The stacked bottleneck
Wafers, CoWoS and HBM Gate Every AI GPU
GPUs became the default AI training chip because their thousands of cores match the matrix math of neural networks, and because NVIDIA built CUDA around them over two decades.
- Wafers: Blackwell runs on TSMC 4NP and Rubin on TSMC N3P, the same wafer pool AMD, Apple, Qualcomm and the hyperscaler chips draw on.
- CoWoS: after fabrication, TSMC joins the GPU dies and HBM stacks on a silicon interposer with bridges (CoWoS-L). This capacity is separate from wafer starts and has limited NVIDIA shipments since the Hopper ramp in 2023.
- HBM: a Rubin GPU carries eight HBM4 stacks, 288GB in all. SK hynix led HBM supply with half of the market in the second quarter of 2026, ahead of Samsung and Micron.
Each input expands on its own multi-year schedule, so relieving one does not relieve the others. Detail is on CoWoS and HBM.
GPU families
GPU Families, Products and Process
| Family | Products | Process and packaging | Memory and suppliers |
|---|---|---|---|
| NVIDIA Blackwell (data center) | B200 and GB200; B300 Blackwell Ultra; GB300 NVL72 racks | TSMC 4NP, two dies per package; CoWoS-L | 192GB HBM3E (B200), 288GB HBM3E (B300); SK hynix lead, Samsung and Micron |
| NVIDIA Rubin (data center) | Rubin GPU paired with the Vera CPU in Vera Rubin racks; Rubin Ultra follows in 2027 | TSMC N3P; CoWoS-L | 288GB HBM4 at 22 TB/s; HBM4 from SK hynix, Samsung and Micron |
| NVIDIA GeForce RTX 50 (consumer) | RTX 5090, 5080 and down | TSMC 4N | GDDR7 from Micron, Samsung and SK hynix, a separate memory chain from HBM |
| AMD Instinct (data center) | MI355X; MI455X in Helios racks (2026) | TSMC N3 (MI355X); TSMC N2 compute dies on N3 base dies with hybrid bonding and CoWoS-L (MI455X) | 288GB HBM3E (MI355X); 432GB HBM4 (MI455X), Samsung the primary HBM4 supplier |
| AMD Radeon (consumer) | Radeon RX 9070 series | TSMC N4P | GDDR6 |
| Intel data center GPUs | Gaudi 3; Crescent Island inference GPU announced for 2026 sampling | TSMC 5nm (Gaudi 3) | HBM2E (Gaudi 3); LPDDR5X (Crescent Island) |
| China GPUs | Moore Threads MTT; MetaX; Biren; Huawei Ascend (see AI Accelerators) | SMIC; some earlier designs at TSMC before export controls | Moore Threads and MetaX listed on Shanghai's STAR Market in late 2025 |
Supply risk
Deployment and Supply Risk
| Platform | Where it is used | Primary supply risk |
|---|---|---|
| NVIDIA Blackwell and Rubin | AI training and inference clusters at hyperscalers, neoclouds and labs | TSMC wafers, CoWoS-L capacity and HBM, all needed at once; HBM4 qualification |
| AMD MI355X and MI455X | Microsoft, Oracle, Meta and OpenAI clusters | TSMC N2 and N3 shared with AMD EPYC; HBM4 supply; ROCm software maturity |
| Consumer GPUs | Gaming, workstations, local AI inference | Lower allocation priority than data center parts; GDDR7 supply |
| China GPUs | Domestic AI clusters | SMIC capacity and yield; no EUV; HBM access |
CUDA lock-in
CUDA Concentrates Demand on One Architecture
NVIDIA's strongest position is its software: PyTorch, JAX and most AI research code assume CUDA, cuDNN, NCCL and TensorRT. AMD's ROCm now runs the major frameworks, but porting custom kernels and finding engineers who know ROCm still cost time. That switching cost concentrates AI demand on NVIDIA hardware in a way a second chip supplier alone does not undo, which is why the hyperscalers build their own accelerators (see AI Accelerators).
Export controls
Export Controls Split GPU Supply in Two
- United States: controls since October 2022 bar NVIDIA's and AMD's top data center GPUs from China. The H20, NVIDIA's China-compliant part, needed a license from April 2025. In January 2026 Washington allowed H200 exports under license.
- China: Beijing first held back H200 purchases while it promoted domestic chips, then let selected companies buy limited quantities. NVIDIA's first H200 shipments came to less than 1% of its $89 billion data center revenue in the quarter ended July 26, 2026.
- Domestic alternatives: Huawei's Ascend 910C is made at SMIC, though a TechInsights teardown found TSMC-made dies obtained before controls tightened. Cambricon's revenue more than doubled to about $890 million in the first half of 2026, and Moore Threads and MetaX GPUs reached the STAR Market in late 2025.
China's wider position is on the China Bifurcation spotlight and the Huawei and HiSilicon spotlight.
Record
Page Record
- Covers
- Data center and consumer GPUs, the wafer, packaging and HBM constraints, CUDA, and export controls
- Sources
- NVIDIA and AMD product disclosures; South China Morning Post on H200 shipments (August 2026); Caixin on H200 licensing (January 2026); TechInsights via Bloomberg on Ascend 910C; Cambricon filing (August 2026); Counterpoint HBM share
- Last verified
- October 2026