Google has introduced two systems in its eighth-generation AI accelerator lineup, targeting workloads that need stronger data protection or more memory capacity. The TPU 8tx is designed for confidential computing, while the TPU 8i Max combines two chips to serve memory-intensive applications. Google says the additions bring its total to four TPU systems shipped in a single year.
New systems use existing processors
The new variants do not introduce a new chip generation. Instead, Google has changed how existing processors are packaged and connected, along with their memory configurations. The company describes this modular approach as a way to bring new system options to market in months rather than years.
That distinction matters for customers choosing AI hardware. A processor is only one part of a system: packaging, networking and memory can shape how well it handles a particular job. Google’s latest designs apply those system-level changes to two different needs, rather than asking customers to wait for entirely new silicon.
TPU 8tx protects data during use
The TPU 8tx is aimed at enterprises and government agencies that work with confidential information. Data is typically encrypted while stored or being sent, but must be decrypted in memory while a processor handles it. Google says the 8tx is designed to protect and isolate data and model weights during active processing as well.
Nvidia already offers a comparable capability for GPUs through its confidential computing architecture. Google’s system uses the same low-latency Boardfly network it developed for real-time inference. The company says the TPU 8tx can deliver up to twice the latency reduction compared with the TPU 8t.
Google says organizations can use one system to fine-tune models, distill larger models into smaller ones, and train AI agents with reinforcement learning. The stated benefit is that sensitive data does not need to move between separate clusters for those tasks. For customers with strict confidentiality needs, keeping the work within one system could simplify how they handle data across development and use.
TPU 8i Max addresses memory limits
The TPU 8i Max focuses on workloads where memory bandwidth can become a bottleneck. The source describes AI agents working simultaneously across large codebases or hours of video as examples in which the demand for memory can constrain performance.
Google’s response is to place two TPU 8i chips on a single package. According to the company, that configuration doubles compute power, memory and memory bandwidth while keeping the same rack format. This can provide more performance per square meter of data center space, particularly for applications that need to keep large amounts of information accessible while they run.
Software support broadens access
Google is also working to make its TPUs easier for developers to use. It has released TorchTPU as open source, with the goal of allowing PyTorch code to run on the hardware with minimal changes. The company also supports inference frameworks vLLM and SGLang.
Software compatibility can influence whether customers adopt a new accelerator. If teams can use familiar code and frameworks, moving workloads to different hardware may require fewer changes. Google’s software work therefore complements the new system designs: the company is addressing both the kinds of jobs the hardware can handle and the effort involved in using it.
Inference is shaping the chip competition
The expansion reflects Google’s shift from treating TPUs mainly as an internal tool to marketing them as a platform for outside customers. The company now counts nearly 100 external TPU customers. Its latest systems are part of a broader effort to make its in-house accelerators useful beyond Google’s own operations.
As AI applications spread, attention is shifting from expensive training runs toward inference: the day-to-day operation of models. Specialized chips such as TPUs can be suited to that work, and Google is not alone in pursuing custom accelerators. Amazon, Microsoft and OpenAI are also developing their own chips.
Nvidia is strengthening its inference offering through an acqui-hire and licensing deal with Groq. The competition is increasingly about giving customers practical options for running AI models, including systems designed for particular workload constraints. Google’s two new configurations address confidentiality and memory capacity while its software efforts aim to make the hardware easier to adopt.