Amazon just proved the fastest way to build a $25 billion business is to give away the picks and shovels during a gold rush.
The Summary
- Amazon's AI business hit a $25 billion annual run rate, growing faster than AWS did in its early years despite massive infrastructure costs
- Trainium3 chips now have native PyTorch support, removing the friction that kept AI researchers locked into Nvidia's ecosystem
- Amazon is betting that owning the infrastructure layer for AI agents matters more than short-term profitability
The Signal
Amazon's AI infrastructure business crossed $25 billion in annual revenue, a milestone that took AWS years longer to reach. The speed matters because it signals where the compute economy is heading: not toward centralized models, but toward distributed agent workloads that need elastic infrastructure.
The business grew this fast while bleeding money on data center buildouts and chip development. Amazon isn't optimizing for margins yet. They're optimizing for default status, the same playbook that made AWS the infrastructure layer every startup builds on without thinking.
"Amazon's rapid AI growth highlights the strategic importance of AI infrastructure investment, despite financial pressures from rising costs."
The native PyTorch support for Trainium3 is the tactical move that makes the strategy real. Before this, researchers had two choices:
- Use Nvidia chips with PyTorch (easy, expensive, good luck getting allocation)
- Use Amazon's custom silicon and rewrite everything (cheap, painful, nobody did it)
- Wait for Google's TPUs to maybe support your framework next year
Now there's a third path. PyTorch is the de facto standard for AI research. If you can run PyTorch natively on Trainium without translation layers or compatibility hacks, Amazon just removed the moat around Nvidia's researcher ecosystem. Every PhD student who learns to train models on Trainium instead of H100s is a future enterprise buyer who defaults to AWS.
This isn't about competing with OpenAI or Anthropic on foundation models. Amazon doesn't need to win the model race. They need to be the infrastructure that everyone uses to deploy agents at scale. The $25 billion run rate suggests they're already there.
The Implication
Watch where the AI researchers go in the next six months. If PyTorch on Trainium actually works without the usual cloud provider jank, Amazon will own the training infrastructure for the next generation of agent frameworks. That matters because agents are inference-heavy and training-light compared to foundation models. Cheaper chips that run the standard tools win.
For anyone building agent companies: your infrastructure costs just became negotiable again. Nvidia's pricing power only holds if there's no alternative that works with your existing codebase. Amazon just shipped that alternative.