Our eighth generation TPUs: two chips for the agentic era
blog.google
April 22, 2026
7 min read
🔥🔥🔥🔥🔥
61/100
Summary
Google Cloud has introduced the eighth generation of Tensor Processing Units (TPUs), featuring two distinct architectures: TPU 8t for training and TPU 8i for inference. These chips are designed to enhance custom-built supercomputers, facilitating advanced model training, agent development, and large-scale inference workloads.
Key Takeaways
- Google introduced its eighth generation Tensor Processor Units (TPUs), featuring two architectures: TPU 8t for training and TPU 8i for inference, designed to handle demanding AI workloads.
- TPU 8t delivers nearly 3x the compute performance per pod compared to the previous generation, supporting massive scale with up to 9,600 chips and 121 ExaFlops of compute.
- TPU 8i is optimized for latency-sensitive inference workloads, enhancing efficiency in interactions between AI agents at scale.
- Both TPU 8t and TPU 8i are specialized to unlock significant efficiencies, while still capable of running various workloads.
Community Sentiment
Positives
- The TPU 8t and TPU 8i deliver up to two times better performance-per-watt over the previous generation, indicating significant advancements in energy efficiency for AI workloads.
- Google's vertically integrated approach in AI hardware seems to be paying off, allowing them to capture consumer market share without major infrastructure issues.
- Gemini models are proving competitive with other open models, especially in smaller sizes, which could enhance AI applications on mobile devices.
- Gemini's ability to use drastically fewer tokens than competitors suggests a more efficient processing model, potentially leading to cost savings in AI operations.
Concerns
- Despite their efficiency, Gemini models struggle with reasoning and execution, often producing broken tool calls and failing at complex 'agentic' tasks.
- Some users feel that Google is lagging behind competitors like Claude and ChatGPT, raising concerns about the relevance of their models in a rapidly evolving AI landscape.
- The Gemini CLI experience is reportedly frustrating, with issues like getting stuck in loops, which undermines the potential of Google's advanced hardware.
