Broadcom cut its total 2026 CoWoS wafer allocation from 250,000 to 215,000 wafers as TSMC's advanced packaging capacity fills up. On Monday, two independent reports confirmed Google is working with AMD on a 10th-generation Tensor Processing Unit. The AI supply chain's hardest bottleneck is not the chip. It is the packaging.
The Packaging Constraint at the Center of Everything
CoWoS (Chip-on-Wafer-on-Substrate) is TSMC's process that integrates logic chips and high-bandwidth memory into a single dense module. Every frontier AI accelerator depends on it. TSMC is expanding CoWoS output from approximately 70,000 wafers per month in 2025 toward 120,000 to 130,000 per month in 2026, with up to $56 billion in capital expenditures allocated for the year.
The expansion has not kept pace with demand. Nvidia alone has reserved 800,000 to 850,000 CoWoS wafers for 2026, roughly 60% of TSMC's total output, per Morgan Stanley analysis. The remaining 40% is split among Broadcom, AMD, Google, Amazon, Apple, and every other ASIC vendor. CoWoS-S costs $750 per chip. CoWoS-L, used in the Blackwell B200, costs $1,100 per chip. Packaging slots book 52 to 78 weeks out. HBM memory is sold out through 2026.
Broadcom, which generated over $20 billion in AI ASIC revenue in FY2025 from six confirmed hyperscaler customers including Google, Meta, and OpenAI, has revised its 2026 CoWoS wafer projection from 250,000 to 215,000 and its Google TPU unit shipments from 3.2 million to 2.7 million.
Google's Hedge
Reports from Dataconomy and Data Center Dynamics published Monday confirm Google is working with AMD on a 10th-generation TPU. Google's existing supply agreement with Broadcom runs through at least 2031, so AMD's role is likely supplemental.
The practical read: when TSMC allocates packaging capacity by customer priority, a supplier with more scheduling flexibility becomes operationally valuable. Google is hedging, not defecting.
AMD's MI350X matches Nvidia's B200 on FP8 throughput in benchmarks. But the CoWoS constraint applies to AMD as well. Nvidia's priority booking means AMD also competes for a minority share of a capacity-constrained process.
What to Watch
The bearish case on Broadcom is simple: if CoWoS supply expands faster than AI model training demand, pricing pressure on custom ASICs follows, and Broadcom's 2026 AI revenue guidance faces cuts. The bull case is that TSMC's $56 billion in capex is years away from yielding production capacity at scale, keeping the bottleneck intact through 2027.
Google's AMD partnership adds a second sourcing relationship but does not solve the packaging problem. The physical constraint governing the entire AI buildout is TSMC's advanced packaging lines, not the chip design sitting on top of them.
Follow @UTXOMacro for daily Bitcoin, macro, and AI breakdowns.