Google is accelerating the schedule for its custom AI processors, moving away from a roughly two-year refresh cycle toward a faster rollout that includes introducing two chips a year.
Nikkei Asia reported the shift came during a keynote by Google senior vice president Amin Vahdat at Semicon Taiwan and was framed as necessary to "stay ahead in the artificial intelligence race".
Asian expansion
The company also plans to expand its Taiwan research-and-development footprint by about 60% expand local space 60%, a local capacity bet tied to faster design and deployment timelines.
The move continues a multi‑year push at Google to own more of the AI stack: recent TPU generations split training and inference silicon and aimed to improve price‑performance for large models, and Google runs both its custom accelerators and Nvidia systems in cloud offerings.
Hyperscalers
The faster cadence echoes a wider hyperscaler pattern of building proprietary AI silicon to control costs and capacity while still relying on merchant GPUs for some workloads.
Vahdat told the audience the company expects the cadence to increase beyond two chips a year as demand for agentic, token‑heavy AI workloads grows eventually even more.
