Global GPU Computing News【20260920】

JD Cloud and Moore Threads reach cooperation on a 100,000-GPU-scale domestic GPU cluster. On September 9, JD Cloud announced plans to build a 100,000-GPU intelligent computing cluster based on Moore Threads' full-function GPUs as its computing foundation, focusing on large-model training, inference, and embodied intelligence, and opening the computing power to the entire industry. This marks the first time domestic GPUs have entered a top AI cloud company's 100,000-GPU-scale core intelligent computing cluster.

State Council executive meeting holds a special study on computing power network construction. On September 11, Premier Li Qiang chaired a State Council executive meeting to study work related to computing power network construction. The meeting proposed promoting computing-electricity coordination and computing-network integration, strengthening monitoring and scheduling of computing resources and backbone optical fiber network construction, and accelerating the implementation of projects such as green electricity direct connection.

Nvidia Vera Rubin NVL72 makes its MLPerf debut, with performance up to 3.7 times higher. Preliminary results from MLPerf Inference v6.1 announced on September 16 show that Vera Rubin NVL72 achieved inference throughput up to 3.7 times that of the previous-generation GB300 NVL72 on the Qwen3-VL workload, while DeepSeek-R1 improved by up to 2.5 times.

CoreWeave deploys a multi-rack Vera Rubin NVL72 cluster. AI computing rental provider CoreWeave announced on September 16 that it has put seven Nvidia Vera Rubin NVL72 racks into production in two regions, linking hundreds of Rubin GPUs into a single scale-out system, providing 216 TB per second of NVLink bandwidth and 3,600 PFLOPS of sparse NVFP4 inference compute.

Cloud service provider Nebius raises GPU prices for the second time this year, with increases of up to 21%. On September 17, Nebius announced that starting October 1 it will raise on-demand rental prices for major Nvidia GPUs: H100 from $3.85 per hour to $4.50, H200 from $4.50 to $5.40, B200 from $7.15 to $8.50, and B300 from $7.85 to $9.50, with overall increases of about 17% to 21%.

Jiushi builds China's first L4-level 10,000-GPU computing cluster. On September 17, Jiushi announced that its computing cluster exceeded the 10,000-GPU scale, building a GPU computing cluster of nearly 15,000 GPUs, fully covering the entire R&D process including model training, scene reconstruction, data annotation, and online inference, becoming the first L4 company in the industry to build a 10,000-GPU cluster.

Paratera subsidiary signs a 1.913 billion yuan GPU computing service contract. On September 18, Paratera announced that its wholly owned subsidiary Wuhan Paratera signed a Technical Service Framework Agreement with a client, agreeing to provide GPU computing resources and supporting IT equipment. The total contract amount including tax shall not exceed 1.913 billion yuan, with a service term from January 2027 to January 2032, totaling 60 months.

Kanghui Shares signs a 1.72 billion yuan five-year computing service contract. On September 16, Kanghui Shares announced that its wholly owned subsidiary Beijing Kanghui Zhichuang signed a five-year computing service contract with a client worth about 1.72 billion yuan including tax, and also signed a server procurement contract worth about 1.141 billion yuan, with delivery planned in batches from the end of the third quarter of 2026 to the end of the first quarter of 2027.

Huawei releases the Ascend 960 supernode, with a single node scalable to 4,096 cards. On September 18, during Huawei Connect 2026, Huawei released its first supernode using an NPO optical engine, the Ascend 960 supernode. It adopts an orthogonal architecture and full liquid cooling design, can scale to 4,096 cards, and delivers 8 EFLOPS FP8 and 16 EFLOPS FP4 computing power, intended to match the training and inference needs of 10-trillion-parameter models.

Jensen Huang says Nvidia's chip sales will double next year, with Vera Rubin efficiency greatly improved. On September 17, Nvidia CEO Jensen Huang said at an event in Scotland that he expects the company's chip sales in the coming year to reach twice this year's level. Nvidia's stock price rose 2.5% the same day, with the market focused on Vera Rubin NVL72's MLPerf results showing inference throughput up to 3.7 times higher.