It is increasingly likely that AMD could turn to its old rival Intel for packaging, as TSMC’s capacity squeeze starts making the AI chip queue look daft.
According to OnMSFT UBS expects AMD, Google, the Fruity Cargo Cult Apple and SpaceX to shift to Chipzilla’s EMIB-T packaging through separate foundry agreements. That would give Intel a stronger position in advanced semiconductor packaging, where everyone suddenly wants capacity yesterday.
Chipzilla plans to offer EMIB-T packaging in higher volumes in 2027. Package yields are already close to 90 per cent, while substrate yields remain much less impressive at around 50 per cent.
EMIB-T is reckoned to cost about 50 per cent less than TSMC’s CoWoS packaging. It uses Through-Silicon Vias to improve power delivery and high-speed communication across stacked components.
The numbers still show Nvidia miles ahead in AI racks. Goldman Sachs expects Nvidia to ship about 50,000 data centre racks in 2026, while AMD is expected to manage only 5,000. That gap is not expected to vanish quickly. Nvidia is projected to ship 92,000 racks in 2027 against AMD’s 13,000, then 148,000 racks in 2028 against AMD’s 15,000.
AMD has been trying to strengthen its AI infrastructure pitch through partnerships and acquisitions. It recently teamed with Cerebras, whose Wafer-Scale Engine puts large amounts of compute and memory on one slab of silicon to reduce communication bottlenecks.
The company has agreed to buy Taalas, which builds chips that physically encode AI model parameters into transistor pathways. The idea is to reduce the need to keep pulling model data from external memory during processing.
For AMD, using Chipzilla packaging would not be a surrender so much as a sign that AI supply chains are getting tight everywhere. TSMC remains the key player, Nvidia keeps the rack lead, and AMD is still hunting for every shortcut it can find.
