Multiple Chinese companies signed large GPU compute procurement orders in a concentrated manner.
Aoni Electronics' wholly owned subsidiary Aoni Intelligent successively signed a RMB 1.67 billion GPU compute card procurement contract with Company A, and another procurement contract worth about RMB 1.556 billion, or USD 232 million, with Company B. Paratera announced plans to procure GPU compute servers and supporting IT equipment and graphics cards for a total of no more than RMB 1.31 billion, and additionally planned to lease GPU compute resources, with the contract amount expected to be no more than RMB 1.48 billion. Xingyun Technology signed three compute service contracts, purchasing 43 sets of GPU compute services, with a total tax-inclusive amount of RMB 453 million and a service term of five years. Xianhe Environmental Protection's subsidiary Xingtai Intelligent Computing planned to purchase 63 GPU server complete machines through financial leasing, at a total price of RMB 168 million.
Overseas GPU rental prices saw another round of significant increases.
Compute leasing vendor Nebius issued a price adjustment notice to customers, planning to raise on-demand rents for major NVIDIA GPUs by about 17 percent to 21 percent starting October 1, 2026. H100 rose from USD 3.85 per hour to USD 4.50 per hour, B200 from USD 7.15 per hour to USD 8.50 per hour, and B300 from USD 7.85 per hour to USD 9.50 per hour. Zheshang Securities analysis noted that on-demand prices are the pricing anchor for new customers and small and medium-sized customers, and this round of increases marks an overall upward shift in the marginal retail pricing system for compute.
Crusoe and Thinking Machines signed a USD 65 million annual compute contract.
Data center provider Crusoe announced an annual USD 65 million contract with Thinking Machines Lab, founded by Mira Murati, to provide managed inference services. The compute load will rely on Nvidia B200 server chips, mainly deployed in Crusoe's small 1 MW Spark data center, with the chips mainly used for model inference rather than training.
Musk announced that the number of GPUs in Colossus 2 will double by the end of the year.
Musk said SpaceX's AI supercomputer Colossus 2 project plans large-scale expansion. The data center currently has 110,000 Nvidia GB200 chips and 440,000 GB300 chips. Next week, 220,000 GB300 chips will be put into use, and another 220,000 are expected to come online in November. By the end of the year, the total number of chips will more than double from the current level. The total chip value of the two Colossus projects is estimated at more than USD 100 billion, providing more than 2 GW of computing capacity.
AMD and Anthropic reached a gigawatt-scale GPU strategic cooperation.
AMD and Anthropic announced a strategic cooperation to deploy up to 2 GW of AMD Instinct MI450 series GPUs in AMD Helios rack-scale solutions, with the first gigawatt deployment to start in the first half of 2027. The two sides will also launch multi-year engineering cooperation using Claude to accelerate AMD software development and ROCm optimization. AMD committed to making a strategic equity investment of up to USD 5 billion in Anthropic in the future.
Microsoft Azure expanded its cooperation with AMD and deployed hundreds of thousands of Grace Blackwell GPUs.
Microsoft further expanded its strategic cooperation with AMD and will deploy AMD Helios rack-scale solutions at scale on the Azure cloud platform, integrating Instinct MI455X GPUs, EPYC Venice CPUs, Pensando network chips and ROCm software for frontier model AI inference workloads. Meanwhile, Azure has deployed hundreds of thousands of liquid-cooled Grace Blackwell GPUs globally and lit up the Vera Rubin NVL72 system in its laboratory.
Amazon AWS plans to procure more than 3 million Nvidia GPUs by 2029.
AWS and Nvidia reached an agreement to add 2 million Blackwell Ultra, Rubin and Rubin Ultra GPUs before March 2026, after previously committing to add more than 1 million. By 2029, Amazon will purchase more than 3 million Nvidia GPUs in total. This procurement scale reflects that frontier model training, inference and Agentic AI workloads are growing exponentially.
Loongson released the first software version of its self-developed general-purpose GPU accelerated computing platform.
Loongson Technology officially released the first software version of the Loongson accelerated computing platform supporting its self-developed general-purpose GPU, building a complete compute software product system for the intelligent application era. This progress marks a substantive step for domestic general-purpose GPUs from hardware R&D to software ecosystem construction.
Biren Technology won a RMB 174 million domestic GPU order as the sole bidder.
The inquiry procurement results for the Kunlun Lingyi Intelligent Computing Center project, Phase I, were announced. Biren Technology won Package One with RMB 174 million, becoming the sole winning supplier for that package, and is required to complete supply within 45 calendar days after contract signing. The core products are Bili series GPUs, based on the classic GPGPU architecture, and have formed a three-generation product matrix of BR106, BR110 and BR166. It is noteworthy that Package Two, with a budget of about RMB 452 million, was scrapped because fewer than three suppliers submitted response documents, reflecting the concentration and scarcity in the current domestic AI compute supply chain.
The National Computing Interconnection Beijing node entered trial operation, and policies are promoting the expansion and quality improvement of computing facilities.
On September 21, the launch ceremony for the National Computing Interconnection (Beijing) node was held in Beijing. After receiving approval from the Ministry of Industry and Information Technology, the node achieved phased construction results and entered trial operation. It will build a Beijing-Tianjin-Hebei compute coordination hub and promote standardized interconnection and efficient flow of compute resources across regions, entities and architectures. In the same period, the Ministry of Industry and Information Technology and four other departments jointly issued a notice to organize the 2026 National Green Computing Facility recommendation work and support the development of green, low-carbon, intensive and circular computing facilities. Vice Minister of Industry and Information Technology Yu Xiaohui said at the 2026 China International Information and Communication Exhibition that it is necessary to promote the expansion and quality improvement of computing facilities and systematically advance the construction of new-generation communication networks.