Google Cloud TPUs Aim at Nvidia With TPU 8
Google Cloud TPUs launched at Cloud Next '26 with TPU 8t/8i and a $750 million partner fund, shifting AI compute economics and cloud competing with Nvidia.
KEY TAKEAWAYS
- Google split TPU 8 into TPU 8t for training and TPU 8i for inference, lowering AI compute costs.
- Company paired the launch with a $750 million partner fund for agentic AI adoption and partner credits.
- The moves position Google Cloud to challenge Nvidia's GPU dominance in cloud AI compute.
HIGH POTENTIAL TRADES SENT DIRECTLY TO YOUR INBOX
Add your email to receive our free daily newsletter. No spam, unsubscribe anytime.
Alphabet Inc.'s Google Cloud unveiled its eighth-generation Tensor Processing Units (TPUs) at Cloud Next ’26 on April 22, 2026. The new lineup includes TPU 8t for AI model training and TPU 8i for inference, designed to reduce AI compute costs and enhance cloud competitiveness. The launch coincided with a $750 million partner fund to support agentic AI development.
TPU 8 Architecture and Strategy
Google Cloud detailed in a blog post that the TPU 8 series splits into two specialized chips: TPU 8t for training and TPU 8i for inference. These chips feature higher on-chip memory, lower latency, and cost efficiencies tailored for AI agent workloads. The company described an integrated infrastructure combining TPU, GPU, and CPU resources. It expects more than half of 2026’s machine-learning compute investment to flow to its Cloud business. Alphabet CEO Sundar Pichai wrote, "To deliver the massive throughput and low latency needed to concurrently run millions of agents cost-effectively."
Partner Ecosystem and Customer Commitments
Google Cloud said in a press release it will allocate the $750 million fund to a 120,000-member partner ecosystem focused on agentic AI, providing value assessments, prototyping, deployment support, upskilling, and sandbox credits. The company also introduced new generative AI features for Google Maps and geospatial analytics aimed at enterprise users.
Broadcom disclosed in an April 6, 2026, 8-K filing a five-year agreement through 2031 to co-design TPUs with Google, coordinating hardware development across the stack. Mira Murati’s Thinking Machines Lab signed a single-digit billions deal with Google Cloud for AI infrastructure that will use Nvidia GB300 GPUs for model training, deployment, and reinforcement-learning workflows.
Other customers and partners include Meta, which has a multibillion-dollar TPU testing arrangement; Anthropic, committed to large-scale TPU usage; and adopters such as Citadel Securities and U.S. national labs. Inside Google, about 75% of new code is AI-generated, a share rising as staff increasingly use internal AI tools.
Together, the TPU architecture, co-design agreements, and funded partner ecosystem aim to lower the cost and latency of large-scale AI deployments, positioning Google Cloud to challenge Nvidia’s dominance in GPU compute for cloud AI workloads.
