AMD stretches supplier planning out to five years to keep pace with AI chip demand

Of everything Su said in Taipei, the longer planning window is the part with a cost attached. AMD made that plan with the companies she saw on this trip, her third to Taiwan this year: Acer, Asus, Foxconn and Quanta, then TSMC in Hsinchu ^[5]. Getting those partners to expand capacity for CPUs, GPUs and other AI products was a central purpose of the visit, she said ^[15]. The longer horizon is meant to bring that capacity online in a coordinated way ^[7]. Her case on demand, as Reuters reported it, was plainer. "There's very, very high demand for the next several years," she said ^[2].

The May pledge of more than $10 billion targeted advanced packaging and manufacturing partnerships for AMD's next AI infrastructure ^[11]. Data center revenue was $6.7 billion in the second quarter, up 107% on the year ^[12]. Divide by 2.07 and the year-earlier quarter comes to about $3.2 billion ^[16]. At the current quarterly rate, the original pledge equals about a quarter and a half of data center sales, 10 divided by 6.7 ^[17].

The facts fit three outcomes. In the first, demand stays ahead of output for years and AMD sells whatever its partners can build; Su's own words on 2027 were that AMD "can definitely use more" ^[1]. In the second, the limit is memory. Su said memory supply is constrained across the market, both for high-bandwidth chips in AI servers and for conventional memory in CPUs and PCs, and she went on to South Korea, home to Samsung and SK Hynix, for talks with memory makers ^[9]. Money spent in Taiwan does not add Korean memory. In the third, demand cools before the window closes.

I think the longer window does support the view that AMD expects supply to stay tight beyond this cycle. A company that expected the shortage to end in 2027 would have little reason to plan capacity with its suppliers five years out. The counter-case comes from the same evidence. Every demand statement here is Su's, made while she was asking partners for more capacity, and a seller in that position has reason to describe demand at its strongest. The size of the revised pledge will help separate the two. A large increase means AMD is funding a good part of its own forecast; a small one leaves more of the risk with the partners. The view is wrong if data center growth falls well short of the 107% pace while partners are still adding capacity under the longer plan ^[12].