Zhipu AI is rapidly expanding into upstream compute infrastructure, including building a 1GW data center, acquiring a Chinese Academy of Sciences-affiliated infrastructure company, and developing in-house chips to reduce reliance on Nvidia. These moves signal independent AI model companies shifting from merely using compute to defining and controlling it.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
From acquiring computing power to unleashing it and ultimately defining it, Zhipu AI is building an independent computing system by constructing a hyperscale data center, acquiring an infrastructure company, and developing proprietary chips to reduce reliance on NVIDIA.
According to AI Frontline, Zhipu AI is constructing a 1GW data center. This scale is extremely rare among independent AI companies; typically only hyperscale cloud providers build facilities of this size. A 1GW power capacity can support the deployment of hundreds of thousands of GPUs, providing ample computing power for Zhipu AI's large model training and inference.
Building its own data center means Zhipu AI will shift from renting computing power to owning it, which not only reduces long-term costs but also grants greater resource control. However, the construction timeline and financial pressure are major challenges; a 1GW data center investment typically runs into billions of dollars.
In addition to building its own data center, Zhipu AI has acquired a 'CAS-backed' infrastructure company for hundreds of millions of yuan. This company specializes in data center infrastructure (Infra), potentially involving key technologies such as liquid cooling, power supply, and networking. Through the acquisition, Zhipu AI can quickly gain a mature infrastructure team and technology, accelerating the data center construction process.
'CAS-backed' typically refers to technology companies with ties to the Chinese Academy of Sciences, which have deep expertise in high-end infrastructure. This acquisition indicates that Zhipu AI is not only pursuing scale but also technological autonomy.
Reports indicate that Zhipu AI is developing its own chips. This is a key part of its 'de-NVIDIA-ization' strategy. Currently, large model training heavily relies on NVIDIA GPUs, which are supply-constrained and costly. If successful, in-house chips would allow Zhipu AI to completely break free from NVIDIA, achieving full-stack autonomy from chips to data centers.
However, developing chips is extremely difficult, requiring massive R&D investment and long-term iteration. It is still unclear whether Zhipu AI's chips are for training or inference, and what architecture they will use. This move is more of a strategic reserve and is unlikely to replace NVIDIA in the short term.
Overall, Zhipu AI's computing layout follows the path of 'acquiring computing power → unleashing computing power → defining computing power.' Early on, it acquired computing power by renting cloud GPUs; now it is unleashing computing power through self-building and acquisitions; and in the future, it aims to define computing power through in-house chips. This marks the extension of independent large model companies upstream into the computing layer, attempting to build vertically integrated competitiveness.
This strategy also reflects an industry trend: as large model competition deepens, computing power has become a core barrier. Whoever controls computing power will have an advantage in model iteration and cost control.
This article primarily draws information from AI Frontline reports, which are second-hand sources. Key details such as the 1GW data center, the acquisition for hundreds of millions, and in-house chip development are based on that report and have not been officially confirmed by Zhipu AI. The report does not provide specific timelines, chip architectures, or other details; some content (e.g., acquisition amount) is vaguely stated.
Zhipu AI is building an independent computing system by constructing a hyperscale data center, acquiring an infrastructure company, and developing in-house chips to accelerate de-NVIDIA-ization. If successful, this strategy could reshape the competitive landscape of large models, but it faces multiple challenges including funding, technology, and time.
Primary report
Primary source
Also covered by 1 sources