Intelligence/Brief
The CoWoS bottleneck: why GPU supply still gates 2026 model roadmaps
Every Blackwell, Ironwood, and Trainium2 chip flows through one TSMC packaging line. Capacity doubled in 2026 — and every frontier lab is still constrained.
Published May 20, 2026
Compute supply for the entire frontier AI industry in 2026 is gated by a single TSMC capability: CoWoS-L advanced packaging. NVIDIA GB200 / GB300, Google TPU v7 Ironwood, AWS Trainium2, and Marvell custom silicon all require CoWoS-L for their HBM-attached interposers. There is no second-source option at volume.
What changed in 2026
TSMC's CoWoS-L line ran at ~40,000 wafers/month through 2025. Aggressive capex in 2024 and 2025 brought it to roughly 80,000 wafers/month by Q2 2026 — a doubling, faster than TSMC's usual cadence for new packaging tech. NVIDIA receives the lion's share of allocation; Google, AWS, and Marvell split most of the remainder.
Even at 2x capacity, every major customer is short. GB300 NVL72 racks Microsoft and Oracle ordered for the Stargate Abilene phase 2 buildout are arriving 1–2 quarters later than originally promised. xAI's Colossus 2 GB200 target was rebased from 550k units to ~400k by mid-year.
The downstream effect on models
- Frontier training delays. OpenAI's GPT-5.5 Pro shipped on schedule but the next major model (rumored "GPT-6") was pushed from a target Q3 2026 release to early 2027 in part due to compute timing.
- Inference pricing floor. No vendor has dropped frontier-tier API pricing materially in 2026; gross margins on GPU-side inference are protected by physical scarcity.
- Custom silicon premium. Anthropic (Trainium2) and Google (TPU v7) effectively get higher CoWoS allocation than a NVIDIA-only buyer of equivalent dollar value, because TSMC prices the custom packaging path differently. This is part of the structural reason Anthropic and Google can defend lower API prices.
What to watch
- CoWoS-L capacity targets for 2027 — TSMC's investor-day guidance suggests another doubling, but Apple's M-series transition to CoWoS could absorb meaningful slack.
- HBM4 availability — SK Hynix's ramp is the second bottleneck behind packaging.
- Whether Samsung's CoWoS-equivalent lands at volume — that's the only realistic second source on a 12–18 month horizon.
Linked entities
Underlying signals
- Compute clusterBloomberg / company filings ↗
xAI Colossus 2: targeting 1M GPUs across Memphis + new Mississippi site
Colossus 2 — xAI's expansion target — aims for 550k Blackwell-class GPUs (GB200/GB300) in 2026, scaling toward 1M total accelerators by year-end. A second 2GW campus in DeSoto County, MS, is under construction to host the bulk of the buildout. Power deals announced with the Tennessee Valley Authority and Mississippi Power.
Accelerators: 1M · GB200/GB300Power: 2 GW Google TPU v7 "Ironwood": 9,216-chip pods, 42.5 exaflops, dedicated to inference
Announced at Google Cloud Next April 2026, TPU v7 (Ironwood) is the first Google TPU generation purpose-built for inference rather than training. Each pod scales to 9,216 chips delivering 42.5 exaflops of FP8 compute, with 192GB HBM3e per chip. Powers Gemini 3 Pro / 3.5 inference at Google scale. Annual TPU spend ramped to over $40B for FY26.
Accelerators: TPU v7 Ironwood- GPU supplyNVIDIA GTC + supply chain reports ↗
NVIDIA GB300 ships at volume Q2 2026, replacing GB200 in hyperscaler deals
GB300 NVL72 began volume shipments in April 2026, six months ahead of original schedule. Each rack delivers 1.4x the FP8 throughput of GB200 NVL72 with the same power envelope. Microsoft, Oracle, and xAI are the largest Q2 2026 takers. CoreWeave disclosed first GB300 deployment May 7.
Accelerators: GB300 - GPU supplyDIGITIMES / TSMC investor day ↗
TSMC doubles CoWoS-L capacity to ~80k wafers/month for 2026 AI demand
TSMC's CoWoS-L advanced-packaging capacity — the binding constraint on Blackwell + Ironwood + Trainium2 supply — was doubled to ~80,000 wafers/month for 2026. Even after the expansion, every major customer (NVIDIA, Google, AWS, Marvell) remains capacity-constrained. CoWoS-L allocation has become the single most important variable in vendor compute roadmaps.