On the evening of 12 October, Microsoft chairman and CEO Satya Nadella outlined the company’s latest in-house chip plan: a large-scale deployment of Azure Maia 300 in 2027. Microsoft shares rose 1.14 per cent to US$541.15 that day, pushing its market value back above US$4 trillion.
The Maia case
Nadella’s point is that everyone is now building their own chips. The gold-rush rule, he quipped, was to sell the shovel first and then the shovel factory. Cloud vendors are now picking up the hammer themselves. Maia 300 targets the three clearest problems in AI compute: cutting data movement to raise utilisation, packing high-bandwidth memory for longer contexts, and using a chiplet layout to ease leading-edge manufacturing. A single Maia 300 accelerator carries 18 TB/s of memory bandwidth at a 950 W thermal design, and a 96-accelerator rack delivers 27.6 TB of HBM and 1.7 PB/s of aggregate bandwidth. Microsoft claims 40 per cent more AI token throughput per megawatt than a leading GPU system.
China’s parallel track
The same build-it-yourself race started earlier in China. Huawei’s Ascend line, paired with the MindSpore framework, CANN stack and Atlas hardware, now reaches cloud, industry and energy. Cambricon pursues dedicated AI silicon. Baidu runs Kunlun in search and cloud, and Alibaba ships the Hanguang 800 inference chip. The harder problem is no longer the silicon. It is the software that makes different chips actually run, which is why AI Infra has become the new focus. Wubu Infra closed a RMB 2.3 billion C round this week, with RMB 4.3 billion cumulative, a national record for the category.
Microsoft still cannot fully quit Nvidia. Azure keeps selling Nvidia GPUs, and CUDA remains the default dev base. The real change is the nature of the relationship: one of Nvidia’s largest customers is becoming a rival of consequence on its own tuned workloads.
Editor’s note: This is an adapted translation of the original Sohu report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1087924397_116132.