🔍 Read the full analysis: Google’s AI Hardware Push: Four Chips In A Year, More Choice Ahead? on ThorstenMeyerAI.com
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
TL;DR
Google announced two specialized systems for its eighth-generation TPUs: TPU 8tx for confidential inference and TPU 8i Max for workloads with very long inputs. The additions bring Google’s total to four TPU systems announced in 2026, but none is currently available; the two new variants are expected in 2027, with prices and independent benchmarks still unknown.
Google announced two new eighth-generation TPU systems on November 9, expanding its AI chip lineup to four systems introduced in 2026. The company says TPU 8tx is designed for confidential inference and TPU 8i Max for workloads with very long inputs, but neither is available yet; Google’s TPU page lists all four systems as coming soon.
The additions build on TPU 8t, intended for training AI models, and TPU 8i, designed for inference, or running trained models to produce responses. Google introduced those systems at its Cloud Next conference on April 22. The November announcement extends the lineup with versions aimed at specific customer needs rather than introducing two entirely new chip designs.
Google says TPU 8tx uses the same silicon as TPU 8t, paired with fifth-generation AMD EPYC processors, and is optimized for secure, confidential inference. The company says it can protect sensitive data and model weights while they are being processed, and claims up to two times lower latency than TPU 8t with a near-zero performance penalty. Those figures and security statements are Google’s claims, not independently verified results.
TPU 8i Max combines two TPU 8i chips on one package. Google says the system doubles the compute, on-chip SRAM and high-bandwidth memory bandwidth of TPU 8i while fitting twice the silicon into the existing rack footprint. The company is targeting tasks such as processing a large codebase or long stretches of video. It has not published absolute performance or memory figures for either new system.
Google’s AI hardware push Four systems, more choice ahead?
Google has expanded its eighth-generation TPU lineup with systems for confidential inference and very long inputs. Their practical value will depend on availability, pricing, independent testing and software support.
Four systems, distinct roles
The November additions extend the training and inference systems announced in April.
TPU 8t
Designed to train AI models.
TPU 8i
Designed to run trained models and produce responses.
TPU 8tx
Targets confidential inference; Google says it uses the same silicon as TPU 8t.
TPU 8i Max
Combines two TPU 8i chips for workloads such as large codebases or long video.
What customers gain
More providers may give organizations additional ways to access AI computing capacity.
- Regulated data needing confidential processing
- Applications handling especially long inputs
- Organizations seeking options beyond Nvidia GPUs
- Availability, pricing and capacity
- Performance on real workloads
- Software support and migration effort
Evidence still to come
The new systems have no published absolute specifications or independent benchmarks in the source material.
Google expects the new variants in 2027; specific dates, regions and prices are unannounced.
Google’s performance, latency and security statements remain company claims until independently tested.
Buyers will need to assess capacity, software support and the engineering needed to move applications.
A Wider Route Beyond Nvidia
The announcement adds another option for organizations seeking AI computing capacity beyond Nvidia GPUs. Google says nearly 100 customers use its TPUs, while Alphabet has described the market as supply-constrained. Oracle is also set to offer Google TPUs through its cloud starting in calendar 2027, according to the source material. More providers could give customers additional ways to access chips and negotiate capacity.
But adding a product to a lineup does not establish that it will be a practical substitute for Nvidia hardware. Buyers will need to compare availability, pricing, performance and software support, then account for the engineering work involved in moving workloads between platforms. The new systems’ specifications and availability could matter most to customers with regulated data or applications that need to handle especially long inputs.
For now, the evidence is limited. Google’s published comparisons are vendor claims, and there are no independent benchmarks for TPU 8tx or TPU 8i Max. The announcement signals a more specialized product strategy, but does not yet show how the systems will perform in customers’ workloads or how much capacity Google can supply.
Products Worth Considering
Bose QuietComfort Earbuds (2nd Gen) – True Wireless Bluetooth Earbuds with Active Noise Cancellation
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
DJI Osmo Mobile 8 3-Axis Phone Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
From Two TPUs to Four Systems
Google’s April announcement divided its eighth-generation TPU offering between training and inference. The company says TPU 8t can be organized into superpods of 9,600 chips and lists 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. For TPU 8i, Google lists 288 GB of high-bandwidth memory, 384 MB of SRAM, 10.1 petaflops of peak FP4 compute and 8,601 GB/s of bandwidth.
Those figures are for the two base systems, not the November additions. Google has claimed up to a 2.7-times performance-per-dollar improvement for TPU 8t over its previous generation, Ironwood, in large-scale training, and up to an 80% improvement for TPU 8i in inference. These are Google’s own comparisons; independent testing of the new systems has not been reported in the source material.
The company has also pointed to software work intended to make TPUs easier to use. It says TorchTPU supports PyTorch code and Hugging Face models, and cites backends for the open-source inference engines vLLM and SGLang. Google’s Cloud Pathways runtime, used by Google DeepMind behind Gemini, is generally available on Google Kubernetes Engine, according to the source material.
“This is the first time we have introduced four TPU systems in a single year.”
— Google, in its November 9 announcement
Price, Capacity and Testing Remain Unknown
Google has not announced prices, cloud regions or specific launch dates for the new variants. Amin Vahdat told Bloomberg they would be available in 2027, but the company has not given a more precise timetable in the material provided. Google’s TPU page lists all four eighth-generation systems as coming soon.
There are also no independent benchmarks for TPU 8tx or TPU 8i Max, and Google has not published their absolute specifications. Its performance, latency, memory and security statements remain vendor claims until independently tested. The available figures do not show how the systems compare with current Nvidia hardware or how they perform across different customer workloads.
It is also unclear how much TPU capacity Google will be able to offer as demand grows, or how readily customers can shift existing applications to the systems. Software support may reduce some friction, but migration costs and real-world availability will affect whether the new options change purchasing decisions.
Availability Expected During 2027
The next milestones are Google’s expected 2027 availability for TPU 8tx and TPU 8i Max and Oracle’s plan to offer Google TPUs in its cloud beginning that calendar year. Google has not yet said which regions will receive the new variants first or how customers will obtain them.
Potential buyers will be watching for pricing, capacity details and independent performance testing once systems can be evaluated. Those details will help show whether the confidential-computing and long-context designs provide meaningful advantages for customers, and whether Google’s expanded lineup offers a workable alternative for particular AI workloads.
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 TPU 8t for training and TPU 8i for inference, bringing the eighth-generation lineup to four systems.
When will TPU 8tx and TPU 8i Max be available?
Google’s Amin Vahdat told Bloomberg the two new variants would be available in 2027. Google has not provided a specific launch date, pricing or regional availability in the source material.
Are the new systems new chip designs?
Not according to Google. Vahdat told Bloomberg the variants are not completely new silicon designs. Google says TPU 8tx uses the same silicon as TPU 8t, while TPU 8i Max packages two TPU 8i chips together.
Have the performance claims been independently tested?
The source material provides Google’s claims, not independent benchmark results, for the two new systems. Google has not published their absolute specifications, and no independent testing is reported there.
Why might customers choose a TPU?
TPUs give customers another potential source of AI computing capacity, with systems aimed at training, inference, confidential workloads and long inputs. Whether a TPU suits a particular buyer will depend on availability, cost, tested performance and the work needed to move software from another platform.
Source: ThorstenMeyerAI.com
