AI Daily Brief — 22 April 2026

Google Cloud Next ’26 opens in Las Vegas with two flagship announcements: the Gemini Enterprise Agent Platform — Google’s end-to-end build/scale/govern stack for autonomous agents — and the eighth-generation TPUs shipped as two distinct chips for training (TPU 8t) and inference (TPU 8i).

Top stories

• Gemini Enterprise Agent Platform. Build, scale, govern and optimise agents from a single workspace. Agents get long-term memory for multi-step processes; cryptographic identities and anomaly detection bake in security. via Google
• TPU 8t for training. New Inter-Chip Interconnect (ICI) topology scales up to 9,600 TPUs and 2 PB of shared high-bandwidth memory in a single superpod. 3× processing power vs Ironwood, 2× better perf/Watt. via Google
• TPU 8i for inference. “Boardfly” topology connects 1,152 TPUs in a pod; 3× more on-chip SRAM hosts larger KV caches entirely on-silicon; Collectives Acceleration Engine delivers 80% better perf-per-dollar for inference vs the prior generation. via The Decoder

By the numbers

• TPU 8t superpod: 9,600 chips · 2 PB shared HBM · 3× Ironwood perf · 2× perf/W
• TPU 8i pod: 1,152 chips · 80% better inference perf-per-$ vs prior gen
• Cloud Next ’26 expected attendees: 32,000+

Compiled by AI Feed’s editor from publicly reported announcements on 22 April 2026.