According to reports, Google is set to launch two new Tensor Processing Unit (TPU) chips, the TPU 8tx and TPU 8i Max, to meet the demands of various AI applications.
Among them, the TPU 8tx targets confidential computing needs for government agencies and regulated industries, working in conjunction with AMD chips; the TPU 8i Max is primarily aimed at AI applications requiring long context windows, boosting computational power and memory capacity by packaging two TPU chips together.
Google's chip head, Amin Vahdat, indicated that the two new products are expected to launch next year. Industry research predicts that Google's TPU-related chip, software, and hardware sales will exceed $135 billion by 2028, up from an estimated $12 billion this year.
Google states that the TPU business currently has nearly 100 customers, with its application scope extending to customers' own data centers.
Industry research estimates that TPU chip sales, excluding those leased externally via Google Cloud, will grow from approximately $12 billion this year to over $135 billion by 2028, representing an increase of more than 11-fold in two years.
However, NVIDIA still holds products with superior performance. NVIDIA's next-generation Rubin chip can deliver up to 35 petaflops (PFLOPS, quadrillions of floating-point operations per second); Google's TPU 8i offers around 10 petaflops, while the TPU 8i Max provides approximately 20 petaflops.
Google emphasizes its cost-effectiveness advantage. Industry analysis firm SemiAnalysis points out that when running AI agent tasks, Google's existing Ironwood TPU delivers nearly three times the AI processing power per dollar compared to NVIDIA's Blackwell chip; of course, if ultimate absolute speed is the priority, NVIDIA's flagship products still hold the edge.