China Misses the HBM3 Launch Again, Putting Local AI Accelerator Plans at Risk

China has again run into a delay in launching its own HBM3 production. HBM3 is the memory that modern AI accelerators are hard to build without. The delay hits the plans of local developers to cut their dependence on imports and speed up import substitution in AI infrastructure.

Why HBM3 is critical

HBM3 is high-speed memory that accelerators need to train and run large AI models. For such chips, performance is limited not only by the compute units but also by how fast data moves into and out of memory. That is why even a strong processor loses its value without modern HBM: it cannot reliably deliver the required bandwidth, especially in training, inference and serving large clusters.

This is an especially sensitive point for Chinese accelerator developers. The chip itself can be designed in China or manufactured through available production chains, but without HBM3 the final product ends up weaker or more expensive than its competitors. As a result, dependence on memory supply remains as critical as dependence on lithography, packaging and advanced interconnects. Without local HBM, the whole strategy of replacing foreign accelerators starts to stall.

Where it went wrong

The latest attempt to set up HBM3 production in China has again run into new problems, and they affect both the schedule and the cost of the project. For this kind of product, it is not enough to simply make DRAM dies: manufacturers also have to master multilayer assembly, stack bonding, thermal behavior and a stable yield of good chips. Any failure at these stages quickly turns an ambitious project into a pilot line without sound mass-production economics, and the AI hardware market is not willing to wait.

• Timelines for local AI accelerators are slipping
• Dependence on imported memory persists
• Costs and supply risks are rising
• Scaling server clusters is getting harder

The problem is also that HBM3 is not an auxiliary component but one of the key elements of the whole platform. If memory supply is unstable, accelerator developers cannot confidently plan production volumes, board configurations or the price of the final system. Even with their own architectures and server assembly, a cascading delay follows: first engineering validation stalls, then customer pilots, and then mass deployments in data centers.

A blow to accelerators

For the Chinese market, this means another failure of import substitution in exactly the segment where demand is currently highest. Companies building AI infrastructure want local alternatives to foreign accelerators, but without modern memory such products either reach the market later or lose out on speed, power consumption and compute density. As AI workloads grow quickly, the difference between HBM2E, HBM3 and newer generations stops being marketing and becomes a direct constraint on business.

Hence the main risk: manufacturers may formally unveil their own accelerator but fail to turn it into a mass-market, economically sustainable product. That is a bad scenario for enterprise customers, because they care not only about specs on a slide but also about real availability, service, repeatable supply and a clear roadmap. While HBM3 production in the country is being postponed again, Chinese players will have to either keep depending on imports or accept compromises in performance.

What this means

The HBM3 story shows that in AI hardware the bottleneck is often not the accelerator itself but the memory next to it. Memory has now become the main point of technological dependence. Until China masters stable production of next-generation HBM, its plans for its own AI accelerators will run up against imports, timelines and limited supply volumes.