💢💢💢

🚨 Amazon may be about to declare war on NVIDIA.

Amazon is exploring direct sales of its self-developed AI chips to data centers and enterprise customers, including AWS Trainium.

Amazon runs a three-chip AI architecture: Trainium for training, Inferentia for inference, and Graviton CPU for general-purpose computing.

Trainium has so far been limited to AWS infrastructure, and reported plans will expand its availability to external buyers.

This marks Amazon's shift from a cloud provider using custom chips to a vertically integrated AI computing supplier, offering cloud services and the hardware that powers them.

NVIDIA's business model relies on every company having to buy GPUs from NVIDIA regardless of which cloud service they use. If Amazon starts selling Trainium directly to data centers, companies will no longer need AWS or NVIDIA.

They can buy Amazon's chips and run them on their own infrastructure. This would completely cut NVIDIA out of the deal.

Amazon's Trainium AI accelerator has already attracted major clients including OpenAI, Anthropic, and Uber, all of whom access it via AWS.

Anthropic has committed to using up to 5 gigawatts of current and future Trainium capacity.

OpenAI has agreed to use about 2 gigawatts. These are workloads that would likely have run on NVIDIA GPUs.

Amazon said in April that the Trainium chip has generated over $225 billion in revenue commitments. Trainium2 is largely sold out.

Trainium3 begins shipping in early 2026 and is almost fully booked. Trainium4 is about 18 months away from broad availability but is mostly booked.

This level of demand indicates that real customers, even when limited to AWS, are choosing Trainium over NVIDIA GPUs.

Andy Jassy said in his April shareholder letter that the chip business (including Trainium, Graviton, and Nitro) has annualized revenue of approximately $20 billion, growing at a triple-digit year-over-year rate.

He said that if the chip business were sold externally like NVIDIA's, that alone would represent a $50 billion annualized revenue business.

AWS executive Peter DeSantis, in charge of chip efforts, confirmed these talks, citing the need for local control over AI computing, especially Europe's push for sovereign infrastructure. He dismissed concerns about cannibalizing AWS cloud revenue, pointing to generally low AI compute utilization across the industry.

This follows Alphabet's move in April, when CEO Sundar Pichai said Google would begin offering its TPUs to external data centers through a limited customer program.

Two of the world's three largest cloud providers are now acting independently to sell their own AI chips beyond their own cloud walls. This is the part that should worry NVIDIA the most.

This is not one competitor. It's the entire hyperscale cloud layer simultaneously deciding to no longer rely exclusively on NVIDIA as a customer.

NVIDIA remains the dominant force today.

Data center revenue grew 92% year-over-year to a record $75.2 billion, and Amazon itself remains one of NVIDIA's largest customers, committed to deploying over 1 million NVIDIA GPUs starting in 2026.

The world's two largest cloud companies are now actively building infrastructure to compete directly with NVIDIA, using the very customers NVIDIA currently relies on.