Global GPU Computing News【20261025】

Google Cloud and NVIDIA expand Vera Rubin cooperation, with plans to deploy up to 960,000 Rubin GPUs. Google Cloud announced that A5X cloud instances based on the Vera Rubin NVL72 system will be commercially available next month and will gradually expand to the European market, targeting AI agent inference and large-scale reinforcement learning workloads.

NVIDIA unveiled the Vera Rubin system architecture and the DSX POD AI factory solution at GTC Berlin. Vera Rubin uses a fully liquid-cooled design, with each rack able to accommodate 72 GPUs; DSX POD is a prefabricated AI factory reference solution, with Lenovo, Dell, HPE and 17 other companies as ecosystem partners, compressing deployment time from about six months to 14 days.

ByteDance, Alibaba and Tencent reportedly ordered more than 3 million NVIDIA RTX PRO 5500 GPUs in total, worth over US$35 billion. The report said the three Chinese internet companies have placed large-scale orders with NVIDIA, with purchase volumes far exceeding NVIDIA's original production plan; if confirmed, this would significantly affect the global GPU supply landscape.

China's Ministry of Industry and Information Technology disclosed that domestic intelligent computing power reached 2,592 EFLOPS, up 146% year on year. As of the end of September 2026, China's intelligent computing power reached 2,592 EFLOPS; MIIT said it will accelerate the expansion and upgrading of computing facilities and plan a batch of major strategic projects.

A subsidiary of Meishi Technology signed a RMB 839 million GPU computing card procurement contract. Meishi Technology announced that its wholly owned subsidiary Hainan Xingwei signed a sales contract with Company X, with a tax-inclusive procurement amount of RMB 839 million, more than three times the company's revenue last year, for AI model training, inference deployment and large-scale visual data processing.

Biren Technology completed an HK$4 billion share placement in October, with Goldman Sachs bullish on ramping of its next-generation GPUs. The placement raised net proceeds of HK$4 billion, with 70% for product commercialization, supply chain procurement and production preparation, and 20% for R&D and software ecosystem; Goldman Sachs said Biren is advancing internal testing of BR200 and aims to enter customer validation after completion.

Loongson Technology slimmed down its private placement plan, with fundraising for its general-purpose GPU project cut by about 54%. The total fundraising amount was reduced from no more than RMB 2.3 billion to RMB 1.621 billion, and the planned fundraising for the general-purpose GPU R&D project was cut from RMB 360 million to RMB 165 million, while a RMB 483 million working capital replenishment project was cancelled; the company said the overall strategic goal of GPU R&D remains unchanged.

PyTorch officially integrated the FlyDSL backend, accelerating inference and training on AMD MI350 series GPUs. PyTorch and AMD integrated the FlyDSL high-performance kernel backend into the ROCm ecosystem, open to the Instinct MI350 series; on MI355X, MXFP8 scaled matrix multiplication was accelerated by an average of 1.58x and MXFP4 by up to 1.68x, while in vLLM end-to-end tests the MXFP4 quantized version raised request throughput by up to 3.01x.

Vultr and AMD cooperated to deploy a 20MW MI355X supercomputing cluster in Chennai, India. The cluster will be India's largest AMD Instinct deployment and Vultr's second global flagship AMD computing cluster after Springfield, Ohio; each MI355X GPU has 288GB of HBM3E memory and 8TB/s bandwidth.

MetaX set up another RMB 200 million industrial fund as an LP, making strategic around the GPU ecosystem. Chinese GPU company MetaX, through a wholly owned subsidiary, participated in setting up the Shanghai Fuxi Huiqi Venture Capital Fund with a contribution of about RMB 200 million; the investment covers GPU-related IP, chip software and hardware, AI computing applications and industry solutions. This is its second move as an LP this year.