🔍 Read the full analysis: Google’s New AI Chips And The Growing Search For Nvidia Alternatives on ThorstenMeyerAI.com
Get smart everyday buys 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 new configurations of its eighth-generation TPUs on Nov. 9: TPU 8tx for confidential computing and TPU 8i Max for workloads with very long inputs. Google says both are due in 2027, but has not released prices or independent test results. The additions give cloud customers another potential option alongside Nvidia GPUs, though availability and real-world performance remain unconfirmed.
Google announced two new systems built around its eighth-generation AI chips on Nov. 9, adding configurations for confidential computing and very long model inputs as cloud providers and customers seek alternatives to Nvidia GPUs. Google says the TPU 8tx and TPU 8i Max will be available in 2027; neither has a published price or independent performance test.
The new systems extend the TPU 8 generation Google introduced at its Cloud Next conference on April 22. That launch included TPU 8t, designed for training AI models, and TPU 8i, designed to run models and return answers, a process known as inference. The November additions adapt those base systems for more specific tasks rather than introducing entirely new silicon designs.
Google says TPU 8tx is designed for confidential inference, aimed at customers such as regulated businesses and governments. It uses the same silicon as TPU 8t and is paired with fifth-generation AMD EPYC processors. Google claims it can provide up to twice the lower latency of TPU 8t with a near-zero performance penalty. These figures are vendor claims; outside testing has not been reported.
TPU 8i Max puts two low-latency TPU 8i chips on a package. Google says it doubles TPU 8i’s compute, on-chip SRAM and high-bandwidth memory bandwidth, while fitting twice the silicon within the existing rack footprint. The company is targeting workloads that need to process very large amounts of input, such as an AI agent working across a codebase or hours of video. Google has not provided absolute specifications for either new system.
Google’s New AI Chips Add Options, Not Proof
Two new TPU configurations target confidential inference and very long inputs. They could widen cloud customers’ choices, but availability and real-world performance remain unconfirmed.
Two configurations, distinct jobs
Both build on Google’s eighth-generation TPU silicon and adapt existing systems for specific workloads.
TPU 8tx
Designed for sensitive workloads at regulated businesses and governments. Google claims lower latency, but independent testing has not been reported.
TPU 8i Max
Targets tasks such as working across a codebase or hours of video. Google says it doubles TPU 8i’s compute, on-chip SRAM and memory bandwidth.
What cloud buyers gain
A second platform could give customers another way to secure AI capacity and negotiate cloud services.
- Confidential workloads handling sensitive data
- Inference with very long prompts or large files
- Customers exploring options alongside Nvidia GPUs
- Can existing software and workloads move easily?
- Will capacity and cloud regions meet demand?
- How will operating costs compare?
Proof still has to come
The announcement outlines intended uses; buyers still lack the details needed to compare systems.
Prices, specifications, regions and precise release dates have not been announced.
Google expects both configurations in 2027; customers have not tested the new hardware.
Comparable workload results are needed to judge performance, cost and security against Nvidia systems.
More Options for AI Compute Buyers
The announcement adds to the options cloud customers may have as they seek computing capacity for AI training and inference. Nvidia remains the dominant supplier of AI GPUs, and demand for its processors has exceeded available supply, according to the source report. A viable second platform could give customers another way to secure capacity and negotiate for cloud services, though that benefit depends on the new systems becoming available and meeting customers’ needs.
Google’s focus on distinct configurations also reflects the different demands of AI workloads. Training large models calls for substantial compute, while inference can be constrained by latency, memory capacity or the time needed to move information through a system. A confidential-computing option could be relevant to organizations handling sensitive workloads; a higher-memory system could help with long prompts or large files. Those are intended use cases, not independently verified results.
For buyers, switching hardware is not just a matter of comparing chip specifications. Software compatibility, engineering work, availability and the cost of running a workload all affect whether a move from Nvidia GPUs makes sense. Google points to support for PyTorch, Hugging Face models and open-source inference tools, but the new hardware has not yet been tested by customers. The announcement signals a broader choice in principle, not an immediate substitute for existing systems.
