Custom Silicon Development Accelerates Among Hyperscalers
Amazon, Google, and Microsoft have all expanded internal custom AI accelerator programs, developing chips like Trainium, TPU, and Maia specifically to reduce dependence on merchant GPU suppliers for a portion of their internal training and inference workloads. Each custom silicon program requires enormous upfront design and fabrication investment, but hyperscalers with sufficient internal workload volume can justify that cost through improved price-performance on workloads their own software stack controls completely. Broadcom and Marvell have both expanded custom silicon design services supporting these hyperscaler programs directly. The shift is creating a genuine second procurement channel alongside merchant GPU purchases.
Market Impact: Adds 60% inference compute share








