推文

@uqbariss · 2026-10-12 21:09

NVIDIA may put LESS memory in its next-gen AI GPUs. TrendForce forecasts a 121% jump in HBM's blended selling price in 2027. Rubin Ultra is now being evaluated with 8-layer HBM stacks instead of the original 12-layer plan. Less GPU memory means less room for KV cache. The next AI infrastructure gold rush could be in memory offloading. Who's building that layer?

曝光 79 · 评论 1 · 点赞 0 · 书签 0 · 曝光/时 16.6899945889044