AWS Neuron

license

mixed/proprietary-core

source

partial(aws-neuron-sdk is a documentation and issue-tracker repo carrying no Runtime or Compiler source, while the kernel driver, NKI kernel library, vLLM-for-Neuron and NKI samples do ship as source)

runtime

proprietary(Neuron Runtime+Compiler not open sourced)

open-glue

Linux kernel driver(GPL-2.0),vLLM-Neuron integration(Apache-2.0,vllm-project org),NKI library,samples(MIT-0)

hardware-lock

AWS-Inferentia/Trainium-only

Source-available rather than closed or open. Neuron is not a managed service, since the SDK installs and runs on the customer's own EC2 instance, but it installs from proprietary packages and the headline inference path is not published. The aws-neuron-sdk repository is a documentation and issue-tracker repo whose compiler and neuron-runtime directories hold documentation rather than code, and AWS's own open-source catalog lists six public repositories with neither the Neuron Runtime nor the Neuron Compiler among them. What does ship as source is the GPL-2.0 Linux kernel driver, the NKI kernel library, the Apache-2.0 vLLM-for-Neuron integration and the MIT-0 samples, real open components arranged around a closed core.

• https://awsdocs-neuron.readthedocs-hosted.com/en/latest/general/oss/index.html recorded 2026-08-11

  AWS Neuron provides open source code and samples for some of its components, libraries, and tools under the Apache 2.0 license, and lists the public repositories open to contribution. The URL redirects to /about-neuron/oss/index.html, where the catalog lists exactly six public repositories -- TorchNeuron PyTorch Extension, Neuron Agentic Development, Neuron Explorer CDK, Neuron Kernel Library, vLLM for Neuron, NKI Samples -- and neither the Neuron Runtime nor the Neuron Compiler appears among them.

• https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/what-is-neuron.html recorded 2026-08-09

  Neuron Compiler and Neuron Runtime entries in the 'AWS Neuron Core Components' list carry no 'Open Source' bullet, unlike the Native PyTorch entry ('Open Source : Released as an open source package on GitHub under Apache 2.0') and the vLLM entry ('Open Source : Source code released under the vLLM project organization with source code on GitHub').

• https://github.com/aws-neuron/aws-neuron-sdk recorded 2026-08-11

  "path":"LICENSE-DOCUMENTATION","preferredFileType":"license","tabName":"License" alongside "path":"LICENSE-SAMPLECODE",..."tabName":"MIT-0" -- repo carries a generic, GitHub-unclassified license for docs and a separate MIT-0 license for sample code, i.e. no single OSI license covers the SDK repo. The repo's top-level directories are .github, compiler, frameworks, deploy, general, libraries, neuron-runtime, neuron-customops, nki, release-notes, setup and tools -- the "compiler" and "neuron-runtime" directories hold documentation for those components, not their source. The repo points readers at the hosted Neuron SDK documentation site and carries an open issue tracker.

• https://github.com/aws-neuron/aws-neuron-driver recorded 2026-08-09

  "license":{"spdxId":"GPL-2.0","name":"GNU General Public License v2.0"}

• https://raw.githubusercontent.com/aws-neuron/aws-neuron-driver/master/LICENSE recorded 2026-08-09

  GNU GENERAL PUBLIC LICENSE Version 2, June 1991

• https://github.com/vllm-project/vllm-neuron recorded 2026-08-09

  "license":{"spdxId":"Apache-2.0","name":"Apache License 2.0"}

AWS publishes no customer count, download count or usage figure for Neuron anywhere on the product page, names no customers and makes no quantified scale claim. What the page does establish is integration breadth: PyTorch and JAX natively, plus Hugging Face Transformers and Optimum Neuron, vLLM, PyTorch Lightning and TorchTitan, "without modifying your code". The band is therefore qualitative and only as good as that description, since nothing here is countable: the breadth can support a level but cannot confirm that 3 is precisely the right one.

• https://aws.amazon.com/ai/machine-learning/neuron/ recorded 2026-08-13

  PyTorch and JAX natively, plus HuggingFace, vLLM, PyTorch Lightning, TorchTitan and Optimum Neuron, "without modifying your code". No customer count, download count, usage figure or named customer anywhere on the page - the absence is what the reported_traction instrument rests on.

The score rests on the feature set: the Neuron runtime supports the vLLM V1 API on Trainium and Inferentia with expert parallelism for mixture-of-experts models, disaggregated inference architectures and speculative decoding, alongside native PyTorch and JAX and a compiler/runtime/profiler stack. That is a server-class inference feature set, but one locked to AWS Inferentia and Trainium hardware. No MLPerf Inference figure has surfaced for Neuron, so there is no benchmark result behind the score.

• https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/what-is-neuron.html recorded 2026-08-09

  The runtime supports Expert Parallelism for MoE models, disaggregated inference architectures, and speculative decoding.