were domestic debuts or first showcases. All major domestic GPU manufacturers were present, and four entirely new architectural designs competed side by side.
2. Huawei Ascend 950 SuperNode Makes Its First Public Appearance
At WAIC 2026, the Huawei Atlas 950 SuperPoD supernode was unveiled to the public for the first time. Based on the Lingqu interconnect protocol, it achieves the industry’s largest scale of 1,024 cards, delivering 1 EFLOPS FP8 and 2 EFLOPS FP4 compute power, with a 256‑TB unified global memory address space. The cluster scale built on this supernode can be expanded to 500,000 cards. In addition, the Ascend 384 supernode has been commercially deployed in over 750 systems, covering internet, telecom, finance, education, healthcare, and other sectors.
3. China’s First Fully Domestic 100,000‑Card AI Supercluster “Sugon 8000” Runs at Full Capacity in Its First Week
Sugon’s 8000 (Dengfeng) supercluster, awarded the “Treasure of the Pavilion” at WAIC, reached full utilisation in its first week of official operation, with an average of over 150,000 jobs processed per day and a single‑day peak exceeding 500,000. The cluster adopts a globally pioneering high‑density cabinet design, with compute density per unit being 20 times higher than other supernodes, and has achieved stable interconnection at the 100,000‑card scale.
4. Muxi Unveils Xijing S‑Series Supernode and Xisuo X300 GPU
On July 17, Muxi launched the Xijing S‑series supernode and the Xisuo X300 series scientific AI GPU at WAIC. The Xijing S600 enables 64‑card high‑density deployment within a single cabinet, covering the full large‑model training‑to‑inference pipeline. The Xisuo X300 series, built on a fully domestic supply chain and a self‑developed GPGPU architecture, supports scientific computing scenarios such as weather and ocean modelling, fluid dynamics, and molecular dynamics.
5. Moore Threads Demonstrates Three Major “AI Factory” Achievements
Moore Threads showcased its Kua’e intelligent computing cluster, powered by its full‑function GPUs, at WAIC, presenting several “domestic chips training domestic models” results, including the fully trained MoE‑236B foundation model from scratch, and the world’s first 5D world model “Peking University EvoPhys‑World” trained entirely on its own stack. Founder Zhang Jianzhong stated publicly that domestic models can confidently be trained on domestic chips, and that domestic chips have moved from inference to training and integrated training‑inference.
6. Domestic Computing Power Enters the “Ecosystem First” Era, Leaving Behind Single‑Chip Competition
WAIC this year sent a clear signal: competition has moved beyond single‑chip performance into system‑level, full‑stack synergy. Almost all computing vendors launched supernode products. ZTE, together with multiple domestic GPU manufacturers, exhibited the OEX supernode, while Alibaba Cloud presented the Zhenwu M890 and Pangu AL128 chip‑based supernode solution. Companies are no longer simply displaying chip hardware but are focusing on integrated delivery of compute foundations, whole‑rack clusters, software stacks, and industry‑specific ecosystems.
7. NVIDIA GeForce RTX 5090 SE Graphics Card Leaked
Multiple tech media outlets reported that NVIDIA is planning to launch the GeForce RTX 5090 SE, positioned between the RTX 5080 and RTX 5090. Based on the Blackwell GB202 core, it features 14,080 CUDA cores, 24 GB GDDR7 memory with a 384‑bit bus, a power consumption of about 500 W, and an expected launch price of around $1,500.
8. NVIDIA Introduces New Jetson Thor Products and Announces a National‑Level AI Infrastructure in Japan
NVIDIA launched the new Jetson Thor computers (T3000 and T2000 modules), integrating Blackwell GPUs with octa‑core Arm CPUs, delivering 865 FP4 TFLOPS of AI compute in a compact form factor. Separately, NVIDIA announced a partnership with Noetra to build the world’s first national‑level AI infrastructure in Japan, equipped with 27,500 Rubin GPUs and a data centre capacity of 140 megawatts.
9. High‑End GPU Rental Prices Continue to Rise, Supply‑Demand Remains Tight
According to SemiAnalysis, the combined spot and contract rental price for the B200 has risen to approximately $5.0–5.2 per GPU hour, a cumulative increase of about 25‑30% since August 2025. One‑year contract rates for the H100 have also recovered from their lows. Large‑scale long‑term orders worth tens of billions of yuan, such as those from Dongyangguang and Xingyun Technology, are accelerating, with clients pre‑locking premium compute resources. Leading domestic model developers like DeepSeek, Zhipu, and MiniMax are actively raising substantial funding rounds, with capital clearly directed toward model R&D and compute procurement.
10. Oriental Computing Core Globally Debuts High‑Performance AI Chip DF1000
On the eve of WAIC, Oriental Computing Core globally launched its first high‑performance AI chip, the DF1000. Built on a mature domestic 14nm process, the chip claims to achieve performance comparable to 4nm‑based chips, and demonstrated a 512‑card cluster at the event. This marks the official start of China’s independently developed computing chip roadmap.