Google Unveils 8th Gen TPUs, Splits Training and Inference
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Google Cloud introduced the TPU 8t (training) and TPU 8i (inference) at Cloud Next '26, with both chips slated for general availability later in 2026 — a workload-specific split from previous unified TPU generations.
- Sundar Pichai and Google VP Amin Vahdat framed the chips as purpose-built for the "agentic era" of AI, positioning the new TPUs as the compute substrate for autonomous AI agents rather than just model training.
- MarketWatch, CNBC, and the Wall Street Journal all cast the launch as Google's latest direct challenge to Nvidia's dominance in AI accelerators, with the dual-chip approach mirroring how Nvidia segments H100/B200-class products.
- Google announced the Virgo Network megascale data center fabric and Axion CPUs alongside the TPUs, pressing a full-stack AI infrastructure play that extends well beyond the chip itself.
- Inc.com reported the launch is paired with a $750 million plan to move AI from experiments into real-world deployment, signaling Google's intent to commercialize agentic AI workflows at scale.
- Blockonomi noted the eighth-generation TPUs were co-developed with Broadcom, Google's longstanding TPU manufacturing partner, underscoring that the Nvidia challenge remains an asymmetric foundry-and-design effort rather than a fully in-house chip play.
Why it matters: By splitting training and inference into separate TPU SKUs and tying them to a $750 million push into production agentic AI, Google is no longer pitching its custom silicon as a general-purpose Nvidia alternative — it's carving out a workload-specific niche, the same segmentation strategy that has let Nvidia defend premium pricing. Customers building agentic pipelines now have a vertically integrated Google stack (TPU + Axion + Virgo fabric) to anchor procurement decisions against Nvidia-powered systems.
Ask SkimNews
