TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Google announced TPU 8tx for confidential inference and TPU 8i Max for workloads with very long inputs, extending its eighth-generation lineup to four systems. Google says the new variants are due in 2027; prices, regions and independent performance results have not been published.
Google announced two new eighth-generation TPU systems on November 9: the TPU 8tx for confidential inference and the TPU 8i Max for workloads that need to process very long inputs. The additions give Google Cloud customers more specialized alternatives to Nvidia-based computing, but they are not available yet: Google lists all four eighth-generation systems as “Coming soon,” and its chip chief told Bloomberg the two new variants are expected in 2027.
The TPU 8tx is designed for secure and confidential inference, with Google targeting regulated businesses and government users. Google says it protects sensitive data and model weights while they are in active use. The system uses the same silicon as the TPU 8t, paired with fifth-generation AMD EPYC processors. Google claims up to two times lower latency than the TPU 8t with a near-zero performance penalty; that figure has not been independently tested.
The TPU 8i Max is designed for large context windows, such as an AI agent processing a software codebase or hours of video. Google says it puts two TPU 8i chips on one package, doubling the base system’s compute, on-chip SRAM and high-bandwidth memory bandwidth within the existing rack footprint. Those specifications and intended uses come from Google; no independent tests of the new system are available.
Google introduced the base TPU 8t and TPU 8i at its Cloud Next conference on April 22. The 8t is aimed at model training, while the 8i is designed for inference, or running a trained model to produce answers. Google’s November announcement describes the new systems as configurations built from existing silicon, with changes to packaging and other system components. Google has not published prices or which cloud regions will offer them.
Google adds two AI chips to its TPU lineup
The new systems target confidential inference and very long inputs, giving Google Cloud customers more specialized options alongside Nvidia-based computing.
Two systems, distinct workloads
Google says each variant is configured for a specific kind of AI work.
TPU 8tx
Designed to protect sensitive data and model weights while in use, for regulated businesses and government users.
TPU 8i Max
Pairs two TPU 8i chips on one package for workloads such as analyzing a codebase or hours of video.
A broader TPU family
The original eighth-generation systems focus on training and inference.
- Targets model training
- Designed for superpods of up to 9,600 chips
- Targets low-latency inference
- Designed to run trained models and produce answers
What a second source could change
More hardware choice may help large buyers plan capacity and match systems to workloads.
Assess whether confidential inference or very long inputs match the new systems.
Software support may help compatibility, but migration effort and costs remain unclear.
Independent results and workload-specific cost comparisons are not available yet.
Availability and evidence
Google lists all four eighth-generation systems as coming soon; specific dates and cloud regions have not been published.
| What businesses need | What is known |
|---|---|
| Launch timing | New variants expected in 2027 |
| Pricing and regions | Not announced |
| Independent benchmarks | Not available for the new systems |
| Security assessment | No external assessment of 8tx protections provided |
More Options for AI Infrastructure
The announcement expands the choices for organizations that need AI computing and may want an alternative to relying on Nvidia GPUs. Specialized systems could let customers match hardware to a particular workload, including confidential inference or handling extensive inputs. Google says it introduced four TPU systems in one year, a faster release pace than its earlier pattern of releasing a processor every two years, as Bloomberg reported.
A second supplier can give large buyers another option when negotiating capacity and planning deployments. But the existence of a competing chip does not by itself make a workload easy or economical to move. Companies may need to adapt software, assess performance and retrain staff. Google points to support for PyTorch and Hugging Face models through TorchTPU, as well as TPU support in the open-source inference engines vLLM and SGLang. These tools may help with compatibility, but they do not establish how much effort a particular customer would need to switch.
Availability is also central to the business case. Google says nearly 100 customers use TPUs, while Alphabet has described a supply-constrained environment. Oracle has said it plans to offer Google TPUs in its cloud from calendar year 2027. These developments suggest demand and distribution are part of the story, but the new systems’ prices, capacity and customer access remain unknown.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
As an affiliate, we earn on qualifying purchases.
From Two Base Chips to Four
Google announced the eighth-generation TPU family in April 2026 with two systems for different stages of AI work. The TPU 8t targets training and is designed for superpods of up to 9,600 chips. The TPU 8i targets low-latency inference. Google claims up to 2.7 times better performance per dollar for 8t training and up to 80% better performance per dollar for 8i inference compared with the previous generation, Ironwood. These are vendor claims, not independent benchmark results.
Google’s published figures for the base systems show different hardware priorities. The 8t has 216 GB of high-bandwidth memory, 128 MB of SRAM, 12.6 petaflops of peak FP4 compute and 6,528 GB/s of memory bandwidth. The 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 memory bandwidth. The figures help explain why Google presents the 8i as suited to inference workloads that need to move data quickly, while the 8t is positioned for training. They do not show how either system performs across independent real-world tests.
Google says systems can be adapted faster than entirely new chip designs can be developed. In comments to Bloomberg, Google chief technologist for AI and infrastructure Amin Vahdat said the new variants are not completely new silicon designs, but configurations intended for tasks customers requested. The company also cites Cloud Pathways, a runtime used by Google DeepMind for Gemini, which it says is now generally available on Google Kubernetes Engine.
““This is the first time we have introduced four TPU systems in a single year.””
— Google, in its November 9 announcement
Pricing, Access and Performance
Prices, launch regions and specific availability dates have not been announced. Google’s TPU page lists the four eighth-generation systems as “Coming soon,” while Vahdat told Bloomberg the two new variants will be available in 2027. The information provided does not specify when customers can reserve capacity or how much will be available.
Google’s performance and security descriptions are vendor claims. There are no independent benchmarks for the 8tx or 8i Max in the source material, and no external assessment of the 8tx’s confidential-computing protections. It is also unclear how the new variants will compare with Nvidia systems on customer workloads, or what software changes and costs customers would face when moving from GPUs to TPUs.
Awaiting 2027 Availability Details
The next concrete milestones are Google’s rollout details for 2027 availability, including pricing, cloud regions and access to capacity. Google may also publish further technical information as launch approaches. Customers will need independent performance results and workload-specific cost comparisons to judge whether either system fits their needs.
Oracle’s planned TPU offering in calendar year 2027 could provide another route to access Google’s chips beyond Google Cloud. The timing and scope of that offering, like the rollout details for the new systems, remain to be confirmed.
Key Questions
What did Google announce?
Google announced two eighth-generation TPU systems: TPU 8tx for confidential inference and TPU 8i Max for workloads with very long inputs.
When will the new TPUs be available?
Google’s Amin Vahdat told Bloomberg the two new variants are expected in 2027. Google has not provided specific launch dates or cloud regions.
Has Google published independent performance results?
No independent benchmarks for the TPU 8tx or TPU 8i Max are included in the source material. Their stated performance figures are Google claims.
How is the TPU 8i Max different from the TPU 8i?
Google says the 8i Max combines two TPU 8i chips on one package, doubling compute, SRAM and memory bandwidth within the same rack footprint. It is aimed at very long-context workloads.
Does the announcement mean businesses can easily replace Nvidia GPUs?
No. The new systems add another option, but customers still need to assess software compatibility, engineering effort, capacity and costs for their own workloads. Google has not published prices or independent comparisons for the new variants.
Source: Thorsten Meyer AI
