Google Unveils 8th-Gen TPUs and Gemini Enterprise

Google used its Cloud Next 2026 keynote on Wednesday to unveil its most aggressive artificial intelligence infrastructure push yet, splitting its next-generation Tensor Processing Unit line into separate chips for training and inference, announcing general availability of the seventh-generation Ironwood TPU, and introducing Gemini Enterprise as an end-to-end agentic platform.

Ironwood Goes GA

The Ironwood TPU, the seventh generation of Google's custom silicon, is now generally available. Each chip delivers 4.6 petaFLOPS of compute and 192 GB of HBM3e memory, and a 9,216-chip superpod reaches 42.5 exaFLOPS, according to details confirmed by TechCrunch, Google's official blog and TheNextWeb.

Training and Inference Split

For the first time, Google split its eighth-generation TPU architecture into two purpose-built chips. The TPU 8t, aimed at training, scales up to 9,600 TPUs with two petabytes of shared high-bandwidth memory in a single superpod, delivering up to three times the processing power of Ironwood and up to two times better performance per watt. The TPU 8i, aimed at inference, introduces a new Boardfly topology that connects 1,152 TPUs, triples on-chip SRAM and delivers 80 percent better performance per dollar versus the prior generation.

TechCrunch cross-verified the top-line claims, citing "up to 3x faster AI model training, 80% better performance per dollar, and the ability to get 1M+ TPUs to work together."

Gemini Enterprise and Agent Tools

Google announced Gemini Enterprise as an "end-to-end system for the agentic era," along with new agent-builder tools that Bloomberg framed as a direct challenge to OpenAI and Anthropic. Strategically, Google positioned TPUs as complementary to Nvidia silicon, confirming it will also offer Nvidia's Vera Rubin platform later in 2026.

The dual-track strategy signals that inference economics, not just training horsepower, has become the central battleground of the AI infrastructure race.