AMD could sign an agreement with Intel Foundry, the competitor's production unit, to overcome capacity restrictions at TSMC's factories, according to a UBS report. The move comes amid Goldman Sachs projections indicating that NVIDIA will maintain leadership in the artificial intelligence server market.
AI server projections
Goldman Sachs data released on Thursday (13) show that NVIDIA should remain ahead of AMD in rack supply. In 2026, the bank projects 50,000 NVIDIA servers against 5,000 for AMD; in 2027, 92,000 to 13,000; and in 2028, 148,000 to 15,000.
In addition to AMD, Google, Apple and SpaceX are expected to sign individual contracts with Intel, according to the report. The main advantage is the EMIB-T technology, a three-dimensional chip integration technology that costs about 50% less than TSMC's CoWoS.
EMIB-T uses direct silicon connections (TSV) drilled into embedded bridges. The vertical channels transmit signals and power from the bottom of the package to processors and memory on top.
Intel plans to offer EMIB-T at scale in 2027. The yield of the integration process is already close to 90%; substrate production, however, operates at about 50%.
AMD is also betting on alternative technologies. With the partnership with Cerebras, the Wafer-Scale Engine platform concentrates hundreds of thousands of cores and dozens of gigabytes of SRAM on a single silicon wafer. The company's acquisition of startup Taalas made it possible to develop chips that incorporate the parameters of an AI model directly into the transistors, eliminating the need to search for external data.
