Tomorrow
← Chips & Compute

AWS Trainium2

AWS · #3 most active of 18 in Chips & Compute

Compare →

Trn2 UltraServers and Project Rainier capacity, programmed through the Neuron SDK.

75

Trust score · Mostly verified

2 data points checked · 2026-11-08

Model stack: DisputedVersion: VerifiedHow this works →

Current version

Trainium2

Entry cost

Pricing is not published on the page. Available via EC2 instances/UltraServers.

Changes / 30d

8

The short answers · verified November 7, 2026

What is the latest version of AWS Trainium2?
The current shipped version of AWS Trainium2 is Trainium2, as of November 7, 2026.
How much does AWS Trainium2 cost?
AWS Trainium2 has no published flat entry price. Pricing is not published on the page. Available via EC2 instances/UltraServers.
What AI model does AWS Trainium2 use?
AWS Trainium2 runs primarily on Neuron compiler.

Source: AWS — aws.amazon.com · Reusable under CC BY 4.0 — cite Tomorrow

Capabilities

  • Trn2 UltraServer
  • 64-chip NeuronLink domain
  • Neuron SDK
  • Rainier cluster
  • Generative AI training
  • Generative AI inference
  • Up to 20.8 FP8 petaflops
  • 1.5 TB HBM3 memory
  • 3.2 Tbps EFAv3 networking
  • NeuronLink chip interconnect
  • EC2 UltraClusters scale
  • AWS Neuron SDK integration
  • Native PyTorch & JAX support
  • Trn2 UltraServers support
  • NeuronLink interconnect
  • 1.5 TB HBM3 accelerator memory
  • Up to 20.8 FP8 Petaflops of compute
  • EFAv3 networking (3.2 Tbps)
  • EC2 UltraClusters scalability
  • Up to 8 TB local NVMe storage
  • Neuron SDK integration
  • Neuron Kernel Interface (NKI)
  • 16 AWS Trainium2 chips
  • Up to 20.8 FP8 PFLOPS compute
  • 46 TBps memory bandwidth
  • NeuronLink chip-to-chip interconnect
  • AWS Neuron SDK support
  • EC2 UltraClusters support
  • 3x more energy efficient than Trn1
  • Configurable FP8 (cFP8) support
  • Stochastic rounding & 4x sparsity (16:4)
  • Local NVMe storage up to 8 TB
  • EC2 UltraClusters scaling
  • EFAv3 networking
  • Neuron SDK support
  • PyTorch support
  • JAX support
  • vLLM support
  • Generative AI Training
  • Inference
  • NeuronLink Interconnect
  • EFAv3 Networking
  • PyTorch Support
  • JAX Support
  • vLLM Support
  • Elastic Fabric Adapter (EFAv3) networking
  • NeuronLink
  • EC2 UltraClusters
  • Generative AI Inference
  • AWS Neuron SDK
  • PyTorch integration
  • JAX integration
  • generative AI training
  • generative AI inference
  • AWS Nitro System encryption
  • FP8 compute
  • HBM3 memory
  • ML Training
  • ML Inference
  • High Bandwidth Memory
  • vLLM integration
  • AI Training
  • AI Inference
  • Gen AI Training
  • AWS Nitro System
  • Elastic Fabric Adapter (EFAv3)
  • FP8 compute support
  • Up to 20.8 FP8 Petaflops (Trn2)
  • Up to 83.2 FP8 Petaflops (UltraServer)
  • Up to 6 TB HBM (UltraServer)
  • Gen AI training
  • Gen AI inference
  • 3.2 Tbps EFAv3
  • 1.5 TB HBM3
  • Hugging Face integration
  • Elastic Fabric Adapter
  • UltraServers
  • inference
  • UltraClusters
  • Elastic Fabric Adapter (EFA)
  • Model Parallelism
  • High bandwidth memory (HBM3)
  • Model parallelism
  • Encryption in transit
  • High-performance networking
  • UltraServers scaling
  • Neuron SDK Integration
  • Nitro System Encryption
  • Neuron Link
  • FP8 Compute
  • High-performance memory
  • Local NVMe storage
  • NVMe storage
  • NKI
  • FP8 support
  • FP8 Support
  • PyTorch Integration
  • JAX Integration
  • Stochastic rounding
  • BF16 support
  • PyTorch & JAX integration
  • FP32 support
  • TF32 support
  • FP16 support
  • cFP8 support
  • Stochastic Rounding

Pricing

Pricing is not published on the page. Available via EC2 instances/UltraServers.

Entry price over time

Not enough pricing history yet — we start charting from the second observation.

