AWS Trn3 UltraServer
Up to 144 Trainium3 chips on an all-to-all NeuronSwitch fabric: 362 FP8 PFLOPS per system.
You're looking at a system record: an assembled machine, a server or a whole rack, with what's inside it and where it's deployed.
The Trn3 UltraServer, generally available since December 2, 2025, packs up to 144 Trainium3 chips into one scale-up domain wired through NeuronSwitch-v1, with 20.7 TB of HBM3E, 706 TB/s of aggregate memory bandwidth and 362 FP8 PFLOPS. AWS cites up to 4.4x the compute and 4x the energy efficiency of Trn2 UltraServers. UltraClusters 3.0 link thousands of them, up to one million chips. Anthropic's April 2026 agreement anticipates scaled Trainium3 capacity later in 2026, as part of nearly 1 GW of Trainium2 and Trainium3 by year end.
Specifications & key facts
Each figure is labeled by basis: vendor spec, vendor claim, reported, estimate, or official. "Not publicly disclosed" means exactly that: we never invent values.
Overview
Memory
Compute
Performance
Customers
How this connects to the rest of the stack
This record's links at a glance: click a group to jump to it.
Designed by
Contains
Available from
Successor to
Competes with
Sources
AWS EC2 Trn3 instances ↗ · Amazon: Trainium3 UltraServers now available (Dec 2025) ↗ · Anthropic: Amazon compute agreement (Apr 2026) ↗
Record reviewed as of 2026-11-16; see Methodology.