Amazon just turned its internal AI chip shop into a $20 billion business, and now it's thinking about selling to anyone who'll buy.
The Summary
- Amazon's chip division is on track to generate over $20 billion annually, and CEO Andy Jassy says they're considering selling these chips to other companies
- Uber expanded its AWS contract to run more ride-sharing features on Amazon's chips, choosing Amazon over Oracle and Google
- Amazon could flip the AI chip market by becoming both the infrastructure provider AND the chip vendor, competing directly with Nvidia on silicon while still being Nvidia's biggest customer
The Signal
Amazon didn't set out to become a chip company. They built Graviton, Trainium, and Inferentia chips because buying from Nvidia was too expensive and too limiting. Classic vertical integration play. But somewhere between "we need cheaper chips for our own cloud" and "holy shit, this is $20 billion in annual revenue," Amazon realized they'd accidentally built a chip business.
The Uber deal is the tell. Uber chose Amazon's chips over Google's TPUs and whatever Oracle was pitching. That's not just about price. It's about the whole stack. When you're already running on AWS, using AWS chips means one vendor relationship, one optimization path, one throat to choke when things break. The switching costs aren't just financial, they're cognitive and operational.
Now Jassy is floating the idea of selling these chips outside the AWS walled garden. That's a different game entirely. If Amazon starts selling Trainium or Inferentia chips to anyone, they're competing with Nvidia on silicon while simultaneously being Nvidia's largest cloud customer. They're also competing with their own cloud customers who might want to build on-prem infrastructure with Amazon chips but not Amazon cloud services.
The $20 billion figure matters because it's real revenue, not projected TAM. Amazon chips are already powering enough workloads to hit hyperscale economics. That's the foundation you need before you go external. You don't sell chips at scale until you've proven them at scale on your own infrastructure.
The Implication
Watch what Amazon does with pricing and availability. If they actually open-source chip designs or make them available through traditional distributors, that's a nuclear option against Nvidia's margins. More likely, they'll keep it tight: AWS-first, select partners second, maybe never truly "open" at all. For companies building AI infrastructure, this creates a new calculation. Do you lock into Nvidia's ecosystem, bet on Google's TPUs, or go all-in on Amazon's stack knowing the chips and cloud are optimized together? The bundling game just got more sophisticated.
Sources: Bloomberg Tech | TechCrunch AI