AMD to Bet $10 Billion on Taiwan to Win the AI Packaging Race Against Nvidia

The investment underwrites AMD Helios, which pairs Venice CPUs with Instinct MI450X GPUs and is slated for multi-gigawatt deployment in late 2026, signaling that the AI race is shifting from chip design to supply-chain capacity.

AMD will invest more than $10 billion in Taiwan’s chip ecosystem. The money will widen partnerships. It will also scale advanced packaging for AI hardware, but the headline number is not the real story.

For years, the limit on top AI chips was transistor density. Now that has shifted. Today, the hard part is packaging. In other words, firms must join chiplets, memory and interconnects into one working rack. AMD is spending to control that step.

At the center sits a technology AMD calls Elevated Fanout Bridge, or EFB. It is a wafer-based 2.5D bridge interconnect. AMD is building and testing it with ASE and SPIL. There are other partners, who are helping too. AMD says the design lifts bandwidth and cuts power use. It will also support the new 6th Gen EPYC server chips, codenamed Venice. As a result, customers get more speed per watt. And they get it within the same power and cooling limits.

ALSO READ: Nasdaq Rebounds as Meta Teams With AMD

AMD also hit a key milestone with partner PTI. Together, they qualified the first 2.5D panel-based EFB interconnect. Panel-based packaging uses large square substrates, not round wafers. So it can cut the cost per chip as volumes rise. AMD says the step delivers high bandwidth at scale. It also improves the economics of AI systems.

“As AI adoption accelerates, our global customers are rapidly scaling AI infrastructure to meet growing compute demand,” said Dr. Lisa Su, chair and CEO of AMD. The company tied the spend to its rack-scale plans. In short, it sees Taiwan as the engine behind its roadmap.

That roadmap runs through AMD Helios. It is the company’s rack-scale platform. Also, it pairs Venice chips with AMD Instinct MI450X GPUs. AMD says Helios is on track for multi-gigawatt rollouts in late 2026. Meanwhile, partners such as Sanmina, Wiwynn, Wistron and Inventec are moving it into mass production. The list also adds Unimicron, Nan Ya PCB, Kinsus and AIC. Clearly, the supply base behind one AI rack has grown wide.

The rivalry is easy to see. For years, Nvidia locked up packaging and memory supply. Those limits shaped what rivals could buy. So AMD is not just competing on chip design. Instead, it is funding its own packaging base. This move secures the output it needs, and in turn, that output turns a strong GPU into a shippable rack. For buyers, the question is changing. It is no longer just whose chip is fastest, but whose supply chain can deliver at scale.