Particle.news
Download on the App Store

AWS Launches Trainium3 and Plans NVLink-Ready Trainium4, Tightening Its Grip on AI Compute

Investor focus shifts to the economics of AI hardware, from depreciation timelines to the payback on massive data‑center spend.

Overview

  • AWS rolled out Trainium3 servers for customers, with clusters of 144 chips that deliver over four times the performance of the prior generation while using about 40% less energy.
  • AWS says Trainium can cut training and operating costs by up to 50% versus comparable GPU-based systems, and named Anthropic among current users.
  • Amazon said its forthcoming Trainium4 will support Nvidia's NVLink Fusion interconnect, part of a broader multigenerational partnership that also expands access to Nvidia GPUs and software on AWS.
  • Nvidia shares pared gains following Amazon’s announcements, reflecting rising competitive pressure even as Nvidia retains leadership and reports surging data‑center revenue.
  • Analysts and investors, including Michael Burry, warn that extended server depreciation schedules and potential hardware obsolescence could strain AI capex and future profitability across hyperscalers.