Geek mode

Models under the hood

  • AWS Trainium2

    AWS

    inferred

  • Llama

    Meta

    inferred

  • Llama 3.1

    Meta

    inferred

  • Llama 3.1 405B

    Meta

    inferred

  • Llama 3.3

    Meta

    inferred

  • Llama 3.3 70B

    Meta

    inferred

  • Neuron compiler

    AWS

    disclosed

  • Qwen3

    Alibaba

    inferred

  • Stable Diffusion

    Stability AI

    inferred

  • Trainium2

    AWS

    disclosed

Context window

—

Public API

yes

Multi-model routing

no

Stack signals

  • aws-nitro
  • aws-neuron-sdk
  • pytorch
  • jax

Shared model stack

Other tracked products running on the same foundation models — a quick read on how much of the catalog moves when one of these models changes.

green disclosed · amber inferred · grey unattributed. Aliases are folded into one model; provider concentration counts only models with an established vendor.

Reported scale

Reference data. Each figure is whatever the source actually said — weekly users, downloads, revenue run-rate — with its own definition and date. These are not comparable between tools and are never used to rank anything.

No public usage figure on record for this product.

Also in Chips & Compute

All Chips & Compute
Google TPU Ironwood

Google

11

Inference-first TPU pods

Version
7th generation (Ironwood)
Cost
from $5.4/unit
Model
Gemini
Cerebras WSE-3

Cerebras

9

Wafer-scale inference and training

Version
CS-4
Cost
Free trial credits available; developer tier has free promotional credits.
Model
Codex-Spark
Tenstorrent Blackhole

Tenstorrent

8

Open RISC-V AI hardware

Version
Blackhole
Cost
from $999/unit
Model
Blackhole Tensix
NVIDIA Blackwell

NVIDIA

7

GB200/B200 rack-scale AI systems

Version
GB200 NVL72
Cost
Pricing is not published (enterprise/quote-only)
Model
Blackwell
Groq LPU

Groq

4

Deterministic low-latency inference

Version
GroqCloud API
Cost
Pricing not published on current page. Custom/usage rates apply.
Model
canopylabs/orpheus-arabic-saudi
AMD Instinct MI355X

AMD

3

High-memory GPU alternative

Version
MI355X
Cost
Enterprise hardware, pricing not published
Model
CDNA

Elsewhere on the same models

Products in other segments that name one of AWS Trainium2's models in their stack.

Fireworks AI

Fireworks AI

Low-latency serving for open models and custom fine-tunes.

Version
2025
Cost
Per-token usage
Model
Llama
Meta AI

Meta

8

Llama-powered assistant across WhatsApp, Instagram, Messenger and standalone app.

Version
Muse
Cost
Free to use
Model
Llama 3.3
Together AI

Together AI

Open-model inference and fine-tuning at frontier speed.

Version
2025
Cost
Per-token, from $0.06/M
Model
Llama

Change history

  • capability

    New capabilities: Stochastic Rounding

    105 tracked106 tracked · +Stochastic Rounding

    source
  • version

    AWS Trainium2 moved to Trainium2

    Trn2Trainium2

    source
  • version

    AWS Trainium2 moved to Trn2

    Trainium2Trn2

    source
  • model

    AWS Trainium2 added AWS Trainium2 to its model stack

    Llama, Llama 3.1, Llama 3.1 405B, Llama 3.3, Llama 3.3 70B, Neuron compiler, Qwen3, Stable Diffusion, Trainium2AWS Trainium2, Llama, Llama 3.1, Llama 3.1 405B, Llama 3.3, Llama 3.3 70B, Neuron compiler, Qwen3, Stable Diffusion, Trainium2

