Google Takes Aim at Nvidia With New Tensor Chips to Power AI Boom

3 weeks ago 11

In brief

Google introduced eighth-generation Tensor Processing Units with 2 architectures: TPU 8t for grooming and TPU 8i for inference.
TPU 8t delivers astir 3x the compute show per pod implicit erstwhile generations, scaling to 121 ExaFlops.
TPU 8i features 3x much on-chip representation to grip the iterative demands of AI agents.

Google unveiled two AI processors astatine its Cloud Next 2026 league successful Las Vegas connected Wednesday, marking the company's eighth procreation of customized silicon designed to situation Nvidia's AI spot dominance.

The training-focused TPU 8t delivers astir 3x the compute show per pod compared to its predecessor, with a azygous superpod scaling to 9,600 chips and delivering 121 ExaFlops of compute capacity. The architecture besides offers 2.8x amended price-to-performance, according to Google.

The TPU 8i takes a antithetic approach, optimizing for inference workloads with 3x much on-chip SRAM than erstwhile generations—384 MB of on-chip SRAM paired with 288 GB of high-bandwidth memory. The spot delivers up to 80% amended show per dollar and 2x the show per watt, the institution claimed.

Both chips leverage Google's caller Boardfly architecture, which achieves up to a 50% betterment successful latency for communication-intensive workloads by reducing web diameter, the method documentation shows.

The hardware announcement follows Google's expanded concern with Anthropic earlier this month, which volition supply the AI startup with aggregate gigawatts of next-generation TPU capacity. The woody highlights however Google is leveraging its customized silicon to pull large AI companies seeking alternatives to Nvidia's GPUs successful the progressively competitory infrastructure market.

Google CEO Sundar Pichai positioned the chips arsenic purpose-built for AI agents, stating they present the monolithic throughput and debased latency needed to concurrently tally millions of agents cost-effectively. The institution has already secured adoption from Citadel Securities, with the fiscal services steadfast choosing TPUs to powerfulness their AI workloads.

The dual-chip strategy reflects the diverging computational needs of modern AI systems: monolithic parallel processing for grooming frontier models versus rapid, memory-intensive operations for deploying those models arsenic interactive agents.

Pichai said Wednesday that Google is connected way to spend up to $185 cardinal this twelvemonth alone to powerfulness AI infrastructure for the “agentic era,” with the steadfast already generating astir 75% of its caller codification with AI nether the watchful oculus of engineers.