Trainium (Trainium3) by Amazon: status and timeline

Trainium (Trainium3)

First AWS AI accelerator on 3nm. Trn3 UltraServers for large-scale generative AI training. 2.52 PFLOPS FP8 per chip, 144 GB HBM3e memory.

ScalingStage 5 of 5

GA Dec 2025. 2.52 PFLOPS FP8/chip. 362 PFLOPS UltraServer. 50% cost, 4.4x perf vs GPU.

Updated Dec 2025Checked 10 Oct0 updates this week

Milestones

No announced next step

Trainium3 announced at AWS re:InventDec 2025Complete.
Trn3 UltraServers generally availableDec 2025Complete.

Most important updates

AWS launches Trainium3 UltraServers: first 3nm AI accelerator, 144 chips per server, 362 PFLOPS FP8, 50% cost savings vs. GPUDec 2025

Current obstacles

Ecosystem maturityPyTorch and TensorFlow support for Trainium is mature but still trails NVIDIA CUDA's decades of optimization and library breadth.

Physics limits

Transistor size limitsBelow 3nm, quantum tunneling effects become significant, where electrons pass through barriers that should stop them. Physical limits approach.

How it works

3 parts

3nm process

Latest semiconductor technology

Higher density, lower power per op, higher clock speeds vs. previous generations.

HBM3e memory

High-bandwidth memory stack

Stacked memory on-chip eliminates slow external path. Critical for training where bandwidth is bottleneck.

Dense 144-chip packing

Exaflop-scale compute per server

Trn3 UltraServer integrates 144 chips for massive parallel training runs.

Spec sheet

Spec	Trainium (Trainium3)
Manufacturing process	3 nanometers^R (reported)
FP8 compute per chip	2.52 PFLOPS^R (reported)
HBM3e memory per chip	144 GB^R (reported)
Memory bandwidth per chip	4.9 TB/s^R (reported)
Max chips per UltraServer	144^R (reported)
FP8 per server	362 PFLOPS^R (reported)
Cost reduction vs. GPU	Up to 50%^R (reported)
Performance improvement vs. GPU	Up to 4.4x^R (reported)

^R reported by the company

Update log

1 update

Dec 2025

Major: AWS launches Trainium3 UltraServers: first 3nm AI accelerator, 144 chips per server, 362 PFLOPS FP8, 50% cost savings vs. GPUPressHardware

About Amazon

The team behind Trainium (Trainium3)

Amazon

Prime Air drones, AWS silicon, cloud computing, robotics

Amazon operates AWS, the world's leading cloud platform with custom AI accelerators (Trainium, Inferentia), quantum chips and servers. Amazon Robotics runs the largest robot fleet deployed. Prime Air delivers packages by drone.

AWS leads enterprise cloud with 1M+ deployed robots. Prime Air delivers to 500 US cities, hundreds of thousands of packages in 2026.
Custom silicon: Trainium/Inferentia accelerators, Graviton ARM CPUs, cat-qubit chips. Project Rainier (500k Trainium2) trains Claude.
Pieter Abbeel leads frontier model work, due re:Invent 2026. Kiro agentic IDE inverts coding: specs are source, code is artifact.

amazon.com (opens amazon.com)Company profile →

Founded: 199432 yrs
Headquarters: United States
Status: Public
Listed: AMZN (opens google.com)NASDAQ
Valuation: $2.1Tmarket capOct 2026
Staff: ~1,570,000Sep 2026
Works in: AviationDelivery dronesAIRoboticsSpaceflightQuantumComputing
Coverage: 8 programs · 7 updateslatest 1 Sep 2026checked 10 Oct
Customers: Anthropic (opens anthropic.com)GXO Logistics (opens gxo.com)
Partners: Caltech (opens caltech.edu)
Suppliers: United Launch Alliance (opens ulalaunch.com)
People: Pieter Abbeel (opens en.wikipedia.org)AI Researcher, Frontier Model ResearchAndy Jassy (opens en.wikipedia.org)CEOChris Weber (opens linkedin.com)VP of Leo Business

Spec

Trainium (Trainium3)

Manufacturing process

3 nanometers^R (reported)

FP8 compute per chip

2.52 PFLOPS^R (reported)

HBM3e memory per chip

144 GB^R (reported)

Memory bandwidth per chip

4.9 TB/s^R (reported)

Max chips per UltraServer

144^R (reported)

FP8 per server

362 PFLOPS^R (reported)

Cost reduction vs. GPU

Up to 50%^R (reported)

Performance improvement vs. GPU

Up to 4.4x^R (reported)