A Google TPU v4 package — the new TPU 8tx and TPU 8i Max systems build on the eighth-generation TPU line rather than a new chip design. Image: Wikimedia Commons / Jouppi et al., CC BY 4.0.
📌 KEY HIGHLIGHTS
- Announced 9 November 2026: Google unveiled TPU 8tx and TPU 8i Max on its official blog, completing a four-system eighth-generation TPU portfolio in a single year.
- TPU 8tx — confidential inference: Built on the same silicon as the training-focused TPU 8t, paired with 5th Generation AMD EPYC hosts, it keeps data and model weights protected while they are being processed — with up to 2x lower latency than TPU 8t.
- TPU 8i Max — massive context: Two TPU 8i chips on a single package, delivering double the compute, double the SRAM and double the HBM bandwidth of TPU 8i, in the same rack footprint.
- Not new chips: Both are new systems built around existing eighth-generation silicon — packaging, networking and memory reconfigured to meet customer requests in months, not years.
- No price, no date yet: Google's announcement gives neither; Google's Amin Vahdat told Bloomberg both systems are expected next year, in 2027.
What happened?
Google announced two additions to its eighth-generation Tensor Processing Unit (TPU) family on 9 November 2026, in a blog post by Amin Vahdat, its senior vice-president and chief technologist for AI infrastructure. The TPU 8tx targets secure, confidential inference; the TPU 8i Max targets massive context windows and the highest compute density per rack.
The announcement matters because of what TPUs have become. Designed more than a decade ago for Google's own voice recognition and translation work, the chips now train and serve Gemini models, run agentic workloads and power Google Cloud customers — nearly 100 of them, Google says, including frontier AI labs, trading firms, retailers and universities.
This is the second eighth-generation expansion this year. At Google Cloud Next in April 2026, Google introduced the TPU 8t for large-scale training and the TPU 8i for real-time reasoning, claiming up to 2.7 times better training performance and 80% better inference price-performance than the previous generation. The two new systems complete the set — and Google says 2026 is the first year it has introduced four TPU systems.
Google TPU 8tx TPU 8i Max explained: two systems, two jobs
Google is careful to draw a line between a new chip and a new system. Designing a new chip means laying out billions of transistors over several years; a system is how the compute, memory, networking and security are packaged around the chip. Because Google built its TPUs as modular building blocks, it can ship specialised systems in months. Both of today's announcements are systems, not chips.
| System | Built for | What is new |
|---|---|---|
| TPU 8t | Large-scale training | April 2026; up to 2.7x better training performance vs previous generation |
| TPU 8i | Real-time reasoning | April 2026; 80% better inference price-performance; cuts inter-chip delays by up to half |
| TPU 8tx | Secure, confidential inference | Same silicon as 8t, 5th Gen AMD EPYC hosts, Boardfly network; up to 2x lower latency than 8t |
| TPU 8i Max | Massive context windows, dense racks | Two 8i chips on one package; 2x compute, 2x SRAM, 2x HBM bandwidth vs 8i |
TPU 8tx is aimed at regulated enterprises and governments that run frontier or proprietary models next to their most sensitive data — banks, hospitals, public agencies. Normally, data is encrypted when stored or sent over a network, but it has to be unlocked in memory while a computer works on it. The 8tx builds security into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use. It also lets teams fine-tune models on private data, distil large models into smaller ones, and train agents with reinforcement learning on a single system, without moving sensitive data between separate clusters. Notably, the 8tx pairs the TPU with AMD CPUs — the first AMD silicon inside a TPU system, where the April systems used Google's own Axion Arm CPUs.
A TPU 3.0 board — the new systems repackage eighth-generation silicon rather than designing a new chip. Image: Wikimedia Commons / Zinskauf, CC BY-SA 4.0.
TPU 8i Max attacks a different bottleneck: the memory wall. When dozens of AI agents work together across an entire software codebase or hours of video at once, memory bandwidth — not raw compute — becomes the limit. Google put two low-latency TPU 8i chips on a single package, doubling the compute, the on-chip SRAM and the high-bandwidth memory (HBM) bandwidth of a single 8i, while fitting twice the silicon in the same rack footprint. Network World reports the package carries 384 GB of HBM, about a third more than the 8i's 288 GB. Google built the 8i Max for the million-token context windows its Gemini models support.
