I could argue that it will only reach 2,781.6 EFLOPS by 2028...
Compute capacity is projected to reach 1,037.3 EFLOPS in 2025 and 2,781.9 EFLOPS by 2028 <Grin> Clearly the writer never learned to round to significant digits to reflect the least accuracy of the assumptions used. I doubt ANY were 5 digits...
Otherwise, I still feel foolish taking profits in my AMAT and LRCX shares, but they pay for remodeling and generate cash in my ROTH to buy back on the dips... that have always shown up eventually.
Supernodes, not servers: How Alibaba, Baidu, Huawei are rewiring AI compute
Staff reporter, Taipei; Levi Li, DIGITIMES Asia
Thursday 2 October 2025
At the Apsara Conference 2025, Alibaba introduced its PanJiu AI Infra 2.0 with a 128-chip supernode, the most-watched launch of the event. The debut underscores how China's cloud leaders are pushing independent AI infrastructure innovation and moving beyond Nvidia's shadow.
In his keynote "The Road to Super AI," Alibaba CEO Eddie Wu said computing is shifting from CPU-centric to GPU-driven, large-model workloads. He forecast that only five to six global platforms may achieve true hyperscale cloud power, with supernode infrastructure as the critical barrier.
Alibaba's 128-chip supernode
The PanJiu AI server delivers high density with an open architecture, housing 128 AI chips in a single rack. Its orthogonal design removes the midplane, avoiding cable-tray bottlenecks and enabling flexible hot-swapping, easier expansion, and greater reliability.
Alibaba Cloud also launched HPN 8.0, a high-performance network interconnect raising GPU bandwidth to 6.4 Tbps. It supports clusters of up to 100,000 GPUs, enabling large-scale training to function as a "single logical computer" and eliminating traditional latency bottlenecks.
Baidu, Huawei, Inspur join supernode race
In the past six months, other Chinese tech giants rolled out their own supernodes. At its August AI Cloud Summit, Baidu unveiled the Kunlunxin supernode with the Baike AI 5.0 and Qianfan 4.0 platforms. Using Kunlunxin chips and high-speed interconnects, Baidu created a "cloud behemoth" for trillion-parameter training, cutting workloads from hours to minutes.
In April 2025, Huawei launched its CloudMatrix 384 supernode in Wuhu, combining 384 Ascend NPUs with 192 Kunpeng CPUs to reach 300 PFlops. Huawei aims to scale to 432 supernodes, forming a 160,000-GPU cluster able to train thousands of 100-billion-parameter models. The company said its cloud power grew 250% year-on-year in 2025, with customer numbers rising fivefold.
Server maker Inspur also released its Metabrain SD200 supernode, integrating 64 AI accelerators into one unit. Tuned for China's leading open-source models, the SD200 expands local compute supply and signals domestic vendors' determination to compete in AI infrastructure.
China's computing power surge
IDC reported China's AI server market was US$19 billion in 2024 and will grow 36.2% to US$25.9 billion in 2025, surpassing US$55.2 billion by 2028. Compute capacity is projected to reach 1,037.3 EFLOPS in 2025 and 2,781.9 EFLOPS by 2028. China has elevated computing to national infrastructure, alongside energy and transportation as economic pillars.
Globally, Meta, Microsoft, OpenAI, and xAI are accelerating 100,000-GPU cluster deployments. Chinese firms, constrained by GPU supply, are pursuing breakthroughs via domestic AI chip design and high-speed interconnects.
Collectively, Alibaba's PanJiu AI Infra 2.0, Baidu's Kunlunxin supernode, Huawei's CloudMatrix, and Inspur's Metabrain SD200 illustrate the rapid expansion of China's AI compute ecosystem. For China's cloud giants, supernodes are no longer about specs alone—they are the entry ticket to the next era of cloud computing.
Article edited by Jack Wu
digitimes.com |
