What is a custom AI chip (ASIC)?

How it works

JEDEC, a microelectronics standards body, defines an application-specific integrated circuit (ASIC) as an integrated circuit made for one particular application or function and for one customer. A custom AI chip applies that idea to machine learning.

Google's Tensor Processing Unit (TPU) is a long-running example. Google describes TPUs as custom-designed ASICs optimized for the matrix multiplication at the core of neural networks. In a 2017 paper, Google engineers reported the chip had been deployed in data centers since 2015. They reported it ran trained neural networks about 15 to 30 times faster on average than a contemporary GPU or CPU, with 30 to 80 times the performance per watt.

Some companies build these chips with a partner. In October 2025, OpenAI and Broadcom announced a plan for 10 gigawatts of OpenAI-designed accelerators, developed and deployed with Broadcom. The companies targeted rack deployments starting in the second half of 2026 and finishing by the end of 2029.

Why it matters

Companies that design their own chips cite cost and efficiency on their own workloads. In January 2026, Microsoft said Maia 200 delivers 30% better performance per dollar than the latest hardware in its fleet. Microsoft calls it a chip for inference (running trained models) built on TSMC's 3-nanometer process. Amazon Web Services says its Trainium chips are designed to lower the cost of AI training and inference. These are company claims, not independent benchmarks.

AI developers also mix custom chips with processors sold by chip vendors. Anthropic said in April 2026 it trains and runs its models on AWS Trainium, Google TPUs and Nvidia GPUs. That lets it match workloads to suitable chips, it said.

Where things stand in 2026

In April 2026, Google described its eighth-generation TPU as two separate systems: TPU 8t for large-scale pre-training and TPU 8i for post-training and inference. As of October 2026, Google Cloud listed both as coming soon and the seventh-generation Ironwood as generally available. Also in April, Anthropic announced an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, which it expects to come online starting in 2027.

For the quarter ended August 2, 2026, Broadcom reported AI semiconductor revenue of $16.7 billion, up 221% from a year earlier. It forecast $21.7 billion for the following quarter. Chief executive Hock Tan said demand for the company's custom AI accelerators and networking remained very strong.