One honest caveat: every performance figure here is Google's own. No independent test of either system has appeared yet, and Google has published no per-chip specifications, pricing or availability date for either system.
Why it matters for Indian developers and businesses
Most Indian readers will never buy a TPU — they will rent one, indirectly, through Google Cloud or a product built on it. Confidential inference matters most to India's regulated sectors: banks, insurers, fintechs and health-tech firms that want to run AI on customer data without that data ever sitting exposed in memory. If TPU 8tx reaches Google Cloud regions serving India, it gives those teams a hardware-backed answer to a compliance question that today usually ends in "we can't put that in the cloud."
The 8i Max matters to a different Indian audience: developers building AI agents. Agents that read a whole codebase, or hours of meeting video, in one go need enormous context windows — and the memory bandwidth to feed them. Google says the 8i Max is built for exactly that workload, and it pairs the hardware with a software push aimed at ordinary developers: last month Google open-sourced TorchTPU, so PyTorch code and Hugging Face models now run natively on TPUs with no rewrite, and a new TPU agent suite in preview plugs into coding tools such as Claude Code and Cursor. Indian developers who already write PyTorch can try TPUs without learning a new stack — read our explainer on Google TorchTPU and running PyTorch natively on TPUs.
There is also a competitive angle. Google's pitch is specialisation — a purpose-built accelerator for every stage of the AI lifecycle — against Nvidia's general-purpose GPUs. Google's own Gemini models run on these chips, and the Gemini Omni 1.2 Flash model, priced for India, is the kind of workload the 8i Max is built to serve. If Google keeps shipping at this pace, the price of running AI falls — and that shows up in the tools Indian developers pay for.
Should you care?
Price and availability
Google's announcement contains no price and no availability date for either system, and no eighth-generation TPU has a published price. Google's Amin Vahdat told Bloomberg both new systems are expected to arrive next year, in 2027, and Oracle's cloud is expected to start offering TPUs the same year. Google has not said which Cloud regions will get them first, so there is no India date to report yet — we will update this post when Google announces pricing or regional availability.
FAQs
What is Google TPU 8tx? TPU 8tx is Google's third eighth-generation TPU system, announced on 9 November 2026. It is optimised for secure, confidential inference — keeping data and model weights protected in hardware while they are processed — and is aimed at regulated enterprises and governments. It uses the same silicon as the TPU 8t, paired with 5th Generation AMD EPYC host CPUs.
What is TPU 8i Max? TPU 8i Max is Google's fourth eighth-generation TPU system, also announced on 9 November 2026. It combines two TPU 8i chips on a single package, delivering double the compute, double the SRAM and double the HBM bandwidth of a single TPU 8i, for massive context windows and maximum compute density per rack.
Are TPU 8tx and TPU 8i Max new chips? No. Google says both are new systems built around existing eighth-generation silicon — the 8tx reuses the 8t chip with different hosts and networking, and the 8i Max repackages two 8i chips together. New systems can ship in months; new chips take years.
Takeaway: Google's two new TPU systems are not about raw speed — they are about the two problems holding AI back in production: keeping sensitive data safe while it is processed, and feeding agents enough memory to use million-token contexts. Both arrive in 2027, and the real test will be independent benchmarks and prices.
If you write PyTorch, the fastest way to feel this shift is to try the software side today — our Google TorchTPU explainer shows how to run PyTorch natively on TPUs without rewriting your code.
Sources
- Introducing TPU 8tx and TPU 8i Max — Google (official announcement, 9 November 2026)
- Google advances agentic AI custom silicon expanding TPU family with 8tx and 8i Max — SiliconANGLE
- Google Adds TPU 8Tx and TPU 8I Max to Eighth-Generation Lineup — Unite.AI
- Test Your PyTorch Workloads on Ironwood Before Choosing TPU 8tx — Beri.net (2027 timing, AMD SEV detail, Network World HBM figure)
Comments
Post a Comment