LOADING

Google Has Started Selling Full TPU Systems, and the Eighth Generation Comes in Four Configs

News

This year Google began delivering full TPU systems to selected customers, instead of only renting TPU capacity on Google Cloud. The eighth generation is split into four configs for training, real-time inference, confidential computing, and very long context. It is the first year Google has launched four TPU systems. In its 10-Q, Alphabet lists “product sales, primarily TPU systems” inside Google Cloud revenue, and calls those sales an important contributor to cloud growth in the third quarter.

From cloud rental to a full rack at the customer site

TPU started in 2013 as a custom chip to cut the cost of Google’s own AI compute. From 2018 it was rented on Google Cloud, with Google still running the halls. On the first-quarter earnings call, CEO Sundar Pichai said demand was strong from AI labs, capital-markets firms, and high-performance computing, and that selected customers could deploy TPUs in their own data centers. Deliveries of complete systems began in the second quarter. A system includes compute, storage, networking hardware and software, plus training and inference services.

People close to Google say the external sale was decided after an internal argument, pushed by an executive who had been at the company for less than three years, with the aim of selling systems “the way Nvidia does.” The report frames the motive as using outside orders to share data-center capital spending. On the second-quarter call, Pichai still said Google’s own compute need for frontier AGI is a line it will not cross. The other side of the window is customers who cannot get high-end Nvidia GPUs. The rising cost of Nvidia compute is covered in the next round of compute financing.

Google Cloud revenue accelerated through the year: up 63% year over year in the first quarter, 82% in the second, and $30.7 billion in the third, up 103%. The report says third-quarter operating profit rose from $3.594 billion to $11.121 billion. Management expects TPU system sales to keep growing in the fourth quarter. Most of the related contract revenue does not hit the books until 2027.

Four configs, and TorchTPU

Ironwood, the seventh generation from 2025, was built for inference. This year’s eighth generation separates training, real-time inference, confidential computing, and very long context. Two chips announced this week cover confidential inference and long context. With the two from the first half, that is four configs. Amin Vahdat, who leads Google’s chip business, has said “the future belongs to specialization.”

Software is the other gap. TPUs were tied to TensorFlow and JAX for a long time, far from the PyTorch path built around CUDA. In mid-2025 Google started TorchTPU with Meta’s PyTorch team, so a PyTorch model could run on TPU with small code changes. That work was open-sourced at the end of October 2026.

Tokens per dollar are still moving

On 5 November, SemiAnalysis published a set of agent-inference tests. On some runs of models such as Muse Spark, seventh-generation Ironwood produced more tokens per dollar than Nvidia GPUs. Google’s AgentX submission showed TPUv7 ahead of the GB300 NVL72 in FP4 on performance per dollar. Nvidia then ran four rounds of optimization. On one model, the GB300’s tokens per dollar moved back ahead of the TPU, by 1.5× to 2.6×. That is one test swinging back, not a verdict on the whole product line.

FAQ

Q: Is Google selling TPUs to customers, or only renting them in the cloud? A: From 2018 the main offer was rental on Google Cloud. In the first quarter of this year Google said selected customers could deploy TPUs in their own data centers. Complete systems, with compute, storage, networking, and software, started shipping in the second quarter.

Q: How many configs does the eighth-generation TPU have? A: Four: training, real-time inference, confidential computing, and very long context. The two chips announced this week target confidential inference and long context.

Q: Did SemiAnalysis find the TPU cheaper per dollar than the GB300? A: Google’s AgentX result showed TPUv7 ahead of the GB300 NVL72 in FP4 on performance per dollar. After four Nvidia optimization rounds, the GB300 produced 1.5× to 2.6× more tokens per dollar on one model.

© 版权声明

相关文章