AMD lifts its Taiwan supply-chain investment

AMD CEO Lisa Su arrived in Taiwan around October 5 for her second visit in less than five months, mostly closed-door supplier meetings to review delivery schedules into 2027 and reserve more capacity. Su said AMD will increase its previously announced Taiwan supply-chain investment of more than $10 billion as AI demand grows.

In May AMD announced over $10 billion in Taiwan investments through about 2029 for advanced packaging, substrates and manufacturing of rack-scale AI systems, with partners including ASE/SPIL, PTI, Wiwynn, Wistron, Inventec, Unimicron, Nan Ya PCB and Kinsus.

AMD's AI deals have ballooned: reports cite about 14 GW of data center agreements, including 6 GW each with OpenAI and Meta, plus 50,000 MI450 chips for Oracle. Advanced packaging is one of the industry's tightest bottlenecks, and Su is also reported to be talking to Samsung about HBM4 memory.

Some outlets report Su sought capacity coordination even with rivals, an unusual step that shows how tight supply has become.

Modern AI chips combine a compute die with stacks of high-bandwidth memory in one package. Few plants can do this at scale, mostly TSMC and ASE in Taiwan. Without packaging capacity, finished chips cannot become systems, which caused much of Nvidia's recent delivery delays.

The AI bottleneck now runs through packaging, memory, substrates and power, not just GPUs. Securing capacity early is as much a logistics race as a technology one.

Most of the world's most advanced chips are made and packaged in Taiwan, by TSMC and partners serving Nvidia, AMD and Apple. The US and Europe are building their own capacity, but Taiwan's packaging and substrate ecosystem cannot be replaced quickly.

Supply tightness feeds into AI prices. Leave room in budgets for GPU rental swings, factor rising memory prices into hardware refreshes, and compare Nvidia- and AMD-based cloud options for leverage.