NVIDIA's Feynman roadmap points to customized HBM4E demand

NVIDIA's post-Rubin roadmap is already shaping expectations for HBM suppliers. The queued source payload says the company is working on Feynman GPU designs for TSMC's A16 node, with mass production scheduled for the second half of 2028. The platform is described as using 3D chiplets, advanced 2.5D or panel-level packaging, and customized high-bandwidth memory. For RamTrend, the key point is the memory customization. The payload says NVIDIA is expected to work with partners on HBM, likely HBM4E, and that the customized memory could include a specialized base die with logic such as a controller or packet-processing function. That would extend the trend of AI accelerators requiring memory stacks tailored to system architecture, not just more bandwidth. The pricing impact is not immediate because the roadmap is tied to a 2028 production window. Still, the direction supports a longer-term view that premium HBM demand will remain closely linked to GPU platform design and advanced packaging capacity.