Google has added two specialized chips to its AI hardware lineup: one designed to protect sensitive data during AI work, and another built to handle larger workloads. The move extends the company’s effort to supply more of the chips used in its cloud computing business with hardware it develops in-house.
Two chips for different AI needs
The new processors, called TPU 8tx and TPU 8i Max, belong to Google’s eighth generation of tensor processing units. Google introduced that generation in April this year, with more capacity for inference and support for large training requirements.
TPU 8tx is aimed at confidentially training and running AI models. Google says it builds security into the hardware layer to keep sensitive data and model weights protected and isolated while they are being used.
The chip was designed with AMD and is based on the TPU 8t. Google also says its custom inferencing network can run reasoning alongside training, enabling a model to be distilled and trained while work is underway.
Training on proprietary data
That combination could help companies that want to train models using proprietary information without moving the data between devices. The potential benefit described by Google is keeping sensitive inputs closer to where the model work happens, while still carrying out training and inference tasks.
Confidential computing is especially relevant in this scenario because the data and model weights remain important during active use, not only when they are stored. Google’s description centers on hardware-level protection and isolation during that active period.
The source does not detail how customers will deploy TPU 8tx or specify which types of confidential workloads will be supported. It does, however, position the chip for organizations that need to work with private data as they train or run AI models.
More room for large queries
TPU 8i Max addresses a different constraint: the amount of information an AI model can process during a query. It combines two inference-focused TPU 8i chips launched in April into one package.
According to the source, the combined design doubles memory capacity, processing power and bandwidth. More memory can let an AI model handle a broader range of information in a query, while the added processing power and bandwidth accompany that increased capacity.
In practical terms, the two new units extend Google’s TPU range in separate directions. TPU 8tx focuses on keeping sensitive information protected as training and inference happen; TPU 8i Max focuses on scaling the resources available to inference workloads.
A broader push into AI hardware
Google’s latest chips arrive as the company seeks to reduce its dependence on chips made by other companies for its cloud business. Developing its own AI semiconductors gives it more options for matching hardware to workloads such as confidential training or inference with higher memory needs.
Other large cloud providers and AI companies are also working on AI-specific chips and the software that accompanies them. The source names Microsoft, Amazon, OpenAI and Anthropic among those pursuing their own lines of semiconductors.
That activity reflects a wider effort across the industry to rely less on Nvidia’s dominance in AI chips. Google’s two additions illustrate how chip makers are targeting distinct needs within AI computing: protecting data during model work and accommodating larger amounts of information during inference.