Google Cloud Next 2026: Google Announces Two New Chips For Next-Gen AI
Google introduced TPU 8t and TPU 8i chips at Cloud Next 2026, optimised for large-scale AI training and agentic reasoning workloads.
Published : April 23, 2026 at 1:18 PM IST
Hyderabad: Google introduced two specialised custom-built chips for AI workloads, including autonomous AI agents that work on behalf of users to get things done. Announced during the Google Cloud Next 2026 conference in Las Vegas, the eighth generation of Tensor Processor Units (TPUs) has been confirmed to be split into two, where one of them, TPU 8t, is purpose-built for model training, and the other one, TPU 8i, is geared for inference.
"AI agents need to reason, plan, and execute multi-step workflows. TPU 8i is designed specifically to enable AI agents to complete this very quickly to provide a good user experience. Complementing TPU 8i, TPU 8t is optimised for training and can run even the most complex models on a single, massive pool of memory," Google said.
These two chips will drive everything from cutting-edge model training and agent development to massive inference workloads. The tech giant claims that the new TPUs, along with its full-stack infrastructure, will create the engine to help develop "highly responsive agentic AI" for the masses.
Both chips will be generally available later this year and can be used as part of Google’s AI Hypercomputer.
| Feature / Capability | TPU 8t – The Training Powerhouse | TPU 8i – The Reasoning Engine |
|---|---|---|
| Primary Focus | Frontier model training at a massive scale | Complex reasoning and agentic tasks |
| Performance Gain | ~3x compute performance per pod vs previous generation | 80% better performance-per-dollar vs the previous generation |
| Scale | Superpod with 9,600 chips, 2 PB shared high-bandwidth memory, 121 ExaFlops | Designed for collaborative, iterative agent workflows |
| Bandwidth | Double interchip bandwidth vs the previous generation | Doubled ICI bandwidth to 19.2 Tb/s |
| Memory | Massive shared pool enabling complex models | 288 GB high-bandwidth memory + 384 MB on-chip SRAM (3x previous generation) |
| Storage Access | 10x faster storage access with TPUDirect | Optimised NUMA isolation with Axion Arm-based CPUs |
| Networking | Virgo Network enables near-linear scaling up to 1M chips | Boardfly architecture reduces max network diameter by >50% |
| Latency Reduction | Optical Circuit Switching (OCS) reroutes around failures seamlessly | Collectives Acceleration Engine (CAE) reduces on-chip latency up to 5x |
| Reliability & Utilization | Targets >97% goodput with RAS features (telemetry, fault detection, rerouting) | Eliminates “waiting room” effect for agents |
| Use Case | Accelerating frontier model development cycles from months to weeks | Serving nearly twice the customer volume at the same cost |
Google says that every major computing transition has required infrastructure breakthroughs, and the agentic era is no different. The company emphasises that infrastructure must evolve to meet the demands of autonomous agents operating in continuous loops of reasoning, planning, execution, and learning. It has presented TPU 8t and TPU 8i as the answer to this challenge.
In short, TPU 8t is optimised for massive-scale training with extreme throughput and reliability, while TPU 8i is tuned for reasoning workloads to minimise lag and enable efficient agent collaboration.
