The rivalry between AMD and Intel could take an unexpected turn. According to an analysis by UBS, AMD could turn to Intel Foundry for certain stages of production, as TSMC’s available capacity remains under pressure. At this stage, this is not an agreement officially announced by AMD or Intel, but rather a forecast by UBS analysts. It is therefore important to classify this information not as a rumor, but as an “analyst prediction.” This hypothesis has gained traction recently, as Goldman Sachs estimates that AMD is likely to lag far behind NVIDIA in the strategic market for rack-based AI infrastructure. It’s also worth noting that AMD’s interest is genuine, as rumors of discussions regarding the 18A and 14A processes date back to last year.
Intel’s EMIB-T packaging could appeal to AMD
According to UBS, Google, Apple, AMD, and SpaceX could use Intel’s EMIB-T packaging technology as part of individual agreements with Intel Foundry. Mass production of this solution is expected to begin in 2027. Intel has reportedly already achieved packaging yields close to 90%, although substrate yields remain limited to around 50%.
The economic benefits could be significant. According to reports, EMIB-T is approximately 50% less expensive than TSMC’s CoWoS. This technology uses Through-Silicon Vias (TSVs) directly integrated into the silicon bridges. These vertical connections allow power and high-speed signals to flow from the bottom of the package to the processors or memory installed above. This architecture also enables 3D stacking solutions.
As things stand, it’s difficult for AMD to challenge NVIDIA…
However, AMD’s challenge isn’t limited to production. Estimates from Goldman Sachs show a significant gap with NVIDIA in complete AI systems. NVIDIA could ship 50,000 racks in 2026, compared to just 5,000 for AMD. In 2027, forecasts stand at 92,000 versus 13,000 units, respectively, and then 148,000 versus 15,000 units in 2028.
AMD is, however, stepping up its efforts to strengthen its offering. The company recently partnered with Cerebras, whose Wafer-Scale Engine combines memory and computing capabilities on a massive interconnected chip. AMD has also reached an agreement to acquire Taalas, a company developing an approach in which the mathematical values of an AI model are directly integrated into the processor’s circuits.
A potential industrial partnership with Intel would therefore be a particularly symbolic development. But above all, it would address a very concrete need: increasing available capacity and reducing packaging costs so that AMD can accelerate its shipments in the face of NVIDIA, which continues to dominate the AI infrastructure market by a wide margin.
