Google's Eighth-Gen TPUs Claim 3x Training Speed, 80% Better Value
Google announced its eighth generation of tensor processing units at Google Cloud Next on April 22, 2026. Two variants: TPU 8t for training, TPU 8i for inference. The numbers look good on paper.
The Claims
Google says TPU 8th gen delivers up to 3x faster AI model training compared to the previous generation. Performance per dollar improves by 80%.
The cluster scale claim is notable: TPU 8th gen supports configurations of over 1 million TPUs working in parallel. That is a large number. Whether workloads actually benefit from that scale is a separate question.
Still Buying Nvidia Too
Google also announced that its cloud will offer Nvidia's Vera Rubin chip later in 2026. This is not unusual for Google Cloud, which has long sold both its own silicon and Nvidia hardware.
The more interesting piece: Google and Nvidia are collaborating on networking software called Falcon to improve GPU efficiency on Google's infrastructure. Falcon is a Google-created technology, open sourced in 2023 under the Open Compute Project. Software-level efficiency improvements are harder to benchmark than chip specs, but they often matter more in practice.
What to Watch
The 80% performance-per-dollar improvement is the headline that matters for buyers. If it holds under real training runs, that is a meaningful cost reduction. Benchmark conditions and production conditions rarely match exactly.
The Falcon collaboration is worth tracking separately. If Google can squeeze more throughput from Nvidia hardware through networking software, that helps Google Cloud compete on the workloads where customers are already committed to Nvidia tooling.
Real-world numbers will surface once customers start running production workloads on the new hardware.
Source: Techcrunch