Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

08 / Industry · Analysis

Google adds TPUs for more compute and data confidentiality

The reported expansion pairs higher processing performance with protection for sensitive data during AI processing.

Rows of software engineers, seen from behind, work across a vast sunlit IT services floor while a small group in the foreground maps out a business process on a whiteboard. AI-generated image

Google announced two additions to its eighth-generation tensor processing unit (TPU) family, including a product offering twice the compute performance of an existing system, according to AI Times.

The report identified the additions as TPU 8i Max and TPU 8tx. TPUs are specialised processors for AI workloads.

The report presented the expansion as addressing two requirements: processing large workloads and protecting sensitive information. That makes processing performance and confidentiality during processing separate parts of the announcement to assess.

Compute performance and memory bandwidth

According to AI Times, Google announced a product with twice the compute performance and twice the high-bandwidth memory (HBM) bandwidth of an existing system. HBM is memory designed to move large amounts of data quickly to and from a processor.

The report framed the product-family expansion as a move to compete in an AI computing market centred on Nvidia. Its account linked the performance increase to higher memory bandwidth, placing both within the effort to handle large AI workloads.

Confidentiality during data processing

According to AI Times, Google also unveiled a product designed to protect confidentiality while AI processes data. Protection during processing concerns information while it is being used for computation.

The reported line-up therefore addresses processing capacity and sensitive-information protection as distinct requirements. Assessing the expansion means considering the claimed performance increase alongside the protection offered during data processing; the performance multiple alone does not describe both parts of the announcement.