Google Takes Purpose at Nvidia With New Tensor Chips to Energy AI Growth – Decrypt

Google unveiled two AI processors at its Cloud Subsequent 2026 convention in Las Vegas on Wednesday, marking the corporate’s eighth era of customized silicon designed to problem Nvidia’s AI chip dominance.

The training-focused TPU 8t delivers almost 3x the compute efficiency per pod in comparison with its predecessor, with a single superpod scaling to 9,600 chips and delivering 121 ExaFlops of compute capability. The structure additionally presents 2.8x higher price-to-performance, in response to Google.

The TPU 8i takes a unique method, optimizing for inference workloads with 3x extra on-chip SRAM than earlier generations—384 MB of on-chip SRAM paired with 288 GB of high-bandwidth reminiscence. The chip delivers as much as 80% higher efficiency per greenback and 2x the efficiency per watt, the corporate claimed.

Each chips leverage Google’s new Boardfly structure, which achieves as much as a 50% enchancment in latency for communication-intensive workloads by decreasing community diameter, the technical documentation exhibits.

The {hardware} announcement follows Google’s expanded partnership with Anthropic earlier this month, which is able to present the AI startup with a number of gigawatts of next-generation TPU capability. The deal highlights how Google is leveraging its customized silicon to draw main AI firms looking for alternate options to Nvidia’s GPUs within the more and more aggressive infrastructure market.

Google CEO Sundar Pichai positioned the chips as purpose-built for AI brokers, stating they ship the huge throughput and low latency wanted to concurrently run tens of millions of brokers cost-effectively. The corporate has already secured adoption from Citadel Securities, with the monetary companies agency selecting TPUs to energy their AI workloads.

The twin-chip technique displays the diverging computational wants of recent AI methods: huge parallel processing for coaching frontier fashions versus fast, memory-intensive operations for deploying these fashions as interactive brokers.

Pichai stated Wednesday that Google is on monitor to spend as much as $185 billion this 12 months alone to energy AI infrastructure for the “agentic period,” with the agency already producing almost 75% of its new code with AI below the watchful eye of engineers.