Google has unveiled two new artificial intelligence chip systems, the TPU 8tx and the TPU 8i Max, in its boldest move yet to challenge Nvidia's dominance of the booming market for AI hardware. The announcement, made on Monday, 9 November 2026, completes a record year for Google's custom chip programme, with four Tensor Processing Unit systems now launched in a single calendar year.
The two new systems are specialised versions of Google's existing eighth-generation TPUs rather than entirely new chips, the company confirmed. Bloomberg News first reported the rollout before Google published full details in an official blog post written by Amin Vahdat, the company's senior vice president and chief technologist for AI and infrastructure.
The TPU 8tx is built for confidential computing, a technology that keeps sensitive data protected even while it is being processed. It uses the same silicon as the TPU 8t, announced in April, but pairs it with fifth-generation AMD EPYC host processors and the Boardfly networking system from the TPU 8i. Google says it delivers up to twice the lower latency of the TPU 8t at what it calls near-zero performance cost, making secure AI commercially viable for the first time at this scale.
The system is aimed squarely at government agencies and companies in regulated industries, such as banks, insurers and healthcare providers, that need to run AI on confidential data without exposing it. Google describes the 8tx as offering hardware-level confidential computing for private fine-tuning and secure agent execution, a pitch designed to win over customers that have so far been wary of sending sensitive workloads to shared cloud infrastructure.
The TPU 8i Max, by contrast, is built for AI applications that must process enormous amounts of information at once. It packages two TPU 8i chips together in a single unit, doubling the compute power, the on-chip memory and the memory bandwidth of a single 8i chip. Google says the system offers 768 megabytes of on-chip SRAM and 17.2 terabytes per second of high-bandwidth memory, all within the existing rack footprint.
That capacity is aimed at the new generation of AI agents that reason across huge contexts, such as entire software codebases, long documents or hours of video. As AI models move from answering single questions to carrying out long, multi-step tasks, the amount of context they must hold in memory at once has become one of the main bottlenecks in data centres.
The launch marks the first time Google has introduced four TPU systems in one year, Vahdat wrote in the announcement. The TPU 8t, built for training large AI models, and the TPU 8i, built for fast, low-latency inference, were introduced at Google Cloud Next in April 2026. The two new systems complete the eighth-generation lineup.
Neither system is available to customers yet. Google has not published a price or full specifications for either one, and Vahdat told Bloomberg that both will arrive in 2027. All of the performance figures released so far are Google's own claims, and no independent benchmark of either system has been published.
The announcement landed on a dramatic day for the chip market. Alphabet shares rose in pre-market trading after the news, while Nvidia shares fell on the first trading day of a board investigation into its chief executive, Jensen Huang, over reports of a relationship with a suspected Chinese spy. Shares in AMD and Broadcom, meanwhile, rallied, with AMD benefiting directly from its role supplying host processors for the TPU 8tx.
Google's TPUs were originally built for its own products, including Search, YouTube and its Gemini chatbot, but the company is now selling them aggressively to outside customers. Google says its TPUs now have nearly 100 customers, from frontier AI laboratories to trading firms, retailers and universities, and Bloomberg Intelligence estimates that Google's direct TPU hardware and software sales could top 135 billion dollars by 2028, up from an estimated 12 billion dollars in 2026, with TPUs priced about 30 per cent below Nvidia's graphics processors.
The stakes are enormous. Nvidia has dominated the market for the chips that train and run AI systems, but Google, AMD and a string of start-ups are all racing to offer cheaper or more specialised alternatives as data centre spending soars. Whether Google's new systems can turn that ambition into a serious dent in Nvidia's market share will depend on what customers can actually buy in 2027, at what price, and how the chips perform outside Google's own tests. For now, the message from Mountain View is clear: the AI chip race is no longer a one-horse contest.
