Apple leans into Amazon's custom AI chips

Apple famously keeps its infrastructure secrets locked down, almost never letting suppliers brag about having them as a customer.

That made Tuesday's AWS re:Invent conference quite a scene, as Apple's senior director of machine learning and AI, Benoit Dupin, took the stage to publicly back Amazon's custom silicon.

Apple takes the stage at AWS

Dupin revealed that Apple has relied on AWS for over a decade to power services like Siri, Apple Maps, and Apple Music.

More specifically, Apple is already using Amazon's Inferentia and Graviton chips to handle workloads like search, yielding a massive 40% efficiency boost.

Now, Apple is eyeing Amazon's newly launched Trainium2 chips to pretrain its own proprietary models, including Apple Intelligence. Early evaluation numbers suggest pretraining efficiency gains of up to 50%.

On the same day, Amazon announced that Trainium2 is officially generally available for rent.

AWS CEO Matt Garman told CNBC that Apple was an early beta-tester for the silicon, coming to AWS with a specific request for infrastructure to build out its generative AI roadmap.

The anti-Nvidia playbook

While almost everyone else in AI is fighting over expensive Nvidia clusters, Apple is quietly taking a very different route.

Earlier this year, an Apple research paper revealed the company had used Google Cloud TPUs to train Apple Intelligence.

For actual inference, Apple pushes as much processing as possible directly to the chips inside iPhones, iPads, and Macs. When queries get too complex, they head to Apple-operated servers powered by its own M-series chips.

By adding AWS Trainium2 to the mix alongside Google TPUs, Apple is proving it can build flagship AI without writing massive checks exclusively to Nvidia.

Why it matters

We think this is one of the strongest signals yet that Big Tech can successfully break away from the Nvidia tax.

For regular users, swapping backend silicon won't change your daily Siri experience overnight. But cutting pretraining overhead by up to 50% makes running hybrid on-device and cloud models far more sustainable for Apple in the long run.

For AWS, Apple's public endorsement is pure gold. While Microsoft Azure leans on OpenAI and Google pushes its own TPUs, Amazon desperately needed a marquee customer to prove Trainium2 is ready for prime time.

Apple's pragmatic approach here looks brilliant. Instead of joining the frantic bidding war for GPUs, it pits cloud giants against each other and uses custom silicon wherever it makes financial sense. The main question now is whether Trainium2's developer ecosystem can keep up with Nvidia's established software stack as testing progresses.