Ben Bajarin and Jay Goldberg break down NVIDIA’s earnings, examining how supply constraints shape growth and rising component costs pressure margins. They discuss NVIDIA’s financing commitments to neoclouds and the risks of using its balance sheet to support the ecosystem. Marvell’s XPU-attach strategy offers another route into custom silicon, with questions around when its Google opportunity will convert into revenue.

From Hot Chips, Ben shares why competing inference architectures suggest the industry is still working out how to serve AI at scale. The conversation explores the software and system-design challenges facing accelerator startups, along with custom HBM base dies and their implications for memory costs and supplier flexibility. Waymo’s workload-specific silicon raises another question: how much of inference will converge on a common architecture?