I’m bullish on XPUs.
I’ve argued for a while that AI infrastructure is moving from “GPUs for everything” to the right silicon for the workload. Hyperscalers are right-sizing their infrastructure instead of buying one platform for every job. Agentic AI needs different machines for low-latency and long-context inference, and even different CPUs. When I talked with Microsoft’s Saurabh Dighe about Maia 300 on Semi Doped, he said Microsoft “will deploy a heterogeneous fleet to make sure that we cover every point on that frontier efficiency curve.”
Thus, the custom accelerator market will continue to grow, and a bigger market doesn’t have to belong only to folks like Broadcom, Marvell, and MediaTek. Merchant silicon vendors have good reasons to want in this game, and each brings something different. Qualcomm has Alphawave’s SerDes and design team, its own CPU cores and HBC, plus Chris Lattner’s Modular. Intel has x86, advanced packaging and its own fabs. AMD has CPU and GPU chiplets it can tailor for a customer, not to mention Taalas.
Can Qualcomm, Intel, and AMD compete? How do they stack up?
Let’s start with Qualcomm. Could it truly offer custom ASICs competitive with Broadcom or Marvell?