The training-focused TPU 8t delivers nearly 3x the compute performance per pod compared to its predecessor, with a single superpod scaling to 9,600 chips and delivering 121 ExaFlops of compute capacity. The architecture also offers 2.8x better price-to-performance, according to Google.
The TPU 8i takes a different approach, optimizing for inference workloads with 3x more on-chip SRAM than previous generations—384 MB of on-chip SRAM paired with 288 GB of high-bandwidth memory. The chip delivers up to 80% better performance per dollar and 2x the performance per watt, the company claimed.
Both chips leverage Google's new Boardfly architecture, which achieves up to a 50% improvement in latency for communication-intensive workloads by reducing network diameter, the technical documentation shows.
The hardware announcement follows Google's expanded partnership with Anthropic earlier this month, which will provide the AI startup with multiple gigawatts of next-generation TPU capacity. The deal highlights how Google is leveraging its custom silicon to attract major AI companies seeking alternatives to Nvidia's GPUs in the increasingly competitive infrastructure market.
Google CEO Sundar Pichai positioned the chips as purpose-built for AI agents, stating they deliver the massive throughput and low latency needed to concurrently run millions of agents cost-effectively. The company has already secured adoption from Citadel Securities, with the financial services firm choosing TPUs to power their AI workloads.
The dual-chip strategy reflects the diverging computational needs of modern AI systems: massive parallel processing for training frontier models versus rapid, memory-intensive operations for deploying those models as interactive agents.
