Google's TPU and internal AI coding

Google is now using artificial intelligence to generate or approve 75% of its new code as it rolls out a new generation of custom chips built for training and running AI. (blog.google) Chief Executive Sundar Pichai disclosed the coding figure in a post tied to Google Cloud Next 2026, published April 22. In the same update, he said Google was introducing a new Gemini Enterprise Agent Platform for companies trying to manage “thousands” of AI agents. (blog.google) At the conference, Google also unveiled two eighth-generation Tensor Processing Units, or TPUs: TPU 8t for model training and TPU 8i for inference, the stage where a trained model answers prompts and takes actions. Google said both systems are scheduled for general availability later in 2026. (blog.google) A TPU is Google’s in-house AI chip, built to do the matrix math that powers large models faster and with less energy than general-purpose processors. Google said the split design reflects a change in AI workloads, with pre-training, post-training and real-time serving now demanding different hardware. (cloud.google.com) Google said TPU 8t is aimed at massive pre-training runs and can scale to a 9,600-chip superpod using a 3D torus network. TPU 8i is tuned for low-latency inference and reinforcement learning, the feedback-driven training method used to improve model behavior after initial training. (cloud.google.com) The product push comes as Google says customer demand is rising fast. Pichai wrote that Google’s first-party models are now processing more than 16 billion tokens per minute through direct application programming interface use, up from 10 billion last quarter, and that just over half of Google’s machine-learning compute investment in 2026 is expected to go to the Cloud business. (blog.google) Google Cloud Chief Executive Thomas Kurian said nearly 75% of Google Cloud customers are already using the company’s AI products. He also said 330 customers processed more than 1 trillion tokens over the past 12 months, while 35 processed more than 10 trillion tokens. (cloud.google.com) Google is pairing the hardware with more software meant to keep AI systems inside its cloud stack. Kurian said the Gemini Enterprise app now includes Agent Designer, an Inbox for agent activity, long-running agents, Skills and Projects, while Google also announced new storage and networking features around the same platform. (cloud.google.com) The internal coding number extends a trend Pichai has described before, but at a higher level. In October 2024, he said more than 25% of new code at Google was already being generated by AI and reviewed by engineers; the new 75% figure suggests the company has moved from assisted coding toward broader internal automation. (techcrunch.com) Google’s pitch at Next was that AI agents need both faster response times and more computing scale than earlier chatbot systems. The company’s answer was to split the chip roadmap in two and show that its own engineers are already relying on the same kind of code-generating systems it wants customers to buy. (blog.google; blog.google)