Google CEO Sundar Pichai during a meeting with French President Emmanuel Macron on the sidelines of the AI Impact Summit in New Delhi on February 19, 2026.
Ludovic Marin | AFP | getty images
After years of producing chips that can handle inference work as well as train artificial intelligence models, Google Isolating those tasks into separate processors is its latest effort. NVIDIA In AI hardware.
Google said Wednesday that it is making changes to the eighth generation of its Tensor Processing Unit, or TPU. Both chips will become available later this year.
“With the rise of AI agents, we determined that the community would benefit from chips personalized to training and service needs,” Amin Vahdat, Google’s senior vice president and chief technologist of AI and infrastructure, said in a blog post.
In March, Nvidia talked about upcoming silicon that could enable models to respond to users’ questions faster, thanks to technology acquired in a $20 billion deal with chip startup Groke. Google is a big Nvidia customer, but offers TPU as an option for companies using its cloud services.
Most of the world’s top technology companies are doing custom semiconductor development for artificial intelligence to maximize efficiency and so they can build for specific use cases. Apple Apple has included Neural Engine AI components in its in-house iPhone chips for years. Microsoft Announced second-generation AI chip in January. last week, meta Said it is working with broadcom Developing multiple versions of AI processors.
Google was at the beginning of this trend. In 2015, the company began using processors designed to run AI models, and in 2018 began renting them to cloud clients. Amazon Web Services announced the Inferentia chip in 2018 for handling AI requests, and in 2020 unveiled the Trenium processor for training AI models.
Analysts at DA Davidson estimated in September that the TPU business combined with Google’s DeepMind AI group would be worth about $900 billion.
Neither tech giant is displacing Nvidia, and Google isn’t even comparing the performance of its new chips to the AI chip leader. Google said the training chip enables 2.8x the performance of the seventh-generation Ironwood TPU announced in November at the same price, while performance for the inference processor is 80% better.
Nvidia said its upcoming Grok 3 LPU hardware will be based on large amounts of static random-access memory, or SRAM, which is used by AI chip maker Cerebras, which filed to go public earlier this month Google’s new inference chip, called TPU 8i, also relies on SRAM. Each chip has 384 MB of SRAM, three times the amount of Ironwood.
“This architecture is designed to deliver the massive throughput and low latency needed to run millions of agents simultaneously cost-effectively,” Sundar Pichai, CEO of Google parent Alphabet, wrote in a blog post.
Adoption of Google’s AI chips is accelerating. Citadel Securities has created quantitative research software that is based on Google’s TPU, and all 17 U.S. Department of Energy national laboratories use AI co-scientific software built on the chips, Google said. Anthropic has committed to using several gigawatts worth of Google TPUs.
Watch: Broadcom agrees to expanded chip deal with Google, Anthropic
