Google Introduces Specialized AI Chips for AI Training and Inference

In Focus

• Performance of Google’s new training TPU is 2.8x better than that of the Ironwood chip
• Google’s inference TPU offers 80% better performance
• Tech companies are pursuing custom chips to maximize efficiency

Google is changing the way its 8th-generation Tensor Processing Unit (TPU) works. For a long time, the Search giant has been producing chips that can train AI models and perform inference tasks. The tech giant has developed distinct processors for each function. Google launched the TPU AI chips at the Cloud Next event and plans to launch both chips by the end of this year.

“With the rise of AI agents, we determined the community would benefit from chips individually specialized to the needs of training and serving,” Google’s Senior VP and Chief Technologist for AI and Infrastructure, Amin Vahdat noted in a blog post.

Tech Companies Turn to Custom AI Chips

Google introduced specialized TPUs at a time when tech companies are developing custom chips to maximize efficiency. They also want to build specialized use cases. Last week, Meta expanded its partnership with Broadcom to design multiple generations of custom AI chips.

Microsoft is also developing its own chips. In January 2026, the Windows maker introduced its second-generation AI chip. In March this year, NVIDIA said its forthcoming chips would respond to user queries rapidly.

The chipmaker was counting on the AI technology developed by Groq following a licensing deal. In 2018, AWS introduced the Inferentia chip, which is designed to handle AI queries. In 2020, the cloud service provider also launched the Trainium processor, designed to train AI models. Apple, on the other hand, has been using neural engine AI components in its custom iPhone chips for years.

Specialized Chips Come with Higher Performance

Google does not compare the performance of its new chips with NVIDIA’s advanced chips. However, Google said its training TPUs’ performance is 2.8x better than that of its 7th-generation Ironwood chip launched in November. Google’s TPU inference chips also offer 80% better performance.

Google’s inference TPU relies on static random-access memory (SRAM), which NVIDIA is using in its upcoming Groq 3 LPU chip. AI chipmaker Cerebras uses this type of memory. Each TPU chip contains 384 of SRAM, which is 3x the amount in Ironwood.

Google CEO Sundar Pichai said the architecture of the chip is designed “to deliver the massive throughput and low latency needed to concurrently run millions of agents cost-effectively.”

The search giant uses its TPU chips to support its cloud service business. The giant started using its own chips to run AI models back in 2015. In 2018, Google started renting its chips to cloud clients. In September 2025, analysts projected that Google’s TPU chips along with Google DeepMind AI business could well be worth about $900 billion.

Rising Demand for Google TPUs

Demand for Google’s TPU chips is rising gradually. Citadel Securities is already using the chips to support its qualitative research software. Google said that all 17 national laboratories in the U.S. Energy Department use AI co-scientist software that runs on its chips. AI developer Anthropic has also committed to using several gigawatts of Google’s TPUs.