# Google ships four TPU variants in a year as it steps up Nvidia competition

> Google is shipping four TPU variants in a year as it steps up competition with Nvidia in AI computing.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Chips & Infrastructure
- Published: 2026-11-15T20:29:01+00:00
- Updated: 2026-11-15T21:10:00.773694+00:00
- Web page: https://ai-news.ikangai.com/story/google-ships-four-tpu-variants-in-a-year-as-it-steps-up-nvidia-competition/
- Seen in: THE DECODER
- Change: capability (New capability)
- Applies to: TPU 8tx, TPU 8i Max, TorchTPU

## What changed

Google has added the TPU 8tx for confidential computing and the TPU 8i Max for memory-intensive agent work, with TorchTPU released as open source.

Google says it has shipped four TPU systems within a single year for the first time. The two newest variants target confidential data and memory-hungry agent workloads: the TPU 8tx handles confidential computing, and the TPU 8i Max targets memory-intensive work. The Decoder's report on the launch framed the expansion as Google stepping up competition with Nvidia in AI computing.

Neither variant is a new chip generation. Both are built on existing silicon. What changes is the packaging, the networking and the memory configuration around the same processor. Google presents this modular approach as a deliberate design choice that lets it deliver new hardware variants in months rather than years.

The TPU 8tx addresses a gap in how data is normally protected. Data is encrypted in storage and in transit, but a processor has to decrypt it in memory before it can work on it. Google says the 8tx keeps data and model weights protected and isolated during that active use. The report notes that Nvidia already offers a comparable feature for its GPUs through its confidential computing architecture. Google says enterprises could use one system to fine-tune models, distill large models into smaller ones and train agents with reinforcement learning without moving sensitive data between clusters. The system uses the same low-latency Boardfly network Google developed for real-time inference. It is supposed to deliver up to twice the latency reduction compared with the TPU 8t.

The TPU 8i Max takes on memory bandwidth, which the report describes as the chokepoint when many agents work at once across large codebases or hours of video. Google's answer is to put two TPU 8i chips on one package. The company says this doubles compute, memory and memory bandwidth within the same rack format, which it presents as more performance per square meter of data center space.

Google is also working to lower the barrier for developers. It has released TorchTPU as open source, so PyTorch code should run on TPUs with minimal changes, and vLLM and SGLang are supported. The report says the expansion reflects a shift in how Google treats its TPUs, from an internal tool to a platform it markets to outside customers. The company now counts nearly 100 external TPU customers.

The report ties this push to a broader change in the market. As AI applications spread, the focus is shifting from large training runs to inference, the day-to-day running of models. The report argues that specialized chips play to their strengths there. It notes that Amazon, Microsoft and OpenAI are also building their own accelerators. Nvidia has responded, according to the report, by strengthening its own inference offering through an acqui-hire and licensing deal with Groq.

The report gives no absolute performance figures for the 8tx or the 8i Max beyond the relative gains described above, and it does not say how the two variants are priced. Google's own launch material for the eighth-generation chips, which describes the TPU 8t and TPU 8i, presents its performance claims as company figures rather than independent benchmarks.

## Why it matters

More TPU options give buyers another alternative to Nvidia hardware, which can affect pricing and capacity planning.

## What you can do

Check whether your AI workloads can run on TPUs and include them in capacity comparisons.

## Original source

- [Two chips for the agentic era](https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/)

## Further reporting

- [Google ships four TPU variants in one year as it ramps up competition with Nvidia](https://the-decoder.com/google-ships-four-tpu-variants-in-one-year-as-it-ramps-up-competition-with-nvidia/)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Google ships four TPU variants in a year as it steps up Nvidia competition", 2026-11-15, https://ai-news.ikangai.com/story/google-ships-four-tpu-variants-in-a-year-as-it-steps-up-nvidia-competition/
