Alibaba Cloud Releases GPU Pooling Solution, Significantly Reducing Compute Costs
Alibaba Cloud unveiled its computing pooling solution named "Aegaeon" at the SOSP 2025 conference. This solution utilizes technologies like token-level virtualization to allow multiple AI models to dynamically share GPU resources. In tests, the number of GPUs required to serve dozens of large models with 72 billion parameters dropped from 1,192 to 213, a reduction of up to 82%, substantially lowering the hardware costs for serving large models.
Domestic General-Purpose GPU "Xiyun C600" Launched in Nanjing
MetaX Technologies officially launched its first fully domestic general-purpose GPU – the Xiyun C600 – in Nanjing. This chip integrates large-capacity memory and mixed-precision computing power. MetaX's domestic thousand-card cluster, developed in collaboration with the Chinese Academy of Sciences, has successfully completed full-parameter training for multiple large models, demonstrating that domestic computing power now possesses the capability for large model pre-training.
Nanjing Pukou Forms Comprehensive IC Industry Chain Ecosystem
As the birthplace of the "Xiyun C600," the Nanjing Pukou Economic Development Zone has gathered over 300 upstream and downstream integrated circuit companies, forming a complete industry chain encompassing IC design, wafer manufacturing, packaging and testing, and material support. This industrial ecosystem provides crucial support for the continuous innovation and development of domestic chips.
First in Europe! Denmark's DCAI Deploys NVIDIA DGX B300 Systems
Danish AI infrastructure provider DCAI announced it will deploy NVIDIA DGX B300 AI infrastructure in its Gefion supercomputer. This makes them one of the first in Europe and the first in Denmark to acquire this system. This move will provide stronger computing power for AI training and inference for researchers and businesses across Europe.
Hyperscale Data to Launch On-Demand NVIDIA GPU Cloud Service
Hyperscale Data announced plans to launch an on-demand NVIDIA GPU cloud service platform in the first half of 2026 via its Michigan AI data center campus. The service will provide instant, flexible access to advanced NVIDIA GPUs like the H100, B200, and B300, helping customers lower the hardware entry barrier for AI and high-performance computing projects.
Google and Anthropic Seal Tens of Billions Compute Deal
Google will provide AI company Anthropic with up to 1 million of its self-developed TPU chips in a deal valued at tens of billions of US dollars. Scheduled for deployment in 2026, these chips will deliver over 1 gigawatt of computing power, aiming to reduce Anthropic's reliance on NVIDIA GPUs.
Oracle Launches World's Largest AI Supercomputer, Serving OpenAI
Oracle unveiled its AI supercomputing cluster named OCI Zettascale10. It integrates 800,000 NVIDIA GPUs, claiming a peak performance of 16 ZettaFLOPS. It is set to become the computational core for OpenAI's future "Stargate" supercomputer and is planned to become available to customers in the second half of 2026.
Global AI Compute Demand Continues to Surge
The cost of developing cutting-edge AI systems is rising sharply, as they require massive computing power to train and run large language models. The tens-of-billions-dollar compute deal between Google and Anthropic is one of the largest transactions in the current AI hardware race, highlighting this trend.
Diversified Compute Architecture Becomes a Development Trend
The industry is actively seeking diversification in compute supply. For instance, Anthropic utilizes both Google TPUs and Amazon's self-developed chips; Alibaba Cloud's "Aegaeon" solution improves the utilization efficiency of existing GPU resources through pooling technology.
Compute Cost Optimization is a Key Challenge
While the scale of computing power continues to expand, how to effectively control and optimize costs has become a key focus for enterprises globally. Whether through Alibaba Cloud's pooling technology or the on-demand cloud services offered by Hyperscale Data, a core objective is to help enterprises access computing power more economically.