🔍 Read the full analysis: Google’s Two New AI Chips Bring The Year’s Total To Four—Why Nvidia Needs A Second Source on ThorstenMeyerAI.com
Get movie nights 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 TPU 8tx, designed for confidential inference, and TPU 8i Max, designed for workloads with very long inputs. The additions bring the eighth-generation TPU lineup to four systems, but Google says the new variants will be available in 2027; pricing and independent performance tests are not yet available.
Google announced two new AI chip systems on November 9: the TPU 8tx, aimed at confidential computing, and the TPU 8i Max, built for workloads with very long inputs. The additions bring Google’s eighth-generation TPU lineup to four systems announced in 2026, giving cloud customers another potential source of AI computing alongside Nvidia GPUs, although the new variants are not yet available.
The two systems are specialized versions of chips Google introduced in April. The TPU 8tx is intended for confidential inference, including work by governments and regulated businesses. Google says it protects sensitive data and model weights while they are being processed, and claims latency can be up to two times lower than TPU 8t with a near-zero performance penalty. The company says 8tx uses the same silicon as TPU 8t and pairs it with fifth-generation AMD EPYC processors.
The TPU 8i Max is designed for AI tasks that need to handle extensive context, such as analyzing a large codebase or hours of video. Google says it combines two low-latency TPU 8i chips in one package, providing twice the compute, on-chip SRAM and high-bandwidth memory bandwidth of TPU 8i within the existing rack footprint. Google has not published absolute performance or memory figures for either new system.
Google’s chief technologist for AI and infrastructure, Amin Vahdat, told Bloomberg the variants are not entirely new silicon designs, but systems reconfigured around customer needs. Google says both will be available in 2027. Its TPU product page lists all four eighth-generation systems as “Coming soon”; prices, cloud regions and independent benchmarks have not been published.
Google adds two specialized TPUs A second path to AI compute
TPU 8tx targets confidential inference; TPU 8i Max is built for very long inputs. The additions expand Google’s eighth-generation lineup, but customer access and independent testing are still ahead.
A Second Path to AI Compute
Google’s announcement matters because large AI developers and cloud customers have had limited alternatives to Nvidia’s widely used GPUs, which Bloomberg reports remain in demand beyond available supply. A credible second supplier could give buyers more choice and bargaining power, particularly if Google can deliver capacity at scale. It does not yet establish that the new TPUs will be cheaper or faster for a given workload.
The specialized systems also reflect a shift from a single general-purpose accelerator toward hardware configurations aimed at particular bottlenecks. The 8tx targets the protection of data during processing, while 8i Max emphasizes memory capacity and bandwidth for long-context tasks. Those features may matter to organizations handling sensitive information or large inputs, but their practical value will depend on testing, availability and the workloads customers run.
Google says nearly 100 customers use its TPUs, and Alphabet has described an AI-chip supply-constrained environment. Oracle plans to offer Google TPUs through its cloud beginning in calendar year 2027, according to the source material. Wider access could make the hardware relevant beyond customers already using Google Cloud, though the commercial terms and service locations remain unknown.
Popular right now
Bose QuietComfort Earbuds (2nd Gen) – True Wireless Bluetooth Earbuds with Active Noise Cancellation
Vitamix 5200 Blender with 64 oz. Container, Black
DREAME A3 AWD 1000 LiDAR Robot Lawn Mower
As an affiliate, we earn on qualifying purchases.
From Two TPUs to Four
Google introduced the eighth-generation TPU 8t and TPU 8i at its Cloud Next conference on April 22, 2026. The 8t is aimed at training large models, while the 8i is designed for inference, when a trained model generates responses. Google says 8t systems can scale to superpods of 9,600 chips. Its published performance-per-dollar comparisons with the previous generation are vendor figures, not independent test results.
The November announcement extends that product split into more specialized systems. Google has said it can make a system variant by changing elements such as packaging, memory, networking or security without designing a new chip from scratch. Vahdat told Bloomberg the additions respond to tasks customers requested. Google has also highlighted software work, including TorchTPU support for PyTorch and Hugging Face models and TPU backends for the open-source inference engines vLLM and SGLang.
The available figures for the base systems show why memory is part of Google’s pitch. Google lists TPU 8t with 216 GB of high-bandwidth memory and 6,528 GB/s of memory bandwidth. TPU 8i has 288 GB and 8,601 GB/s, while its peak compute figure is lower than 8t’s. Google’s published specifications are company-provided; independent comparisons of the new variants are not yet available.
“This is the first time we have introduced four TPU systems in a single year.”
— Google, in its November 9 announcement
Availability and Test Results
Neither new variant is available for customers to test or buy yet. Google has said they are due in 2027, but has not given specific launch dates, prices or cloud regions. Its product page marks all four eighth-generation systems “Coming soon.”
Google’s claims about latency, compute, memory and bandwidth have not been independently tested for the new systems. The company has not released full specifications for TPU 8tx or TPU 8i Max, and the supplied information does not establish how they compare with Nvidia GPUs on particular workloads, cost or availability. It also remains unclear how readily customers can move existing software and workloads to TPUs; doing so can require engineering work.
Launch Details Due in 2027
The next meaningful developments are Google’s release of pricing, deployment regions and fuller specifications, followed by customer access and independent performance testing. Google has indicated that TPU 8tx and TPU 8i Max are expected in 2027, but has not provided a more precise timetable.
Oracle’s planned Google TPU offering, also expected to begin in calendar year 2027, will provide another milestone for access outside Google’s own cloud. Until launch details and outside benchmarks are available, buyers will have to weigh Google’s stated capabilities against the cost of adapting workloads and the performance of hardware they can already obtain.
Key Questions
What did Google announce?
Google announced TPU 8tx, aimed at confidential inference, and TPU 8i Max, designed for workloads with very long inputs. They join TPU 8t and TPU 8i in Google’s eighth-generation lineup.
When will the new Google TPUs be available?
Google’s Amin Vahdat told Bloomberg the two new variants are expected in 2027. Google has not announced specific release dates, prices or cloud regions.
Are these entirely new chip designs?
No. Vahdat told Bloomberg they are not completely new silicon designs. Google describes them as specialized systems built from existing chip silicon and reconfigured for different tasks.
Has Google’s performance claims been independently verified?
No independent benchmarks for TPU 8tx or TPU 8i Max are included in the source material. Claims such as lower latency or doubled memory bandwidth come from Google and have not yet been independently tested.
Why might the announcement matter to AI customers?
Google’s TPUs could offer customers another source of AI computing beyond Nvidia GPUs. Whether they provide a practical alternative will depend on availability, pricing, measured performance and the work needed to adapt software.
Source: ThorstenMeyerAI.com
