TL;DR
Get everyday essentials 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 TPU systems on November 9: TPU 8tx for confidential inference and TPU 8i Max for workloads with very long inputs. The variants are built from existing chip designs, are due in 2027 and have no published prices or independent benchmarks.
Google announced two new systems for its eighth-generation AI chips on November 9: TPU 8tx, designed for confidential inference, and TPU 8i Max, aimed at workloads with very long inputs. The systems give cloud customers more options alongside Nvidia GPUs, but Google says the new variants will not be available until 2027, and has not published prices or independent benchmark results.
Google’s chief technologist for AI and infrastructure, Amin Vahdat, described the additions in a company blog post. The TPU 8tx is intended for regulated businesses and governments that need to protect data and model weights while they are being processed. Google says it uses the same silicon as TPU 8t, paired with fifth-generation AMD EPYC processors, and claims up to two times lower latency than TPU 8t with a near-zero performance penalty. Those figures are Google’s claims; outside testing has not been published.
The TPU 8i Max combines two low-latency TPU 8i chips in one package. Google says it provides double the compute, on-chip SRAM and HBM bandwidth of TPU 8i within the existing rack footprint. The company describes it as suited to large context windows, including AI agents handling an entire codebase or hours of video. No independent results are available to establish how it performs on those tasks.
The additions follow Google’s April 22 introduction of the eighth-generation TPU family at its Cloud Next conference. TPU 8t is for model training; TPU 8i is for inference, or running trained models to produce responses. Google says the new systems are reconfigurations of existing silicon rather than entirely new chip designs. Vahdat told Bloomberg the variants reflect customer requests. Google has not released prices or launch regions for the four systems, and its TPU product page lists them as coming soon.
Google’s new TPU options could challenge Nvidia
Google added two specialized systems to its eighth-generation TPU lineup. They offer more choices for AI workloads, but pricing and independent performance evidence are still missing.
Two systems, distinct workloads
Each variant targets a specific need for cloud customers.
TPU 8tx
Designed for inference that protects data and model weights during processing.
TPU 8i Max
Combines two TPU 8i chips for workloads such as large codebases or hours of video.
A broader TPU lineup
Google separates training, inference and specialized processing across four systems.
| System | Intended use |
|---|---|
| TPU 8t | Model training |
| TPU 8i | Low-latency inference |
| TPU 8tx | Confidential inference |
| TPU 8i Max | Very long inputs |
What buyers need to weigh
Another supplier could expand access and negotiating options, but switching depends on more than availability.
- More choice alongside Nvidia GPUs
- Systems aimed at specific workloads
- Google cites PyTorch, Hugging Face, vLLM and SGLang support
- Prices and launch regions are not announced
- Independent benchmarks are unavailable
- Software migration effort is unclear
More Choices for AI Compute
Google is trying to make TPUs a more practical alternative for companies that rely on Nvidia GPUs or want another source of AI computing capacity. Specialized systems may help customers match hardware to a particular workload: training, fast inference, confidential processing or long inputs. That could matter as demand for computing power grows and companies weigh availability, performance and cost.
Google says nearly 100 customers use its TPUs, and Alphabet has described the market as supply-constrained. Oracle plans to offer Google TPUs in its cloud starting in calendar year 2027, giving customers another route to access them. A second supplier can give buyers more negotiating options, but it does not make a switch automatic. Companies may need to adapt software and workflows, and the new systems’ prices and performance are not yet known.
Google is also working on software compatibility. It points to TorchTPU, which it open-sourced in October, and support for the open-source inference engines vLLM and SGLang. The company says PyTorch code and Hugging Face models can run natively on TPUs. These efforts could reduce friction for some developers, though the available information does not establish how much migration work a particular customer would face.
Products Worth Considering
Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black
Samsung Galaxy Tab A9+ 11-Inch Android Tablet, 64GB, Graphite
Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
As an affiliate, we earn on qualifying purchases.
How Google Split Its TPU Line
Google introduced TPU 8t and TPU 8i in April, separating training and inference into different systems. Training requires computing resources to develop a model; inference uses a trained model to generate outputs. Google says TPU 8t is designed for superpods of 9,600 chips, while TPU 8i is designed for low-latency responses.
Google’s published figures for the base systems show different hardware priorities. TPU 8t has 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 has 288 GB of high-bandwidth memory, 384 MB of SRAM, 10.1 petaflops of peak FP4 compute and 8,601 GB/s of bandwidth. These are vendor specifications, not independent comparative results.
Google claims TPU 8t offers up to 2.7 times better performance per dollar than its previous generation, Ironwood, for large-scale training. For TPU 8i, it claims up to an 80% improvement in performance per dollar for inference. The company has not supplied comparable independent benchmarks for the new 8tx and 8i Max systems.
“This is the first time we have introduced four TPU systems in a single year.”
— Google, in its November 9 blog post
Prices and Tests Are Pending
The announced availability window for TPU 8tx and TPU 8i Max is 2027, according to Vahdat’s comments to Bloomberg. Google has not named specific launch dates, cloud regions or prices. It has also not published independent benchmarks for either system, so buyers cannot yet compare their real-world performance or cost with Nvidia GPUs or Google’s earlier TPUs.
Google’s security and performance descriptions remain company claims. The available material does not include an outside assessment of TPU 8tx’s confidential computing protections or a detailed test of TPU 8i Max on long-context workloads. It is also unclear how much engineering work customers will need to move existing applications to TPUs.
Availability and Buyer Comparisons
Google says the two new systems are due in 2027. Buyers will need further details on pricing, regions and access before deciding whether either system fits their workloads. Oracle’s planned offering of Google TPUs in its cloud is also scheduled to begin in calendar year 2027.
As launch information and independent testing become available, customers will be able to compare performance, cost and software requirements across systems. Until then, Google’s published specifications and performance claims provide an outline of the products, but not a complete basis for procurement decisions.
Key Questions
What did Google announce?
Google announced TPU 8tx for confidential inference and TPU 8i Max for workloads with very long inputs. They join TPU 8t and TPU 8i in the eighth-generation lineup.
When will the new TPU systems be available?
Google’s Amin Vahdat told Bloomberg the two new variants are expected in 2027. Specific dates and cloud regions have not been announced.
Has Google published prices or independent benchmarks?
No. Google has not published prices or independent benchmark results for TPU 8tx and TPU 8i Max. Its current performance figures are company claims.
Could the TPUs offer an alternative to Nvidia GPUs?
They add another option for AI computing customers, and Google says nearly 100 customers use its TPUs. Whether a system is a practical alternative depends on its eventual price and performance, availability, software needs and the work involved in switching.
Source: Thorsten Meyer AI
