At AWS re:Invent 2017, AWS launched Amazon SageMaker, a managed machine learning platform uniting hosted Jupyter authoring notebooks, containerized multi-node Distributed Training, hyperparameter tuning, and autoscaling inference endpoints into a single cloud workflow.
Amazon & AWS AI
Also known as: AWS · Amazon Web Services · Amazon Bedrock · Amazon Nova · Amazon SageMaker · AWS Trainium
Plain English
Amazon's cloud AI infrastructure, model platform, and first-party foundation model organization—spanning Amazon SageMaker, custom AWS Inferentia and Trainium silicon, the managed Amazon Bedrock multi-model API and agent platform, and the first-party Amazon Nova multimodal, image, speech, and video model family.
Technical Overview
AWS operates across all layers of the foundation model stack: custom AWS Trainium and Inferentia accelerators (programmed via the AWS Neuron SDK and SGLang-Neuron / vLLM) power large-scale Distributed Training clusters including Anthropic's Project Rainier; Amazon SageMaker and Amazon Bedrock provide managed training, RAG, Guardrails & AI Safety, and AgentCore agentic orchestration across third-party and open-weight families; and Amazon's AGI organization trains the first-party Amazon Nova model lineup across text, vision, speech-to-speech, image, and video generation.
Timeline (6)
2017
Release2022
ReleaseFollowing the 2019 rollout of its AWS Inferentia inference ASIC, AWS made Amazon EC2 Trn1 instances powered by custom AWS Trainium silicon generally available—packing up to 16 Trainium accelerators with 512 GB of shared HBM and 800 Gbps EFA networking per node to provide a purpose-built cloud alternative to NVIDIA GPUs for Pre-Training and fine-tuning.
2023
EcosystemAmazon and Anthropic announced a strategic collaboration. Anthropic designated AWS its primary cloud provider and agreed to use Trainium and Inferentia for future model workloads, while AWS customers would gain access to its models through Bedrock.
2024
ReleaseAt AWS re:Invent 2024, Amazon unveiled its first-party Amazon Nova foundation model family on Amazon Bedrock: text-only Nova Micro, multimodal vision-language models Nova Lite and Nova Pro (with 300K context and built-in Distillation & Synthetic Data target support from Nova Premier), Nova Canvas for image generation, and Nova Reel for video generation.
2025
ReleaseAWS made EC2 Trn3 UltraServers available. The system connects up to 144 Trainium3 accelerators with NeuronLink for large-scale training and serving.
2026
ReleaseAWS published a guide to serving language models on Trainium with SGLang-Neuron. The integration brings SGLang scheduling and prefix caching to compatible models running through the Neuron SDK.
See also (20)
Open full graphAmazon & AWS AI
15 more connections
- Custom AI AcceleratorsOutgoing · EnablesIncoming · Mentions
- Distributed TrainingOutgoing · Enables
- Google DeepMindOutgoing · Alternative to
- Microsoft AIOutgoing · Alternative to
- Serverless GPU & Inference CloudsOutgoing · Enables
- Amazon NovaOutgoing · MentionsIncoming · Created by
- Distillation & Synthetic DataOutgoing · Mentions
- Guardrails & AI SafetyOutgoing · Mentions
- Multimodal & Vision-Language ModelsOutgoing · Mentions
- NVIDIAOutgoing · Mentions
- Pre-TrainingOutgoing · Mentions
- RAG & Hybrid RetrievalOutgoing · Mentions
- SGLangOutgoing · Mentions
- Video Generation Models (Sora, Veo, Kling, Wan)Outgoing · Mentions
- vLLMOutgoing · Mentions