AWS re:Invent 2025: Trainium3, Kiro, and the AI Agent Era
Amazon announced a beast of an AI chip, a coding agent that works for days, and new Nova models. Here's the complete breakdown.
Amazon's AI Blitz
AWS re:Invent 2025 delivered major announcements. Let's break them down.
Trainium3: The Numbers Are Insane
Amazon's new Trainium3 UltraServer packs up to 144 custom chips:
- 4.4x more compute than Trainium2
- 4x efficiency improvement
- 4x memory bandwidth
This directly challenges:
- Google's 7th-gen Ironwood chip
- Nvidia's Blackwell Ultra
The message is clear: Amazon is serious about silicon.
Kiro: The AI Agent That Codes for Days
The most exciting announcement: Kiro, an AI coding agent that can:
- Work autonomously for multiple days
- Handle complex, multi-step tasks
- Integrate with existing dev workflows
This is different from Copilot or Claude. Kiro is designed for:
"Here's a feature spec. Build it."
↓
*Kiro works for 3 days*
↓
"Here's your PR, ready for review."
Nova Models: Amazon's Foundation
New Nova models across the spectrum:
- Nova Pro - High capability
- Nova Lite - Cost efficient
- Nova Canvas - Image generation
- Nova Reel - Video generation
Key differentiator: customization. Amazon is betting enterprises want control over their models.
On-Premises AI
In a surprising move, Amazon is offering on-prem Nvidia infrastructure.
A cloud company going on-prem? That's how important AI workloads have become.
Target customers:
- Regulated industries (finance, healthcare)
- Data sovereignty requirements
- Ultra-low latency needs
The Strategy
Amazon's AI strategy is becoming clear:
- Build the silicon (Trainium, Inferentia)
- Build the models (Nova)
- Build the agents (Kiro)
- Own the infrastructure (cloud + on-prem)
What Developers Should Do
If you're on AWS:
- Try Kiro for complex automation
- Benchmark Nova against your current models
- Watch Trainium3 availability for training workloads
Amazon isn't just competing. They're building an entire AI stack.