AMD May Turn to Intel Foundry as TSMC Capacity Pressure and NVIDIA’s AI Rack Lead Intensify
AMD could be preparing for one of the most surprising semiconductor partnerships in years. According to recent analyst commentary, the company may strike a foundry and advanced packaging deal with Intel as ongoing capacity constraints at TSMC continue to pressure chipmakers competing in the booming AI hardware market.
The move would be notable because AMD and Intel have been fierce rivals for decades across CPUs, data center processors, and PC hardware. But in the current AI chip race, access to advanced packaging capacity has become just as important as chip design itself. With demand for AI accelerators, high-bandwidth memory integration, and rack-scale systems surging, AMD may need additional manufacturing and packaging support to keep shipment momentum alive.
Analysts at UBS suggest that AMD could be among several major technology companies looking to use Intel’s EMIB-T packaging technology. Google, Apple, and SpaceX were also mentioned as potential customers for Intel’s advanced packaging services. If these deals materialize, they could provide Intel’s foundry business with a major boost as it tries to position itself as a serious alternative in the global semiconductor supply chain.
Intel is expected to begin offering EMIB-T in volume around 2027. The technology is reportedly progressing well on package yields, which are said to be nearing 90 percent. However, one key challenge remains: substrate yield is still around 50 percent, meaning Intel must improve production efficiency before EMIB-T can scale at the level needed by major AI and high-performance computing customers.
The appeal of EMIB-T is easy to understand. The technology is estimated to be roughly 50 percent cheaper than TSMC’s CoWoS packaging, a crucial advantage at a time when AI chip costs are climbing rapidly. EMIB-T also uses Through-Silicon Vias drilled directly through embedded silicon bridges. These vertical pathways allow power and high-speed signals to move from the bottom of the package through the bridge and into the processors or memory stacked above.
In simpler terms, EMIB-T can support advanced 3D chip stacking, helping processors and memory communicate faster and more efficiently. That matters greatly for AI workloads, where bandwidth, latency, thermal performance, and power delivery can determine how competitive a system is.
While AMD is making aggressive moves to strengthen its AI hardware roadmap, analysts remain cautious about its ability to close the gap with NVIDIA in data center rack shipments. Forecasts from Goldman Sachs suggest NVIDIA is likely to maintain a commanding lead for several more years.
The projected shipment gap is significant. In 2026, NVIDIA is expected to ship around 50,000 AI server racks, compared to about 5,000 for AMD. In 2027, NVIDIA’s shipments are forecast to rise to 92,000 racks, while AMD is expected to reach 13,000. By 2028, NVIDIA could ship approximately 148,000 racks versus AMD’s 15,000.
These estimates imply that NVIDIA may ship around 10 times more AI racks than AMD in 2026, about 7 times more in 2027, and nearly 10 times more again in 2028. For AMD, that highlights both the size of the opportunity and the scale of the challenge.
NVIDIA’s advantage comes from more than just powerful GPUs. The company has built a full AI infrastructure ecosystem around its accelerators, networking hardware, software stack, and rack-scale designs. Customers building large AI clusters often want complete, validated systems rather than individual chips, and NVIDIA has been extremely successful at selling that broader platform.
AMD is not standing still, however. The company has been working to improve the appeal of its rack-scale AI offerings through partnerships and acquisitions that could give it new ways to compete.
One major step is AMD’s partnership with Cerebras. Cerebras is known for its Wafer-Scale Engine, a massive silicon design that places an enormous amount of compute and memory onto a single interconnected wafer. Instead of relying on many smaller chips communicating across external networks, Cerebras’ approach allows hundreds of thousands of compute cores and large amounts of on-chip SRAM to work together with far fewer bottlenecks.
This kind of architecture could help AMD build more differentiated AI systems, especially for customers seeking alternatives to conventional GPU clusters.
AMD has also agreed to acquire Taalas, a company working on AI chips that physically embed a model’s mathematical values into transistor pathways. The concept is highly specialized: rather than constantly loading weights from memory, the circuit itself effectively becomes both memory and computation engine. This can create major efficiency advantages, although the chip is designed for a specific model and lacks the flexibility of general-purpose AI accelerators.
Together, these moves show that AMD is exploring multiple paths to gain ground in artificial intelligence hardware. It is improving its GPU roadmap, pursuing rack-scale systems, working with unconventional AI compute companies, and potentially tapping Intel’s packaging technology to ease supply constraints.
Still, the road ahead remains difficult. NVIDIA’s lead in AI server racks appears likely to remain substantial through at least 2028, based on current analyst forecasts. AMD may continue to grow, but catching NVIDIA in full-stack AI infrastructure will require more than strong chips. It will need manufacturing access, advanced packaging capacity, software maturity, system-level execution, and strong customer adoption.
A potential Intel foundry deal would be a remarkable twist in the semiconductor industry. But in the AI era, old rivalries may matter less than supply chain strength. For AMD, using Intel’s EMIB-T packaging could be a practical step toward staying competitive in a market where every available wafer, package, and rack counts.
