GlobalFoundries Is Building the AI Chip Bottleneck a US Home, in a Bite

An AI accelerator is a stack, not a single chip. The compute dies and the memory sit on a silicon interposer that carries the bus between them, and that interposer has been the quiet gate on how many accelerators ship.

• GlobalFoundries will make silicon interposers for TSMC advanced packaging at Malta, New York, under a five-year deal worth about $2 billion.
• The interposer carries the memory bus between the GPU die and the HBM stacks beside it. The packaging step, not the logic fab, has limited accelerator shipments.
• GlobalFoundries calls Malta the first US source of silicon interposers for advanced packaging, and the new lines will add embedded deep trench capacitor components.
• Volume production is expected to ramp in the first half of 2028, so the current constraint does not ease yet.
• The deal is a supply contract. TSMC still runs the CoWoS packaging process, and GlobalFoundries provides the interposer manufacturing.

Read the full analysis of the interposer bottleneck