Google Launches 8th-Gen TPUs for the Agentic AI Era

Google's eighth-generation TPUs introduce specialized chips for training (8t) and inference (8i) to support the next era of AI agents. These chips offer significant improvements in compute throughput, memory bandwidth, and energy efficiency through custom Axion CPUs and liquid cooling. Available later this year, they are designed to handle the complex reasoning and massive scale required by modern foundation models.

Key Points

• Google introduced two distinct 8th-gen TPU architectures: the 8t for training and the 8i for inference and reasoning.
• The TPU 8t scales to 9,600 chips per superpod, delivering 121 ExaFlops of compute to accelerate frontier model development.
• The TPU 8i features 3x more on-chip SRAM and 80% better performance-per-dollar to handle complex, iterative AI agent workflows.
• Both chips integrate Google's custom Axion ARM-based CPUs and advanced liquid cooling for 2x better performance-per-watt.
• The hardware is co-designed with software frameworks like JAX and PyTorch to support trillion-parameter models and autonomous agents.

Sentiment

Mixed but cautiously optimistic about Google's hardware position. The community broadly acknowledges the impressive TPU specifications and Google's structural advantages in owning the full AI stack, but remains skeptical about their ability to translate hardware superiority into competitive products. Gemini's weaknesses in agentic coding and developer tooling are a recurring frustration, even among commenters who appreciate Google's long-term positioning.

In Agreement

• Google's vertical integration from custom silicon to software gives them a massive structural cost advantage over Nvidia-dependent competitors
• The TPU specs are genuinely impressive — 121 ExaFLOPS per pod with 2x performance-per-watt over the previous generation
• Having separate training and inference chips demonstrates Google's deep understanding of workload optimization that other providers lack
• Google is well-positioned for the long game even if their models aren't currently the best, since they own the entire stack and AI isn't existential to their business
• Google's early investment in TPUs and the transformer architecture was prescient and gives them a durable foundation

Opposed

• Gemini models are significantly weaker at agentic tasks, tool use, and coding compared to Claude and GPT, with commenters reporting death loops, broken tool calls, and unreliable behavior
• Google's developer tooling (Gemini CLI, VS Code extension, Antigravity) is far behind competitors in usability and reliability
• Good hardware is worthless without good software — 'bad software kills good hardware' — and Google has consistently failed to translate its hardware lead into superior products
• Google's aggressive model deprecation policy (one-year EOL) and strict rate limiting create an unreliable platform for building production systems
• Google's management under Sundar is uninspiring, and ChatGPT forced them into the race rather than Google seizing the opportunity on their own