Google plans two custom AI chips a year

Google plans to move from a 2-year custom-chip generation cycle to 2 chips a year, AI infrastructure chief Amin Vahdat said at Semicon Taiwan, according to Nikkei Asia’s report. The faster cadence signals that custom silicon is becoming a workload-specific refresh loop, not a one-time escape from GPUs; it also raises qualification and software-portability costs for cloud buyers. Teams watching Meta’s six-month MTIA cadence should treat hardware lock-in as a scheduling risk: benchmark portability before committing a production path to any one accelerator.