Global GPU Computing News【20261018】

China’s national integrated computing network framework has essentially taken shape. As of June this year, China’s intelligent computing capacity reached 2,185 EFLOPS (FP16), up 177% year on year. More than 70 transmission channels connecting computing hubs have been built, and the China Computing Platform has gathered more than 10,000 registered enterprise users and listed more than 2,000 computing products.

A report released this week by the Planning Institute of the China Academy of Information and Communications Technology (CAICT) noted that local intelligent computing policies are shifting from “construction subsidies” to “token service procurement.” The core efficiency metric is moving from peak computing power to “token output per watt.” Beijing, Zhangjiakou, Shaoguan, Qingyang and other places have already developed large-scale token production capacity.

The world’s first Ascend 950DT 10,000-card intelligent computing cluster has landed at the China Telecom Cloud Computing Inner Mongolia Information Park in Horinger New Area, Inner Mongolia. The project has the conditions for deploying a 30,000-card cluster and is positioned as a full-stack independently controllable computing base for training trillion-parameter large models.

The first 1,024-card Ascend Atlas 950 SuperPoD liquid-cooled super node was officially lit up on October 15 and completed same-day (0Day) adaptation for StepFun’s Step 5 Preview large model. This marks the first deployment of the industry’s largest-scale super node product.

Yunding Technology and Huawei signed a memorandum of cooperation on AI computing centers and intelligent agents on October 14. The two sides will build intelligent computing infrastructure based on the Ascend computing base and provide token services to industry customers.

Dahua Technology’s production line with an annual capacity of 20,000 Ascend servers has been signed to settle in Wangcheng, Hunan. The project involves an investment of RMB 100 million and is expected to generate RMB 1 billion in additional annual output value after reaching full production, making the central China hub a new high ground for expanding domestic computing capacity.

The commercialization of domestic GPUs is accelerating. MetaX expects net profit attributable to shareholders of RMB 587 million to RMB 717 million for the first three quarters of 2026, with non-recurring net profit of RMB 223 million to RMB 302 million, crossing both profitability lines for the first time. Moore Threads has already achieved quarterly book profitability; Biren Technology narrowed its first-half loss by 76.4% year on year while revenue grew nearly 20-fold; and Enflame Technology has set a timetable for consolidated profitability in 2026–2027.

Lenovo Group announced at the OCP Global Summit that it has begun global delivery of NVIDIA’s Vera Rubin NVL72 rack-scale AI system. The system integrates 72 Rubin GPUs and 36 Vera CPUs and uses NVLink 6 interconnect. Throughput per megawatt is up to 30 times higher than the previous generation, inference cost per million tokens can be reduced to as low as 1/45 of the original, and only one-quarter as many GPUs are needed to deliver equivalent AI computing power.

In a podcast interview this week, Jensen Huang said new orders for NVIDIA are pouring in almost weekly, or even daily, with large cloud service providers, emerging AI cloud operators and equipment manufacturers all placing substantial orders. However, power shortages, unfinished data centers and component constraints are still slowing NVIDIA’s supply expansion.

A Mysteel weekly report showed that rental prices for high-end domestic training GPUs were flat across the board, with price levels for H100, H800 and H200 remaining firm. Monthly package prices for RTX 5090 rose noticeably in the Yangtze River Delta and Beijing-Tianjin-Hebei regions. Meanwhile, a 128-card H100 cluster resource was released for lease, with a five-year locked term and fixed price, located in East China, at a quoted price of RMB 84,000 per unit per month. Short-term H100 whole-machine prices are expected to remain firm at high levels.