HBM's generational upgrade not only needs to keep pace with (or even slightly outpace) the iteration cycles of AI chipmakers like NVIDIA and downstream large model vendors,

which means SK Hynix's HBM must provide near-final engineering samples before NVIDIA's tape-out for chip design, verification, and system integration;
and the actual mass production ramp-up of HBM must also occur slightly earlier than the GPU release, otherwise it will slow down the entire AI computing power iteration process.