First reported 2026-10-08 by GlobalFoundries (official announcement via GlobeNewswire)
GlobalFoundries Lands $2B TSMC Deal to Add First U.S. CoWoS Interposer Source
GlobalFoundries announced on October 8, 2026 a multi-year, approximately $2 billion manufacturing agreement to produce silicon interposers — including embedded deep-trench capacitors — for TSMC's CoWoS advanced packaging ecosystem at its Malta, New York fab, with a five-year initial term and volume production ramping in H1 2028. The deal directly targets the interposer bottleneck in CoWoS, which is reported sold out through 2026, and establishes the first U.S.-based interposer source for AI accelerator packaging. Separately, SK hynix and Samsung advanced 16-layer HBM4 (48GB) packaging schedules on Oct. 9, 2026, while Nvidia's Vera Rubin platform baselines on 12-layer, 36GB HBM4 stacks.
Outsourcing interposer supply to a mature-node specialist lets TSMC redeploy capex toward higher-value CoWoS bonding steps while easing a structural choke point for H200/B100-class and Rubin-class accelerators — but relief is dated to 2028, so 2027 packaging allocation and pricing leverage remain firmly with TSMC. The parallel 16-layer HBM4 push confirms 48GB stacks will be a niche high-density tier, as JEDEC's 775µm package ceiling and thermal limits keep mainstream Rubin orders at 12-Hi/36GB.
The brief states no component price move for this cost line, so no BOM change is modeled.
Chips affected — model the economics
This signal touches these AI accelerators. Open any in the cost calculator, pre-filled with its full BOM.
Sources
Related data & tools
Silicon Analysts Research
All benchmarks and data are derived from publicly available sources (earnings calls, press releases, analyst reports, regulatory filings). Figures are estimates for educational purposes only and should not be used as the sole basis for business or investment decisions. Public sources only — no insider data. Generated from the cited reports; verify against the originals before acting. Full Terms & Data Provenance