MW Google debuts two new custom chips in latest bid to challenge Nvidia's dominance

By Britney Nguyen

The new TPUs offer cost advantages and improved storage functions

Google announced its eighth generation of TPUs on Wednesday.

One of Google's big artificial-intelligence advantages is its custom chip business, and the company is stepping up the competitiveness of its products in a potential challenge to Nvidia.

Google on Wednesday introduced the eighth generation of its tensor processing units, with two distinct chips focused on AI training and inference. The TPU 8t for training and TPU 8i for inference have been used to develop AI agents and to run massive inference workloads, Google said in a blog post. Inference is the process of running AI models.

Google has co-developed its TPUs with Broadcom $(AVGO)$ for more than a decade, and the chips have been used to train its Gemini AI models, as well as to power its YouTube and search engine algorithms. The newest chips were designed alongside Google DeepMind, which is Alphabet's $(GOOGL)$ $(GOOG)$ AI research lab.

See more: Broadcom's stock is rising. Here's why its new Google and Anthropic deals are so significant.

The current era of agentic AI, where an AI model can complete tasks with little to no human prompting, requires "a new set of demands on infrastructure," Google said. The new chips were designed to handle larger, more complex workloads, and to be able to adapt to the evolving capabilities of AI models, the company added.

There's been an increased focus on inference as adoption of AI applications grows. Nvidia (NVDA) has dominated the market for AI training and is working to improve its inference capabilities, in part through a nonexclusive licensing agreement with inference-chip maker Groq in December. But it remains to be seen whether the company will have its same advantages in the inference era.

"Several years ago, we anticipated rising demand for inference from customers as frontier AI models are deployed in production and at scale," Google said.

Because agentic AI works "in continuous loops of reasoning, planning, execution and learning," Google said it determined the industry would benefit from separate, specialized chips. Therefore, TPU 8t and TPU 8i were built specifically to support agentic AI, according to the company.

Google said it redesigned the stack for TPU 8i "to eliminate the 'waiting room' effect" during inference, which refers to delays in responses that can happen when AI model usage is high.

With the improvements, Google said the inference chip's performance-per-dollar compared to the previous Ironwood TPU is 80% better, meaning users can meet nearly twice the demand for the same cost.

The training TPU aims to cut down the time it takes to develop frontier AI models from months to weeks, Google said. Now, a single TPU 8t superpod, or chip cluster, can scale to 9,600 chips and two petabytes of high-bandwidth memory - double the interchip bandwidth of Ironwood. TPU 8t can also access storage faster, the company said.

The latest TPUs are "the latest expression of our co-design philosophy," Google said, meaning its hardware and software are designed to work together to optimize performance. As part of that, Google said both chips are running on its Axion central-processing-unit host for the first time, allowing for efficiency on a system level, rather than just on a chip level. A CPU host helps an AI processor stay fully utilized by helping to manage data and tasks.

Additionally, by integrating power management into its TPUs to adjust how much power is being drawn based on demand, Google said the new chips can provide up to two times better performance-per-watt than Ironwood.

Google said both TPUs will be generally available to its cloud customers later this year.

Unlike Nvidia, Google does not sell its TPUs to external customers, though AI startup Anthropic uses the chips.

Don't miss: Google's AI lead is growing in this key area. That's good news for Alphabet's stock.

That said, Google is still a significant Nvidia customer. The company also said Wednesday that it will be among the first cloud providers to offer Nvidia's upcoming Vera Rubin AI platform later this year.

-Britney Nguyen

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

(END) Dow Jones Newswires

April 22, 2026 08:07 ET (12:07 GMT)

Copyright (c) 2026 Dow Jones & Company, Inc.