Google began exploring the idea of specialized artificial intelligence chips as early as 2006, but the real push came in 2013. At the time, the company realized that if it wanted to scale AI broadly, it would have to double the number of its data centers. It therefore began developing TPUs, which entered production in 2016. Unlike Amazon, which was focused on optimizing general-purpose computing at the time, Google targeted AI directly.
Deals with Anthropic and Others
TPUs have proven to be cutting-edge systems for training and running inference on AI models, comparable to Nvidia. For example, Gemini 3, one of the best models in the world, was trained entirely on TPUs. Google now also sells TPUs to external customers such as Anthropic, Meta, Safe Superintelligence (SSI), and xAI, threatening Nvidia's position.
Anthropic signed a massive deal with Google for more than 1 gigawatt of TPU capacity. This includes 400,000 version 7 Ironwood TPUs, which Anthropic is purchasing directly from Broadcom for about $10 billion (approximately CZK 230 billion). Anthropic is leasing the remaining 600,000 units through Google Cloud for an estimated $42 billion (about CZK 966 billion).
This deal helps Anthropic diversify away from Nvidia. Google gave Anthropic special terms, including non-voting investments and a 15% ownership cap. Anthropic trained models such as Sonnet and Opus 4.5 on TPUs, allowing it to reduce API prices by 67% and improve token efficiency. Opus 4.5 requires 76% fewer tokens than Sonnet to achieve the same performance.
Other companies such as Meta, SSI, xAI, and OpenAI are interested in TPUs because they see an opportunity to reduce costs. By merely threatening to switch to TPUs, OpenAI has already saved about 30% on its Nvidia GPU fleet without actually deploying them.
TPUv7 Performance and Cost Compared with Nvidia
TPUv7 Ironwood delivers performance similar to the Nvidia GB200, but at a lower cost. Each TPUv7 chip offers nearly the same theoretical FLOPs as the GB200, with 192 GB of 8-Hi HBM3E memory and bandwidth close to Nvidia's. Unlike Nvidia, which inflates theoretical figures, TPU specifications report more realistic values.
From Google's perspective, the total cost of ownership (TCO) for TPUv7 is 44% lower than for GB200. For external customers such as Anthropic, this means savings of up to 30% compared with GB200 and 41% compared with GB300. This is due to more efficient utilization—TPUs achieve higher model FLOPs utilization (MFU) of up to 40%, reducing the cost per effective FLOP by 52% compared with GB300.
TPUv7 is manufactured using the N3 process and features larger systolic arrays (256x256 versus 128x128 in the previous generation) for better computational performance. It has less memory than GB300 (192 GB vs. 288 GB), but in practice, it compensates with greater efficiency, especially during inference, where memory bandwidth is crucial.
TPUv7 System Architecture
TPUv7 is designed for massive scaling. The basic unit is a rack with 64 TPU chips connected in a 4x4x4 3D torus network. Each chip connects to six neighbors through the Inter-Chip Interconnect (ICI), enabling a superpod of up to 9,216 chips.
The rack contains 16 TPU trays, each with 4 chips, as well as host CPU trays. Liquid cooling with regulated valves is used for greater efficiency—the flow is adjusted according to the workload. Connections within the rack use copper cables, while external connections use optical transceivers and optical circuit switches (OCS).
This design is simpler than the Nvidia Oberon NVL72, with no complex backplanes. Google uses vertical power delivery and battery backups for reliability. All of this enables high availability and fewer outages, which is crucial for large AI models.
TPU software is not as easy to use as Nvidia's CUDA, but this is not a problem for large companies such as Anthropic. They have Google experts who optimize kernels for high efficiency. Google is now opening up the ecosystem, including the XLA compiler, to attract more customers.
This threatens Nvidia, which responded by insisting that it remains ahead. But with models such as Gemini 3, which excels at tasks such as Vending Bench (a business simulation), TPUs are proving their strength. Using TPUs, Anthropic achieved a new SWE-Bench coding record.
Google is thus reshaping the AI hardware market, offering a cheaper alternative and encouraging diversification.
Source: newsletter.semianalysis.com
