Rubin GPU could face delays
A report warned NVIDIA’s Rubin GPU may face possible delays as HBM4 supply falls short while Google TPU demand surges — a supply tension story tied to TSMC capacity competition (benzinga.com). That kind of foundry squeeze elevates urgency for startups to consider locking capacity or on‑prem options.
Samsung and SK hynix were named as primary HBM4 suppliers for NVIDIA’s Rubin program reported), with TrendForce projecting SK hynix to supply roughly half of NVIDIA’s total HBM allocation in 2026 while Samsung leads Rubin‑specific HBM4 shipments.(trendforce.com) Taiwanese supply‑chain reporting says several memory vendors are redesigning HBM4 base‑die and packaging to hit NVIDIA’s higher data‑rate targets, a engineering change that industry sources estimate could push HBM4 shipments—and Rubin wafer starts—back about one quarter.(chinatimes.com) Rubin’s memory architecture has been described in reporting as targeting 16 HBM stacks for a total near 576 GB per GPU and data rates above 10 Gb/s—spec targets that suppliers must qualify to meet.(trendforce.com) TSMC’s CoWoS packaging capacity is being expanded from roughly 75k–80k wafers/month toward a 120k–130k/month target by [late‑2026 reported],(trendforce.com) while multiple industry pieces note NVIDIA has reserved a majority share of advanced packaging slots, tightening availability for TPU and ASIC customers.(wccftech.com) Supply‑side demand from Google’s TPU program has pressured CoWoS allocations—MediaTek/Google TPU orders were described as rising from ~10k to ~20k CoWoS wafers for 2026 in reporting that linked TPU ramp plans to TSMC capacity planning.(trendforce.com) Accounts diverge on Micron’s role: TrendForce and regional press say Micron is not the primary HBM4 supplier for the Rubin flagship and may be positioned for mid‑tier inference modules,(trendforce.com) while Micron’s own releases claim HBM4 variants designed for Rubin entered high‑volume production.(markets.businessinsider.com) TSMC’s multi‑billion dollar capex and CoWoS expansion plans were cited as the industry response to the bottleneck—TSMC budgeting roughly $52–56bn for 2026 capacity build‑out aimed at relieving advanced‑packaging pressure for AI accelerators.(markets.financialcontent.com)
