Browse tools

Chips, compute and data centres

AI accelerator

In short

An AI accelerator is any computer chip designed to speed up AI work, mainly the huge number of multiplications inside neural networks. GPUs, Google's TPUs, Amazon's Trainium and the NPUs in phones are all AI accelerators.

Also called
AI chip, Accelerator chip, Accelerated compute
Topic
Chips, compute and data centres
Level
Beginner
First described
The idea has no single inventor. A landmark case is Google's TPU, in its data centres from 2015 and described by Jouppi and colleagues in 20176.

An ordinary processor can run AI, but slowly. Accelerators are chips built to do the same maths much faster and with less energy. The UK competition regulator defines accelerator chips as high-performance chips, such as GPUs and application-specific integrated circuits (ASICs), used to speed up computation1.

There are two broad families. GPUs are flexible and can run a wide range of software. ASICs are hardwired for a specific job1. Google's TPU is an ASIC2, and so is Amazon's Trainium, which AWS describes as a purpose-built AI chip with 144 GB of high-bandwidth memory per chip in its third generation3.

Accelerators come in very different sizes. A single Nvidia H100 data centre GPU runs at up to 700 watts, and Meta trained Llama 3 on up to 16,000 of them at once7. At the other end, the neural processing unit in a laptop or phone runs small models on the device itself4.

The chip is only part of it. Accelerators need fast memory, fast links between chips, and software that turns a model into instructions the hardware understands. For TPUs that job falls to a compiler called XLA2.

In the cloud, renting time on these chips is often called "accelerated compute"1.

Example

Anthropic said in October 2025 that it spreads its computing across three accelerator platforms: Google's TPUs, Amazon's Trainium and Nvidia's GPUs5.

Why it matters

The supply and cost of AI accelerators sets how fast AI companies can grow. It is also why governments now regulate who can buy the most powerful ones.

Questions people ask

Is a GPU an AI accelerator?

Yes. The UK competition regulator lists GPUs alongside custom chips (ASICs) as accelerator chips1. GPUs were built for graphics but are now the most widely used AI accelerator.

What is the difference between an AI accelerator and an NPU?

An NPU is one kind of AI accelerator, built into a phone or laptop chip to run AI on the device with low power4. The term AI accelerator also covers large data centre chips like TPUs and Trainium.

Who makes AI accelerators?

Nvidia is the best-known GPU maker5. Google designs TPUs2 and Amazon designs Trainium3. Qualcomm, Intel and AMD build NPUs into the chips in Windows laptops4.

Related terms

GPU TPU NPU Inference chip High-bandwidth memory Semiconductor Compute

Sources

  1. Competition and Markets Authority (2025), Cloud services market investigation, final decision glossary
  2. Google Cloud, Introduction to Cloud TPU
  3. AWS, Trainium product page
  4. Microsoft Learn, Copilot+ PCs developer guide (updated August 2026)
  5. Anthropic (23 October 2025), Expanding our use of Google Cloud TPUs and services
  6. Jouppi et al. (2017), In-Datacenter Performance Analysis of a Tensor Processing Unit, arXiv:1704.04760
  7. Grattafiori et al. (2024), The Llama 3 Herd of Models, arXiv:2407.21783, section 3.3.1

Back to the glossary