SK Hynix HBM4 mass production and shipment: memory bandwidth is becoming a new bottleneck in the computing power competition

When global AI companies are competing for the number of GPUs, what really blocks the expansion of computing power is often not the chip itself, but the "blood vessel" that feeds data to the chip - high-bandwidth memory. On July 29, SK Hynix disclosed a development that has direct implications for the AI ​​infrastructure supply chain: the new generation of high-bandwidth memory HBM4 has been mass-produced and shipped in Q2. The production scale will be expanded in the second half of the year, and samples of the upgraded version HBM4E have been delivered to customers.

Why HBM is so critical

HBM (High Bandwidth Memory) is the core key component of AI large model training and reasoning. Its capacity and bandwidth directly determine the performance upper limit of AI chips and computing clusters. The essence of large model training is the repeated reading and writing of massive parameters - no matter how fast the GPU calculates, if the memory cannot keep up with the data feeding speed, the computing power will be idle. This is why the release of each generation of Nvidia’s flagship GPU is almost always accompanied by a simultaneous upgrade of HBM specifications.

The mass production of HBM4 means that high-bandwidth memory technology has officially entered a new generation. For downstream AI chip manufacturers such as Nvidia, this brings higher bandwidth and larger capacity memory support, which is expected to alleviate the increasingly acute memory bottleneck in large model training.

Reconsolidation of the right to speak in the supply chain

From an industrial perspective, SK Hynix has maintained a long-term lead in the HBM market. The mass production progress of HBM4 and the sample delivery of HBM4E have further consolidated its voice in the high-end AI memory supply chain. In the cost structure of AI computing power, the upgrade pace of memory and storage links is directly determining the cost and speed of AI chip shipments and computing power deployment - the "arms race" of chip manufacturers is largely driven by the production capacity rhythm of upstream memory manufacturers.

For the domestic AI ecosystem, this incident is also of reference value: in addition to the manufacturing process and architecture of domestic computing chips, the supply stability of high-bandwidth memory is also an unavoidable issue. HBM4E’s mass production schedule, customer adaptation progress, and the technical roadmap of the next generation HBM are long-term windows for observing the direction of competition in AI infrastructure.

Several directions worth tracking in the future:

1. HBM4E mass production schedule: The distance from sample delivery to large-scale mass production determines the graphics memory specifications of the next generation AI chip.
2. Adaptation progress of NVIDIA and other downstream customers: Whether HBM4 enters the flagship GPU supply chain is the key to its success.
3. HBM Price Trend: Changes in supply and demand relationships will be directly transmitted to the unit cost of AI computing power.
4. Catching up with domestic high-bandwidth memory: The pace of breakthroughs in the domestic supply chain in the HBM link affects the upper limit of independent computing power.