🔍 Read the full analysis: Why Google’s Four AI Chips In One Year Could Mean More Competition For Nvidia on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Google announced TPU 8tx for confidential inference and TPU 8i Max for long-context workloads, bringing its eighth-generation TPU lineup to four systems in 2026. The new variants are not available yet, and their performance claims have not been independently tested; Google says they are due in 2027.
Google announced two additions to its eighth-generation AI chip lineup on November 9: TPU 8tx, designed for confidential computing, and TPU 8i Max, intended for workloads that need very long inputs. The additions bring Google to four TPU systems announced in 2026 and give cloud customers another potential option in a market where Nvidia GPUs remain dominant, though the new systems are not available yet.
The two systems build on the TPU 8t and TPU 8i, which Google introduced in April. The company describes TPU 8t as a chip system for training large AI models and TPU 8i as a system for inference—the process of running a trained model to produce answers. The new versions target different constraints: security for TPU 8tx and memory capacity for TPU 8i Max.
Google says TPU 8tx uses the same silicon as TPU 8t, paired with fifth-generation AMD EPYC processors, and is designed to protect sensitive data and model weights while they are in active use. The company claims it can provide up to two times lower latency than TPU 8t with a near-zero performance penalty. Google says TPU 8i Max puts two TPU 8i chips on one package, doubling that system’s compute, on-chip SRAM and high-bandwidth memory bandwidth within the existing rack footprint.
Those figures are Google’s claims, not independent test results. The company has not published absolute specifications for the two new systems, prices or cloud regions where they will be offered. Google’s TPU product page lists all four eighth-generation systems as coming soon. Amin Vahdat, Google’s chief technologist for AI and infrastructure, told Bloomberg the two new variants would be available in 2027.
Google expands its TPU lineup Competition depends on delivery
Two new systems target confidential inference and very long inputs. They could give cloud customers another source of AI computing capacity, but availability and performance remain unproven.
Four systems, distinct workloads
Google says the new variants build on its TPU 8t and TPU 8i systems.
TPU 8t
Designed to train large AI models.
TPU 8i
Runs trained models to produce answers.
TPU 8tx
Targets sensitive data and model weights while in use.
TPU 8i Max
Pairs two TPU 8i chips for workloads with very long inputs.
What buyers could gain
More options may help companies find capacity and hardware suited to specific workloads.
- Another source of AI computing capacity
- More choice for specialized workloads
- Potentially stronger supplier negotiations
- Software and team changes may be needed
- Compatibility and measured results still matter
- Launch capacity and cloud regions are unknown
Claims await real-world checks
Google has shared relative performance claims, but no independent tests or complete specifications for the new systems.
Google says TPU 8tx and TPU 8i Max are expected in 2027.
Pricing, regions, capacity and technical specifications remain unannounced.
Deployments and independent results can show how performance and software support hold up.
A Second Route to AI Capacity
Google’s announcement matters because access to computing capacity is a practical constraint for companies training and operating AI models. Nvidia GPUs remain the market’s leading option, and Bloomberg has reported that demand for Nvidia processors exceeds supply. If Google can offer TPU systems at scale, customers may gain another source of compute rather than depending on a single chip supplier or waiting for GPU capacity.
The four-system lineup also reflects a shift in how Google is presenting its chips. Rather than relying on one general-purpose processor, it is describing systems tailored to training, inference, confidential workloads and long-context tasks. That could give cloud customers more choice for specific applications. But specialization only changes purchasing decisions if the systems are available, perform as promised and work with customers’ software.
For buyers, a second supplier can improve negotiating leverage and make it easier to match hardware to a workload. It does not make switching effortless. Companies may need to adapt software, retrain teams or evaluate whether a model performs similarly on a TPU. Google is working on compatibility through TorchTPU, which it says allows PyTorch code and Hugging Face models to run natively, and through TPU support in the open-source inference engines vLLM and SGLang. The practical value of those efforts will depend on customer experience and measured results.
