Registers
▶ See it in 3D The Memory Stack region
The handful of numbers the GPU is touching this exact instant — like the figures lit up on a calculator's screen. Far too tiny to invest in on its own; the value sits with whoever designs the chip (NVIDIA, AMD) and whoever prints it (TSMC).
What it is
The fastest memory that exists. Each GPU Streaming Multiprocessor holds ~256 KB of register file — the operands of instructions in flight. ~32 MB aggregate on an H100.
The business
Not separately investable — it ships inside the GPU. The economics flow to whoever designs the GPU (NVIDIA, AMD) and the foundry that prints it (TSMC).
Bottleneck severity
Who makes it
NVIDIANVDA
Buys the most HBM, NVLink and optics on earth and pre-buys supply to lock rivals out: ~60% of CoWoS, $2B each into Coherent and Lumentum (Mar 2026), $2B into Marvell. B300 (288 GB HBM3E) is the shipping reference GPU; Rubin (288 GB HBM4, 22 TB/s) samples Q4 2026, volume Q1 2027.
AMDAMD
MI-series GPUs with leading HBM capacity; the anchor of UALink. Zen 6 EPYC "Venice" launched July 2026 with roughly a year before Intel answers; MI400 carries the first UALink production silicon (H2 2026).
TSMCTSM
Makes the GPU die AND the CoWoS packaging that bonds HBM to it — and now the logic base die under SK Hynix's HBM4 too. CoWoS is heading for ~130k wafers/month by end-2026 and is still sold out into 2027 (NVIDIA alone holds ~60%); it is farming 240–270k wafers/year of front-end steps out to Amkor and SPIL. A Taiwan disruption stops the whole stack.
Metrics that move this layer
- On-die SRAMgrows slowly per node — SRAM scaling has stalled, a quiet driver of HBM reliance