At its Cloud Next 2026 conference held in Las Vegas, Google announced two new AI processors, marking the eighth iteration of their custom silicon aimed at challenging Nvidia’s stronghold on AI chip technology.
The TPU 8t is designed for training applications and boasts a compute performance per pod that is nearly three times higher than its predecessor. With an impressive scalability to 9,600 chips in a single superpod, it offers 121 ExaFlops of compute capacity. According to Google, this architecture provides a 2.8x improvement in price-to-performance ratio.
In contrast, the TPU 8i is optimized for inference tasks, featuring 3 times more on-chip SRAM than previous models—384 MB paired with 288 GB of high-bandwidth memory. The company claims it achieves up to an 80% better performance per dollar and double the performance per watt.
Both chips utilize Google’s innovative Boardfly architecture, which reduces network diameter to achieve up to a 50% reduction in latency for communication-heavy workloads, as detailed in their technical documentation.
Following the hardware announcement, Google expanded its partnership with Anthropic earlier this month, providing the AI startup with significant TPU capacity. This agreement underscores Google’s strategy of using its custom silicon to draw major AI firms seeking alternatives to Nvidia’s GPUs in a competitive infrastructure market.
Sundar Pichai, CEO of Google, highlighted that these chips are specifically designed for AI agents, offering the necessary throughput and low latency to efficiently manage millions of agents. Citadel Securities has already adopted TPUs for their AI workloads.
Pichai stated on Wednesday that Google plans to allocate up to $185 billion this year solely for AI infrastructure in anticipation of an “agentic era,” with the company generating nearly 75% of its new code through AI, under the guidance of engineers.