Products Worth Considering
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
Amazon Kindle Paperwhite (2024, 16 GB) – 7-inch Glare-Free Display, Black
Logitech MX Master 4 for Mac Wireless Ergonomic Mouse, White Silver
As an affiliate, we earn on qualifying purchases.
From Two TPU Systems to Four
Google introduced its eighth-generation TPUs at Cloud Next on April 22, 2026, splitting the generation into the TPU 8t training system and the TPU 8i inference system. Google said TPU 8t was designed for superpods containing up to 9,600 chips. Its April materials claimed up to a 2.7-times improvement in performance per dollar over the previous generation, Ironwood, for large-scale training. For TPU 8i, Google claimed up to an 80% performance-per-dollar improvement for inference. These comparisons are vendor figures, not independent benchmarks of the new systems.
The company says it can add variants more quickly by changing packaging, memory, networking or security features around existing silicon rather than designing a new chip from scratch. Vahdat told Bloomberg the two new systems are not completely new silicon designs, but configurations intended to address customer requests. Google also says nearly 100 customers use its TPUs, while Alphabet has described a supply-constrained environment. Oracle has said it will offer Google TPUs in its cloud beginning in calendar year 2027, potentially extending their reach beyond Google Cloud.
Google’s published figures for the base systems help explain the focus on different workloads. TPU 8t has 216 gigabytes of high-bandwidth memory and 6,528 gigabytes per second of memory bandwidth, according to Google. TPU 8i has 288 gigabytes of high-bandwidth memory and 8,601 gigabytes per second of bandwidth, alongside more on-chip SRAM. For tasks involving long prompts or large amounts of information, moving data through memory can limit performance as much as raw computing power. Google says TPU 8i Max is aimed at such “massive context” workloads.
“This is the first time we have introduced four TPU systems in a single year.”
— Google, in its November 9 TPU announcement
Availability and Performance Still Unproven
The main unknown is how the new systems perform under real customer workloads. Google has shared relative performance claims for TPU 8tx and TPU 8i Max but no independent benchmarks or complete specifications for either. Its April performance-per-dollar claims apply to the base TPU 8 systems and are also company-provided figures.
Google has not announced prices, specific launch dates or regions for the new variants. It is also unclear how much capacity will be available when they launch, or how broadly the systems will support customers’ existing software and models. Google’s statement that nearly 100 customers use TPUs does not specify how many use the eighth generation or how much capacity they consume.
The competitive impact is uncertain as well. Nvidia’s continued supply constraints could make alternatives attractive, but companies may still prefer GPUs if their software, staffing and existing cloud arrangements are built around them. The announcement alone does not show that customers will move workloads from Nvidia to Google.
2027 Launches and Customer Testing
The next milestones are the launch of TPU 8tx and TPU 8i Max, which Vahdat told Bloomberg are expected in 2027, and Google’s publication of pricing, regional availability and fuller technical details. Oracle’s planned 2027 offering will also show whether Google TPUs become accessible through another major cloud provider.
After launch, customer deployments and independent evaluations should provide a clearer picture of performance, software compatibility and supply. Until then, Google has announced a broader set of options, but buyers cannot yet compare the new systems against Nvidia GPUs using public, independently tested results.
Key Questions
What did Google announce?
Google announced TPU 8tx, aimed at confidential inference, and TPU 8i Max, designed for workloads involving very long inputs. They join the TPU 8t training system and TPU 8i inference system announced in April.
When will the two new TPUs be available?
Google’s Amin Vahdat told Bloomberg the two variants are expected in 2027. Google has not yet provided exact launch dates, prices or regions.
Have the new systems been independently tested?
No independent benchmarks were provided in the source material. The performance and latency figures published so far are Google’s own claims.
Does this mean Google is replacing Nvidia?
No. The announcement adds another potential source of AI computing capacity, but it does not establish that customers will switch from Nvidia GPUs. Availability, performance, price and software compatibility remain to be tested.
What is different about TPU 8i Max?
Google says TPU 8i Max combines two TPU 8i chips on one package, doubling the base system’s compute, on-chip SRAM and memory bandwidth. It is intended for workloads such as processing large codebases or long video inputs.
Source: ThorstenMeyerAI.com