Products Worth Considering
Dell XPS Tower 8960 Desktop, Intel Core i7-13700, NVIDIA GeForce RTX 4060, 16GB DDR5, 512GB SSD
Apple Mac mini (M6, 12-core CPU and GPU, 24GB RAM, 1TB SSD)
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
As an affiliate, we earn on qualifying purchases.
From Two TPUs to Four Systems
Google introduced TPU 8t and TPU 8i in April, marking a shift toward offering separate systems for training and inference. The company says TPU 8t can be deployed in superpods of up to 9,600 chips. For the base systems, Google reported 216 GB of high-bandwidth memory and 12.6 petaflops of peak FP4 compute for TPU 8t; TPU 8i has 288 GB of high-bandwidth memory and 10.1 petaflops of peak FP4 compute. These are specifications and performance claims supplied by Google, not independent comparisons.
Google also claimed performance-per-dollar improvements over the prior Ironwood generation: up to 2.7 times for TPU 8t in large-scale training and up to 80% for TPU 8i in inference. The source material provides no independent validation of those figures. Google says it has introduced four TPU systems in one year, while Bloomberg reported that the company previously released a new AI processor about every two years.
The company has also been building out software support, including TorchTPU and backends for the vLLM and SGLang inference engines, according to its announcements. Separately, Oracle plans to offer Google TPUs through its cloud service starting in calendar year 2027. That could make the chips accessible beyond Google Cloud, although the timing and terms of that offering remain ahead.
“This is the first time we have introduced four TPU systems in a single year.”
— Google, in its Nov. 9 announcement
Prices, Tests and Availability Remain Open
The two new systems are not yet available. Google’s TPU page lists the eighth-generation systems as “Coming soon,” and Vahdat told Bloomberg that the new variants are expected in 2027. Google has not announced prices, specific cloud regions or a more precise release schedule.
There are also no independent benchmarks for TPU 8tx or TPU 8i Max. Google has supplied relative claims for the new configurations but not absolute specifications for them, so buyers cannot yet compare their performance, operating costs or security characteristics against Nvidia systems using independent results. It is also unclear how much engineering work customers would need to adapt existing workloads, despite Google’s software compatibility efforts.
The source material says Google reports nearly 100 customers using TPUs and that Alphabet has described a supply-constrained environment. It does not provide a detailed breakdown of those customers, the amount of available capacity, or how much of their computing they run on TPUs. Those unknowns limit what can be concluded about the new systems’ likely reach.
Watch for 2027 Cloud Rollouts
The next milestones are Google’s release of more detailed specifications, prices, supported regions and availability dates. Customers will also need access to the systems before independent researchers or cloud users can report comparable results on real workloads.
Oracle’s planned TPU offering in calendar year 2027 is another development to watch. If it proceeds, customers may be able to access Google’s chips through a cloud provider other than Google Cloud. For now, the announcement establishes Google’s intended product lineup and target uses; it does not establish when buyers can obtain the systems or how they will compare with Nvidia GPUs in practice.
Key Questions
What did Google announce?
Google announced TPU 8tx, a configuration aimed at confidential inference, and TPU 8i Max, designed for workloads with very long inputs. Both build on existing eighth-generation TPU silicon.
When will TPU 8tx and TPU 8i Max be available?
Google’s Amin Vahdat told Bloomberg the two variants are expected in 2027. Google has not provided a more specific release date.
Are the new TPUs proven alternatives to Nvidia GPUs?
Not yet. Google has published its own performance claims, but there are no independent tests of the new systems in the source material. Their real-world performance and cost compared with Nvidia GPUs remain unknown.
What is TPU 8i Max designed to handle?
Google says TPU 8i Max is intended for very long-context workloads, such as processing a large codebase or hours of video. The system combines two TPU 8i chips on one package, according to the company.
Source: ThorstenMeyerAI.com
