Amazon $AMZN ( ▲ 0.87% ) kicked off re:Invent in Las Vegas with a stack of announcements that show exactly how AWS plans to stay on top of the AI infrastructure world. With enterprise demand for compute exploding, AWS made it clear it wants to power both the training and the day-to-day running of AI models for everyone from scrappy startups to giants like Anthropic.
The new Trainium3 chip
The biggest reveal was Trainium3. AWS says the new chip can train models four times faster and at half the cost of the previous generation. It’s also built to handle inference, which is the part where models actually run in the real world. That puts Trainium directly in competition with Nvidia $NVDA ( ▲ 3.74% ) , and the timing isn’t subtle. As more companies look to diversify away from GPUs, AWS is offering a cheaper route with performance that’s good enough for a lot of workloads.
Work on Trainium4 is already underway, and AWS confirmed it will support Nvidia’s NVLink Fusion interconnect, which signals that customers will likely mix and match both chips inside the same clusters.
A new wave of Nova models
AWS also refreshed its in-house model family, Nova, which now comes in four flavors: Nova Lite for everyday tasks, Nova Pro for heavier reasoning, Nova Sonic for speech, and Nova Omni, the all-in-one multimodal option that handles text, image generation, and more. Cheaper and more efficient Nova models let AWS capture customers who want high-quality AI without paying frontier-model prices.
Training made easier with Nova Forge
To help companies build their own custom models, AWS rolled out Nova Forge. It lets customers take the Nova family and train variants using their own proprietary datasets, essentially creating specialized expert models without needing huge GPU clusters or massive ML teams.
AI Factories for enterprises
AWS also introduced AI Factories, which allow customers to run AWS infrastructure inside their own data centers. It’s designed for organizations that want cloud-like scalability but need tight control over their data. These factories support both Trainium and Nvidia GPUs.
A big push into AI agents
The final theme was agents. AWS introduced Nova Act, which lets companies spin up fleets of automated AI helpers for production workflows, and a new DevOps Agent that monitors software systems and fixes issues on its own when engineers are offline. AWS called it an autonomous on-call engineer.
Overall, AWS showed it wants to own the full AI stack from chips to models to agents, all while giving customers cheaper ways to escape runaway GPU costs.