China delays HBM3 launch again: local AI accelerator makers' plans at risk

China has once again encountered delays in launching its own HBM3 memory production — a memory without which it is difficult to manufacture modern AI accelerators. This undermines the plans of local developers to reduce import dependence and accelerate import substitution in AI infrastructure.

Why HBM3 Is Critical

HBM3 is high-bandwidth memory essential for accelerators to train and run large AI models. The performance of such chips is constrained not only by computational units, but also by how fast data flows to and from memory. This is why even a powerful processor without modern HBM loses its purpose: it cannot consistently deliver the required bandwidth, especially in training, inference, and serving large clusters.

For Chinese accelerator developers, this is a particularly sensitive point. The chip itself can be designed domestically or manufactured through available supply chains, but without HBM3, the final product ends up weaker or more expensive than competitors. As a result, memory supply dependence remains as critical as dependence on lithography, packaging, and advanced interconnects.

Without local HBM, the entire strategy for replacing foreign accelerators starts to stall.

Where the Problem Lies

Another attempt to establish HBM3 production in China has once again run into new problems. The issue affects both timelines and project costs. For such a product, it is not enough to simply manufacture DRAM dies: one must also perfect multi-layer assembly, stack bonding, thermal management, and stable chip yields. Any failure at these stages quickly turns an ambitious project into a pilot line without proper mass production economics, and the AI hardware market is not willing to wait.

• Timelines for local AI accelerators are shifting
• Dependence on imported memory persists
• Manufacturing costs and supply risks are growing
• Scaling server clusters becomes more difficult

The problem is further compounded by the fact that HBM3 is not a supporting component but one of the key elements of the entire platform. If memory supplies are unstable, accelerator developers cannot confidently plan production volumes, board configurations, and system pricing. Even with proprietary architectures and server assembly in place, a cascading delay emerges: first, engineering validation stalls, then customer pilots, and finally mass deployments in data centers.

Impact on Accelerators

For the Chinese market, this means another failure of import substitution precisely in the segment where demand is currently highest. Companies building AI infrastructure want to obtain local alternatives to foreign accelerators, but without modern memory, such solutions either reach the market later or fall short in speed, power consumption, and compute density. Against the backdrop of rapid AI workload growth, the difference between HBM2E, HBM3, and newer generations ceases to be marketing and becomes a direct business constraint.

Hence the main risk: manufacturers may formally showcase their own accelerator but fail to develop it into a mass-market, economically sustainable product. For corporate customers, this is a poor scenario because they care not only about specs on a slide but also about actual availability, service, supply repeatability, and a clear roadmap. While HBM3 production in the country is delayed once again, Chinese players will have to either continue depending on imports or accept performance compromises.

What This Means

The HBM3 story shows that in AI hardware, the bottleneck often lies not in the accelerator itself but in the memory beside it. Memory has now become the main point of technological dependence. Until China masters stable production of next-generation HBM, plans for indigenous AI accelerators will be hampered by imports, timelines, and limited supply volumes.