UBS: Broadcom (AVGO.US) — Thoughts on Google's TPU v8 Generation
Summaries are public. PDF access requires an active membership; all members have the same access. Sign in
UBS published its initial analysis of Google's new TPU v8 platform, unveiled at Cloud Next '26. The platform introduces two distinct chips: the v8t (a single-chip training accelerator supplied by Marvell) and the v8i (a dual-chip inference accelerator supplied by Broadcom). Designed specifically for latency-sensitive inference and emerging agentic workloads rather than training, the v8i aligns with UBS's view that Broadcom's TPU shipments in 2027 will significantly outpace Marvell's.
The v8i features 288GB of HBM and 384MB of SRAM—roughly triple its predecessor—enabling larger active model sets to reside on-chip. Google estimates an 80% improvement in performance per dollar. Furthermore, Google expanded the maximum inference pod scale from 256 TPUs to 1,152 TPUs, connecting four chips across eight boards in 36 groups.
UBS highlights that the architectural pivot toward inference is more bullish for Broadcom, as inference workloads scale more broadly and persistently than training. While the v8i doubles the physical CPU host count per server rack, benefiting ARM-based CPUs, UBS notes this dynamic is likely already reflected in ARM's long-term royalty framework. Additionally, the retention of copper in the scaling domain favors Semtech. UBS reiterates a Buy rating on Broadcom with an unchanged target price of $475, applying roughly 12x EV/FCF to CY27E software FCF of $42 billion and 26x to semiconductor FCF of $73 billion.