TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Google added two specialized systems to its eighth-generation TPU lineup on November 9: TPU 8tx for confidential computing and TPU 8i Max for workloads with very long inputs. Google says both will be available in 2027, but has not published prices or independent benchmarks.
Google announced two new eighth-generation tensor processing unit systems on November 9: TPU 8tx, designed for confidential computing, and TPU 8i Max, designed for workloads with very long inputs. The additions give cloud customers more options for running AI tasks on Google’s own chips, though neither system is available yet; Google expects both in 2027.
The new systems build on the TPU 8t and TPU 8i platforms that Google introduced in April. TPU 8t is intended for training large AI models, while TPU 8i is designed to serve models with low latency. Google says the 8tx system targets regulated businesses and government users that need to protect data while it is being processed. It says 8tx can deliver up to two times lower latency than TPU 8t with a near-zero performance penalty. These are Google’s figures; no independent testing of the new system has been published.
Google describes TPU 8i Max as pairing two low-latency TPU 8i chips in one package. The company says the system doubles the compute, on-chip SRAM and high-bandwidth memory bandwidth of TPU 8i while using twice the silicon within the existing rack footprint. Google is targeting tasks such as agents that need to work across large codebases or lengthy video. Prices, deployment regions and independent performance results for both new systems have not been announced.
Google chief technologist for AI and infrastructure Amin Vahdat told Bloomberg the new systems are not completely new silicon designs, but reconfigurations made for particular customer needs. Google says it has introduced four TPU systems in 2026, after previously releasing a new AI processor about every two years, according to Bloomberg. The company’s TPU product page lists all four eighth-generation systems as “Coming soon.”
Google expands its AI chips with specialized TPUs
Two new systems target confidential computing and workloads with very long inputs. Their practical value will depend on availability, pricing and real-world results.
Two needs, two designs
The additions build on TPU 8t for training and TPU 8i for low-latency inference.
TPU 8tx
Designed to protect data while it is processed, for regulated businesses and government users.
TPU 8i Max
Pairs two TPU 8i chips for tasks such as agents working across large codebases or lengthy video.
Different workloads, different trade-offs
Google’s published specifications show the base systems prioritize different capabilities.
| System | Designed for | Published emphasis |
|---|---|---|
| TPU 8t | Training large AI models | Peak compute |
| TPU 8i | Low-latency inference | Memory and bandwidth |
| TPU 8tx | Confidential computing | Protecting data in use |
| TPU 8i Max | Very long inputs | More compute and memory capacity |
What buyers still need
The announcement expands choice, but customers cannot yet assess costs or deployment fit.
More ways to access TPU capacity
Oracle plans to offer Google TPUs in its cloud starting in calendar year 2027; service terms and locations have not been provided.
Google expects the new systems in 2027.
Test performance, software fit and operating effort for your needs.
Check Google and planned Oracle cloud details when published.
More Options for Cloud AI Buyers
Google’s announcement matters to businesses weighing where to run AI workloads because it extends the company’s effort to offer an alternative to Nvidia-based computing. A second source of AI capacity could give customers more choice as demand strains supply. Alphabet described the market as a “supply-constraint environment” on its October earnings call. Google says nearly 100 customers use its TPUs, although that figure does not show how much capacity each customer uses or how many have moved major workloads to the chips.
The specialized systems also address two practical requirements that can shape infrastructure choices: protecting sensitive information during processing and providing enough memory capacity for long inputs. If Google’s claims hold up in real deployments, customers could match hardware more closely to certain tasks. But availability alone does not establish a cost or performance advantage. Businesses would need to compare results for their own workloads, including the engineering time required to port software and operate it on a different platform.
Cloud access could broaden the options beyond Google’s own service. Oracle plans to offer Google TPUs in its cloud starting in calendar year 2027, according to the source material. The planned offering could give some customers another way to access TPU capacity, though pricing, locations and technical details have not been provided.
Products Worth Considering
Samsung Galaxy Tab S7 11-inch Android Tablet 128GB Wi‑Fi with S Pen Mystic Black
Amazon Kindle Paperwhite (2024, 16 GB) – 7-inch Glare-Free Display, Black
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
From Two Base Systems to Four
Google introduced TPU 8t and TPU 8i at its Cloud Next conference on April 22, 2026. The company said TPU 8t could deliver up to a 2.7-fold improvement in performance per dollar over its previous generation, Ironwood, for large-scale training. For TPU 8i inference, Google claimed up to an 80% performance-per-dollar improvement. Those comparisons are vendor claims about the base systems, not independent measurements of the two announced in November.
Google’s published specifications show different design priorities. TPU 8t is listed with 216 GB of high-bandwidth memory, 128 MB of on-chip SRAM, 12.6 petaflops of peak FP4 compute and 6,528 GB/s of memory bandwidth. TPU 8i is listed with 288 GB of high-bandwidth memory, 384 MB of SRAM, 10.1 petaflops of peak FP4 compute and 8,601 GB/s of bandwidth. In those specifications, 8i has less peak compute but more memory and bandwidth than 8t, a configuration Google associates with inference work.
Google has also pointed to software work intended to make TPUs easier to use. The company says PyTorch code and Hugging Face models can run natively through TorchTPU, which it open-sourced in October, and it cites TPU support for the vLLM and SGLang inference engines. Google has also made Cloud Pathways, the runtime it says Google DeepMind uses behind Gemini, generally available on Google Kubernetes Engine. These developments may affect how easily customers can use the hardware, but do not establish the performance of the new systems.
““This is the first time we have introduced four TPU systems in a single year.””
Performance and Pricing Remain Open
The announced systems have not been independently tested, and Google has not released list prices or deployment regions. Its performance figures are company claims, so buyers cannot yet compare the new systems against alternatives using independent results. The source material also gives no detailed availability schedule beyond the expected 2027 release.
It remains unclear how broadly customers will adopt the specialized systems, what workloads will benefit most, and how much effort it will take to move software from existing GPU environments. Google’s customer count and Alphabet’s description of constrained supply indicate interest and capacity pressure, but do not answer those questions. The source material also does not provide deployment details that would let customers assess TPU 8tx’s confidential-computing protections against their particular requirements.
Availability Is the Next Milestone
Google has said TPU 8tx and TPU 8i Max will be available in 2027. Customers will need further announcements for prices, cloud regions and deployment details before they can assess when or where to use them. Independent benchmarks and customer results would also help test Google’s claims about latency, memory capacity and performance per dollar.
Oracle’s planned TPU offering in 2027 is another milestone to watch. The timing and terms of that service have not been specified in the source material. Until Google and Oracle publish more details, the practical choice for buyers remains limited to evaluating the published specifications, software support and their own requirements.
Disclosure: The source article says it was drafted with assistance from Anthropic’s Claude and that Anthropic is a Google TPU customer, as Alphabet said on its October earnings call. It says neither Google nor Anthropic reviewed or sponsored the article.
Key Questions
What did Google announce?
Google announced TPU 8tx, aimed at confidential inference, and TPU 8i Max, aimed at workloads with very long inputs. They join the TPU 8t training system and TPU 8i inference system announced in April.
When will the new TPU systems be available?
Google expects TPU 8tx and TPU 8i Max to be available in 2027. The company has not published a more specific date or cloud region list.
Have the new systems been independently benchmarked?
No independent benchmarks are included in the source material. The published performance claims for the new systems come from Google.
How are TPU 8tx and TPU 8i Max different?
Google says 8tx is designed to protect data and model weights during active use, while 8i Max pairs two TPU 8i chips to increase compute, SRAM and memory bandwidth for large-context workloads.
Source: Thorsten Meyer AI