    source
  • capability

    New capabilities: FP32 support, TF32 support, FP16 support

    101 tracked105 tracked · +FP32 support, TF32 support, FP16 support, cFP8 support

    source
  • version

    AWS Trainium2 moved to Trainium2

    Trn2Trainium2

    source
  • version

    AWS Trainium2 moved to Trn2

    Trainium2Trn2

    source
  • version

    AWS Trainium2 moved to Trainium2

    Trn2Trainium2

    source
  • version

    AWS Trainium2 moved to Trn2

    Trainium2Trn2

    source
  • capability

    New capabilities: PyTorch & JAX integration

    100 tracked101 tracked · +PyTorch & JAX integration

    source
  • version

    AWS Trainium2 moved to Trainium2

    Trn2Trainium2

    source
  • version

    AWS Trainium2 moved to Trn2

    Trainium2Trn2

    source
  • capability

    New capabilities: BF16 support

    99 tracked100 tracked · +BF16 support

    source
  • capability

    New capabilities: Stochastic rounding

    98 tracked99 tracked · +Stochastic rounding

    source
  • capability

    New capabilities: FP8 Support, PyTorch Integration, JAX Integration

    95 tracked98 tracked · +FP8 Support, PyTorch Integration, JAX Integration

    source
  • capability

    New capabilities: FP8 support

    94 tracked95 tracked · +FP8 support

    source
  • version

    AWS Trainium2 moved to Trainium2

    Trn2Trainium2

    source
  • capability

    New capabilities: NKI

    93 tracked94 tracked · +NKI

    source
  • version

    AWS Trainium2 moved to Trn2

    Trainium2Trn2

    source
  • capability

    New capabilities: NVMe storage

    92 tracked93 tracked · +NVMe storage

    source
  • capability

    New capabilities: High-performance memory, Local NVMe storage

    90 tracked92 tracked · +High-performance memory, Local NVMe storage

    source
  • version

    AWS Trainium2 moved to Trainium2

    Trn2Trainium2

    source
  • version

    AWS Trainium2 moved to Trn2

    Trainium2Trn2

    source
  • model

    AWS Trainium2 added Llama, Stable Diffusion to its model stack

    Llama 3.1, Llama 3.1 405B, Llama 3.3, Llama 3.3 70B, Neuron compiler, Qwen3, Trainium2Llama, Llama 3.1, Llama 3.1 405B, Llama 3.3, Llama 3.3 70B, Neuron compiler, Qwen3, Stable Diffusion, Trainium2

    source
  • capability

    New capabilities: Neuron Link, FP8 Compute

    88 tracked90 tracked · +Neuron Link, FP8 Compute

    source
  • version

    AWS Trainium2 moved to Trainium2

    AWS Trainium2Trainium2

    source
  • model

    AWS Trainium2 added Llama 3.1 405B, Llama 3.3 70B to its model stack

    Llama 3.1, Llama 3.3, Neuron compiler, Qwen3, Trainium2Llama 3.1, Llama 3.1 405B, Llama 3.3, Llama 3.3 70B, Neuron compiler, Qwen3, Trainium2

    source
  • capability

    New capabilities: Neuron SDK Integration, Nitro System Encryption

    86 tracked88 tracked · +Neuron SDK Integration, Nitro System Encryption

    source
  • version

    AWS Trainium2 moved to AWS Trainium2

    Trn2AWS Trainium2

    source
  • capability

    New capabilities: High-performance networking, UltraServers scaling

    84 tracked86 tracked · +High-performance networking, UltraServers scaling

    source
  • version

    AWS Trainium2 moved to Trn2

    Trainium2Trn2

    source
  • capability

    New capabilities: Model parallelism, Encryption in transit

    82 tracked84 tracked · +Model parallelism, Encryption in transit

    source
  • model

    AWS Trainium2 added Llama 3.1, Llama 3.3, Qwen3 to its model stack

    Neuron compiler, Trainium2Llama 3.1, Llama 3.3, Neuron compiler, Qwen3, Trainium2

    source
  • capability

    New capabilities: High bandwidth memory (HBM3)

    81 tracked82 tracked · +High bandwidth memory (HBM3)

    source
  • capability

    New capabilities: Model Parallelism

    80 tracked81 tracked · +Model Parallelism

    source
  • capability

    New capabilities: Elastic Fabric Adapter (EFA)

    79 tracked80 tracked · +Elastic Fabric Adapter (EFA)

    source
  • capability

    New capabilities: UltraClusters

    78 tracked79 tracked · +UltraClusters

    source
  • capability

    New capabilities: inference

    77 tracked78 tracked · +inference

    source
  • capability

    New capabilities: UltraServers

    76 tracked77 tracked · +UltraServers

    source
  • version

    AWS Trainium2 moved to Trainium2

    Trn2Trainium2

    source
Subscribe to AWS Trainium2 changes

AWS Trainium2 compared

Straight head-to-head pages against the busiest products in Chips & Compute.

AWS Trainium2 vs Google TPU IronwoodAWS Trainium2 vs Cerebras WSE-3AWS Trainium2 vs Tenstorrent BlackholeAWS Trainium2 vs NVIDIA BlackwellAWS Trainium2 vs Groq LPUAWS Trainium2 vs AMD Instinct MI355X

How to cite this page

Free to cite and reuse under CC BY 4.0. Permalink: https://tomorrow.aliensquad.ai/tools/aws-trainium

APA
Tomorrow. (2026). AWS Trainium2 — version, pricing and model stack [Data set entry]. AlienSquad. Retrieved 2026-11-08, from https://tomorrow.aliensquad.ai/tools/aws-trainium
BibTeX
@misc{tomorrow-tools-aws-trainium,
  author       = {{Tomorrow}},
  title        = {AWS Trainium2 — version, pricing and model stack},
  year         = {2026},
  publisher    = {AlienSquad},
  howpublished = {\url{https://tomorrow.aliensquad.ai/tools/aws-trainium}},
  note         = {Accessed: 2026-11-08}
}