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 announced two additions to its eighth-generation TPU systems: one aimed at confidential computing and another at workloads with very long inputs. The variants are due in 2027, and their performance figures are Google claims; prices and independent tests have not been published.
Google announced two new eighth-generation TPU systems on November 9: TPU 8tx, designed for confidential inference, and TPU 8i Max, aimed at AI workloads with very long inputs. The additions give cloud customers more specialized alternatives to general-purpose AI accelerators, but neither system is available yet; Google’s chip chief told Bloomberg they are expected in 2027.
The new systems build on the TPU 8t and TPU 8i, which Google introduced at its Cloud Next conference on April 22. The 8t is designed for training AI models, while the 8i targets inference: running a trained model to generate answers. Google says the 8tx uses the same silicon as the 8t and adds features for secure, confidential inference. It is paired with fifth-generation AMD EPYC processors.
Google says the 8tx can provide up to twice the lower latency of the 8t, with a near-zero performance penalty. The company describes it as protecting sensitive data and model weights while they are in active use, a feature aimed at regulated businesses and government agencies. Those performance and security descriptions come from Google; no independent testing of the new system is available.
The 8i Max combines two TPU 8i chips in one package. Google says it doubles the compute, on-chip SRAM and high-bandwidth memory bandwidth of the 8i while fitting twice the silicon into the existing rack footprint. The company is pitching it for large context windows, such as AI agents processing a codebase or hours of video. Google has not provided prices or availability regions, and its TPU product page lists all four eighth-generation systems as “Coming soon.”
Why Google’s latest AI chips matter to cloud buyers
Google announced two specialized eighth-generation TPUs for confidential inference and very long inputs. Both are expected in 2027, but prices and independent test results are not yet available.
Two systems, distinct workloads
The additions extend Google’s TPU 8t training and TPU 8i inference family.
TPU 8tx
Adds confidential computing features to the 8t, aimed at businesses and agencies handling sensitive data.
TPU 8i Max
Combines two 8i chips in one package for tasks such as processing a codebase or hours of video.
What Google says changes
The new variants use existing silicon with changes to packaging and system design.
| System | Google’s stated focus | Potential fit |
|---|---|---|
| TPU 8tx | Confidential inference; claims up to twice the lower latency of 8t | Sensitive workloads |
| TPU 8i Max | Double the compute, on-chip SRAM and memory bandwidth of 8i | Very long inputs |
What buyers still need
The announcement describes intended capabilities; practical value depends on access and customer results.
Google has not published prices, regions or a detailed launch schedule.
Software changes and engineering work may be needed to switch accelerators.
Independent benchmarks for the new systems are not yet available.
More Choices for AI Cloud Buyers
For businesses buying AI computing, Google’s announcement signals a move toward specialized systems for different workloads, rather than a single accelerator configuration. Confidential computing may appeal to organizations handling sensitive information, while more memory and bandwidth could help services processing lengthy documents, code or video. Whether either system suits a particular buyer will depend on its workload, cost and measured performance.
The broader issue is access to computing capacity. Google says nearly 100 customers use its TPUs, and Alphabet described a supply-constrained environment on its October earnings call. Oracle has said it will offer Google TPUs in its own cloud starting in calendar year 2027. More sources of computing capacity could give some customers additional purchasing options. But adopting a different accelerator can require engineering work, including adapting software and workflows, so availability alone does not make switching easy.
The announcement also matters because Google says it introduced four TPU systems in one year, a faster release pace than its earlier pattern of a new AI processor about every two years, according to Bloomberg. Google says these latest additions use existing silicon with changes to packaging and system design. That approach may allow the company to respond to customer requests faster than designing an entirely new chip, though the practical value depends on customer results once systems are available.
Products Worth Considering
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
Samsung Galaxy Tab A9+ 11-Inch Android Tablet, 64GB, Graphite
As an affiliate, we earn on qualifying purchases.
From Base TPUs to Specialized Systems
Google introduced its eighth-generation TPU family in April with two systems: TPU 8t for training and TPU 8i for inference. The company said the 8t delivers up to 2.7 times the performance per dollar of its prior Ironwood generation for large-scale training, and that the 8i improves performance per dollar by up to 80% for inference. Those comparisons are vendor claims, not independent benchmark results for the new variants.
Google’s published specifications show different priorities in the base systems. The 8t is listed with 216 GB of high-bandwidth memory, 128 MB of on-chip SRAM and peak FP4 compute of 12.6 petaflops. The 8i is listed with 288 GB of high-bandwidth memory, 384 MB of SRAM and 10.1 petaflops of peak FP4 compute. Its higher listed memory capacity and bandwidth reflect a design focused on moving data efficiently during inference.
Software support is part of Google’s pitch to customers weighing a change in hardware. The company points to TorchTPU, which it says allows PyTorch code and Hugging Face models to run natively on TPUs, as well as TPU support for the open-source inference engines vLLM and SGLang. Google also says Cloud Pathways, the runtime used behind Gemini, is generally available on Google Kubernetes Engine. These tools may ease adoption, but their presence does not establish that a customer can move a workload without cost or engineering changes.
““This is the first time we have introduced four TPU systems in a single year.””
— Google, in its November 9 announcement
Prices and Performance Still Unknown
Independent benchmarks are not available for either new system, so buyers cannot yet compare Google’s claims with results from external testing. The company has not published list prices, regional availability or a detailed launch schedule. Its TPU page marks all four eighth-generation systems as coming soon, and the 2027 timing for the new variants comes from Vahdat’s comments to Bloomberg.
It is also unclear how much engineering work customers will need to run existing applications on these systems, or whether the announced security features and memory gains will improve outcomes for particular workloads. Google’s claims describe intended capabilities; actual performance, cost and adoption remain to be seen.
Availability Expected in 2027
The next milestone is Google’s planned 2027 release of TPU 8tx and TPU 8i Max. Customers and cloud providers will need more detail on pricing, regions and access before they can assess the systems as practical purchasing options. Oracle’s planned offering of Google TPUs in its cloud is also expected to begin in calendar year 2027.
Once systems reach customers, published specifications and independent workload tests can help establish whether Google’s claimed latency, compute and memory gains hold up in practice. Until then, the announcement outlines Google’s product direction, while leaving key buying questions unanswered.
Key Questions
What did Google announce?
Google announced TPU 8tx for confidential inference and TPU 8i Max for workloads with very long inputs. Both extend the eighth-generation TPU family introduced in April.
When will the new TPUs be available?
Google’s Amin Vahdat told Bloomberg the two new variants are expected in 2027. Google has not published a more specific release date or regional availability.
Have the performance claims been independently tested?
No independent benchmarks for the new systems are included in the announcement. Google’s figures, including the claimed latency and memory gains, are vendor claims.
Why might businesses care about these systems?
The 8tx is aimed at organizations that need confidential inference, while the 8i Max targets applications that process very long inputs. The systems could give cloud buyers more workload-specific options, though prices and measured performance are not yet known.
Source: Thorsten Meyer AI
