Most AI models today are trained on NVIDIA GPUs. Yet the four largest buyers of those GPUs — Google, Amazon, Microsoft and Meta — all design chips of their own. The reasons are mostly economic.

The chips

  • Google TPU. Google announced its Tensor Processing Unit in 2016 after running it internally in its data centers. TPUs train and serve Google’s Gemini models and are rented to customers through Google Cloud.
  • AWS Trainium and Inferentia. Amazon’s chips target training (Trainium) and inference (Inferentia). AWS offers them as cheaper alternatives to GPU instances.
  • Microsoft Maia. Microsoft unveiled its Maia 100 AI accelerator in November 2023 for Azure workloads.
  • Meta MTIA. Meta’s Training and Inference Accelerator is designed first for the recommendation models behind its feeds and ads.

Why build instead of buy

  1. Cost. At the scale of hundreds of thousands of chips, even modest efficiency gains save billions.
  2. Supply. GPU shortages in 2023 and 2024 showed the risk of depending on one supplier.
  3. Fit. A chip designed for a company’s own models and software stack can skip features it does not need.
  4. Negotiating power. A credible in-house alternative strengthens a buyer’s position with any vendor.

Why NVIDIA still leads

Custom chips are hard to program well. NVIDIA’s advantage is not just hardware but CUDA, the software ecosystem that most AI research is written for, plus fast networking that ties thousands of GPUs together. Custom silicon tends to work best for a company’s own, well-understood workloads, while GPUs remain the default for everything else.

The takeaway

Expect a mixed world: GPUs for flexibility and frontier research, custom accelerators for high-volume inference inside the largest platforms.

A short history of custom AI chips

Google announced its Tensor Processing Unit (TPU) in 2016, revealing it had been using the chips in its data centers to accelerate neural networks. Since then, TPUs have gone through multiple generations and power Google’s own models as well as cloud customers. Amazon developed Trainium for training and Inferentia for inference; Microsoft announced its Maia AI accelerator; and Meta developed its MTIA chips for recommendation and other workloads. Apple’s Neural Engine accelerates AI on iPhones and Macs.

How custom chips are made

Big technology companies design chips but generally don’t manufacture them. They work with partners for parts of the design and rely on foundries — primarily TSMC for leading-edge chips — for fabrication, plus suppliers of high-bandwidth memory and advanced packaging. This means custom chips face the same supply constraints as other accelerators.

The economics

Designing a chip costs hundreds of millions of dollars and takes years, so it only makes sense at enormous scale. For the largest cloud and AI companies, even modest savings per chip multiply across hundreds of thousands of units, and controlling the hardware lets them optimize it for their own models and software. Smaller companies are better served buying from chip vendors or renting cloud capacity.

What it means for customers

For cloud customers, custom chips add choice: some workloads run cheaper on a provider’s own accelerators, though porting code may require effort. Test performance and cost on your own models before committing, and prefer software frameworks that support several hardware types.

Related: GPU vs TPU vs NPU and why NVIDIA dominates AI.

What to watch next

  • New generations of TPUs, Trainium, Maia and MTIA.
  • The share of AI workloads running on custom chips versus GPUs.
  • Software tools that make models portable across chip types.
  • Foundry and memory capacity for advanced AI chips.

Key takeaways

  • Google, Amazon, Microsoft, Meta and Apple design their own AI chips.
  • Custom chips can reduce costs, improve efficiency and lower dependence on suppliers.
  • They are manufactured by foundries such as TSMC and face shared supply constraints.
  • NVIDIA remains dominant thanks to performance and its software ecosystem.
  • Custom silicon only makes economic sense at very large scale.

Frequently asked questions

What is a TPU?

A Tensor Processing Unit, Google’s custom chip designed to accelerate machine learning workloads.

Will custom chips replace NVIDIA GPUs?

They are taking some workloads, especially inside the companies that build them, but GPUs remain widely used and preferred for flexibility.

Can I use these chips?

Some are available through cloud services, such as Google Cloud TPUs and AWS Trainium and Inferentia instances.