Summary in Plain Language
This news reports that in 2026, JD Cloud announced a major industry breakthrough at its technology conference: they plan to build a super computing cluster equipped with 100,000 domestic GPU chips, all using Moore Threads chips from domestic manufacturers as the core computing power. This cluster will be dedicated to training large models and powering “physical AI” applications such as intelligent robots and industrial robotic arms that can operate in the real world, and it will be made available to the entire industry for use.
This is the first time that a leading domestic cloud provider has replaced all the core computing components in a cluster of this scale with domestically produced chips, precisely aligning with the policy objectives of the Ministry of Industry and Information Technology (MIIT) during the 14th Five-Year Plan period to promote the development of domestic AI computing clusters. It marks a significant milestone for domestic GPUs, which were previously considered to have average performance and only suitable for limited use. With this breakthrough, the cost of AI computing power will significantly decrease, lowering the barriers for various industries to adopt AI technologies.
---
Detailed Explanation
1. What exactly is this 100,000-chip domestic GPU cluster?
Many people may not have a clear idea of what such a cluster entails. For comparison, all previous domestic cloud providers using 100,000-chip clusters relied on NVIDIA chips from overseas. Domestic GPUs could at most handle a few thousand chips, and even managing 10,000 chips stably was considered top-tier performance in the industry. Achieving 100,000 chips is not just a matter of multiplying the number by ten; it’s like connecting 100,000 high-performance gaming computers simultaneously. If even one of them fails, the training of a large model could be halted for months. Domestic GPUs previously lacked the stability and interconnectivity required for this level of performance. This time, the performance of domestic GPUs has been increased by tenfold, creating a real, practical super computing resource that can handle large model training.
2. Why now?
This initiative is the result of a perfect alignment of policy and industry trends:
- The MIIT’s 14th Five-Year Plan explicitly encourages the construction of domestic AI computing clusters with thousands or tens of thousands of chips and provides various forms of support. Previously, companies were hesitant to use domestic chips due to concerns about potential issues, but now the policy provides clear guidance.
- The demand for “physical AI” applications (such as logistics robots, industrial robotic arms, and household service robots) that can operate in the real world has surged, increasing the need for computing power by several orders of magnitude. Importing chips was not only expensive but also subject to supply restrictions. JD Cloud, with its presence in retail, logistics, and industrial sectors, has a real need for such a large computing resource and is not just showcasing it for show.
3. It’s more than just buying chips – it’s a comprehensive integration from the ground up:
Previously, when cloud providers used domestic GPUs, they were merely meeting minimal requirements. The chips were domestic, but the accompanying cloud platforms and model training software were still based on overseas technologies, resulting in only 30% of the chip’s potential being utilized, leading to slow performance and frequent issues. This time, JD Cloud and Moore Threads are collaborating from the chip level, adjusting the chip parameters to meet the needs of JD Cloud’s large model training systems. The cloud platform tools are customized for domestic chips, and the entire training process is jointly developed. This ensures that the chips can perform at their full capacity, and the needs of JD Cloud’s real-world applications can be used to continuously optimize the chips, creating a positive cycle of “real-world applications → model training → chip improvement.”
4. The real benefit: a significant reduction in AI computing costs for the entire industry:
Once the 100,000-chip cluster is operational, the price of AI computing power will drop significantly. Training a medium-sized large model with imported chips previously cost millions; now, with domestic chips, the cost is already half lower. The scale of the cluster further reduces the overall cost. Small factories and startups that couldn’t afford AI before can now utilize it for tasks like intelligent scheduling of robotic arms in workshops. Ordinary people can also use AI to create long videos or 3D content more affordably. In the future, household service robots that can perform practical tasks may be available for a few thousand yuan, rather than the current “toy-like” models that can only perform basic tasks.
5. An industry milestone: domestic GPUs have moved from being marginally useful to being ready for widespread commercial use:
The industry’s stereotype of domestic GPUs as having poor performance and being suitable only for simple AI tasks has changed. With JD Cloud’s achievement, it’s proven that domestic GPUs can handle large-scale clusters and support core business applications. Other cloud providers and companies no longer need to struggle with the choice between imported and domestic solutions. The entire domestic AI computing ecosystem, from chips to servers to cloud platforms, is now thriving, free from the constraints of overseas manufacturers.