Amazon’s AI Numbers: Trainium Hits Multi-Billion Run Rate, Nova 2 Matches Frontier Models

At re:Invent 2025, Jassy revealed Trainium’s scale: a multi-billion dollar revenue run-rate business, 1 million+ chips in production, 100,000+ companies using it, and 30-40% cost savings vs. NVIDIA GPUs. These aren’t research metrics – they’re infrastructure — as explored in the economics of AI compute infrastructure — dominance.

Trainium’s Traction

The numbers tell a story of scale that few appreciate:

– Multi-billion dollar revenue run-rate
– 1 million+ chips in production
– 100,000+ companies using it as majority of Bedrock usage
– 30-40% cost savings vs. NVIDIA GPUs

Trainium3 (launched at re:Invent 2025) delivers the next performance jump. Critical signal: Trainium4 is already in development with support for NVIDIA NVLink Fusion – meaning Amazon’s next-gen chips will interoperate with NVIDIA GPUs. This hybrid approach lets customers hedge while Amazon builds switching costs.

Nova 2: The Model Portfolio

The Nova 2 family launched two weeks before the reorg:

Nova 2 Lite: Fast, cost-effective reasoning. Equal or better on 13/15 benchmarks vs. Claude Haiku 4.5

Nova 2 Pro: Amazon’s most intelligent model. Matches or beats Gemini 2.5 Pro on 15/19 benchmarks

Nova 2 Omni: Industry-first unified multimodal – processes and generates text and images simultaneously. 750K words, hours of audio, 200+ languages

Nova Forge: “Open training” service – companies can build custom models by mixing proprietary data with Nova checkpoints

The Economics

AWS Trainium offers 40% cost reduction compared to competitors, creating a powerful economic moat. When your AI training costs 40% less than your competitors’, you can afford to be patient while others burn through capital.

Key Takeaway

These aren’t announcement metrics – they’re production scale. As AI data center analysis shows, the companies winning on infrastructure economics will win the next AI phase.

Source: Amazon’s AI Superstructure on The Business Engineer