AI, Data and Analytics Services Overview

Four terms are often used interchangeably but are nested. Artificial intelligence is the broad goal of software doing tasks that normally need human judgement. Machine learning (ML) is the subset that learns patterns from data instead of following hand-written rules. Deep learning (DL) is ML with many-layered neural networks, behind most image, speech and language models. Generative AI is deep learning that produces new content (text, images, code, audio) from a prompt, typically with a pre-trained foundation model that you call rather than train. AWS offers three levels of involvement for all of them: call a prebuilt AI service (Rekognition, Textract, Polly), call a foundation model through Amazon Bedrock, or build and train your own model on Amazon SageMaker AI. Pick the highest level that solves the problem; each lower level adds flexibility and operational work.

Amazon Bedrock is a fully managed service that exposes more than a hundred foundation models from Amazon and other providers (the catalogue changes monthly) through a single, IAM-protected API, so an application can switch models by changing a model or inference-profile identifier. New code should use the bedrock-runtime endpoint: the Converse API is the provider-neutral call for text and multimodal chat, InvokeModel sends a provider-specific body, and the endpoint also offers Anthropic Messages-compatible and OpenAI-compatible (Responses and Chat Completions) routes. Cross-Region inference profiles (identifiers prefixed global., us., eu. and so on) spread requests over Regions for throughput. Around inference, Bedrock adds Guardrails (content, prompt-attack, denied-topic and sensitive-information filters, also callable on their own with apply-guardrail), model evaluation, batch inference (create-model-invocation-job), prompt management, customization (fine-tuning, distillation) and provisioned throughput for steady high-volume use. Access is controlled by IAM like any other service; short-term and long-term Bedrock API keys exist for quick starts, but production code should use IAM roles and the default credential chain (see Identity and access).

Amazon SageMaker AI is the managed ML platform for building, training and hosting your own models or customising foundation models: notebooks and IDEs in SageMaker Studio, managed training jobs, real-time, serverless, asynchronous and batch inference endpoints, Pipelines for MLOps, JumpStart (pre-trained models and solution templates), Canvas (no-code ML) and HyperPod for large training clusters. Two naming facts cause most confusion. First, on 2024-12-03 the original Amazon SageMaker was renamed Amazon SageMaker AI; APIs, the aws sagemaker CLI group, IAM prefixes (AmazonSageMaker…​), CloudFormation AWS::SageMaker:: types and console URLs did not change. Second, the same day AWS launched a next generation of Amazon SageMaker: a unified data, analytics and AI platform that brings SageMaker AI, SageMaker Lakehouse (open-table-format access across S3 and Redshift), SageMaker Catalog and governance (built on DataZone), SQL analytics (Redshift), data processing (Athena, EMR, Glue) and SageMaker Unified Studio into one environment, alongside Bedrock. For a plain "train and host a model" job, you still want SageMaker AI; for a data-and-AI team workspace, look at Unified Studio. A group of lesser-used SageMaker AI features entered maintenance during 2026, so check the lifecycle page before adopting Ground Truth, Model Monitor, Debugger, Clarify or Augmented AI.

First step. Console: Amazon SageMaker AI → Studio (or Unified Studio for the platform), create a domain with an IAM role, open JumpStart and deploy a pre-trained model to an endpoint. CLI: inspect what exists.

See What is Amazon SageMaker AI? and the sagemaker command reference. Real-time endpoints bill while they exist, so delete them when finished (see Clean up).

Amazon Q Developer was the AWS coding assistant (IDE plugins, console chat, command-line help). AWS announced that the Q Developer IDE plugins and paid subscriptions reach end of support on 2027-04-30, that new sign-ups stopped in May 2026, and that Kiro, a spec-driven, agentic IDE with Model Context Protocol support, is the recommended replacement. Amazon Q Business (the workplace assistant) moved into maintenance in July 2026 with Amazon Quick as the successor. None of these are needed to run the services on this page; they are developer tools, so see Developer tools and DevOps for the IDE and CI/CD side. First step. Install Kiro from its official site and follow the Q Developer migration guide linked below; there is no AWS CLI command.

See Amazon Q Developer IDE plugins end of support.

The prebuilt services are single-purpose APIs: you send an image, document, audio file or text and get structured JSON back, with no model to train. They remain the cheapest and most predictable route when the task matches one of them, and a foundation model through Bedrock is the alternative when it does not. All of them read inputs from S3 (so the bucket policy and KMS key must allow the calling role) and publish asynchronous results to S3 or SNS.

The AWS analytics pattern is a data lake on S3 (cheap, durable, open formats such as Parquet and Apache Iceberg) described by a catalog, queried by whichever engine fits, and secured by one permission layer. S3 itself is covered in S3 object storage.

AWS IoT Core connects devices to AWS over MQTT (also MQTT over WebSockets and HTTPS), with X.509 certificates and IoT policies for per-device authentication, a thing registry, device shadows (last reported and desired state), rules that route messages to Lambda, S3, DynamoDB, Kinesis, SQS and others, and jobs for fleet updates. AWS IoT Greengrass runs the same logic on edge devices. Bookshelf has no devices, but a connected book-return kiosk would publish scans to bookshelf/kiosk/<id>/scans and a rule would write them to Firehose. Services built around IoT also moved: AWS IoT FleetWise went into maintenance on 2026-04-30 (closed to new customers), AWS Panorama ended support on 2026-05-31, and Amazon Lookout for Equipment ends on 2026-10-07.

Console: IoT Core → Connect → Connect one device walks through certificates, policy and a test client (MQTT test client). See What is AWS IoT?.

Several services that appear in AWS service catalogues are neither AI nor analytics; they get one line each so no name is left unexplained. Amazon Connect is a cloud contact centre (aws connect list-instances) that integrates with Lex and Bedrock. Amazon WorkSpaces provides managed virtual desktops (aws workspaces describe-workspaces); Amazon WorkSpaces Applications (formerly AppStream 2.0, CLI group still appstream) streams individual desktop applications to a browser; and Amazon WorkSpaces Secure Browser (aws workspaces-web list-portals) gives a managed, isolated browser for internal web apps, which moves to maintenance on 2026-10-29 (closed to new customers). AWS AppSync is the managed GraphQL and publish/subscribe API layer (aws appsync list-graphql-apis); the REST/HTTP counterparts are in API Gateway. AWS Amplify is the front-end and full-stack hosting toolkit, summarised in Developer tools and DevOps, while static sites that you assemble yourself are in S3 security and static hosting. Amazon Pinpoint (marketing messaging) reaches end of support on 2026-10-30; use SES for email (SES email) and AWS End User Messaging for SMS and push (Messaging and integration). Hybrid and migration services (DMS, Schema Conversion, Snow Family, Outposts) are in Architecture, migration and hybrid. Each has an official page under docs.aws.amazon.com linked in the references below.

All services on this page follow the same rules as the rest of the section:

Prices change and differ by Region, so none are stated here; use the official pricing pages in the references (checked 2026-10-01) and the AWS Pricing Calculator. The cost drivers by service:

The recurring surprises are idle resources that bill continuously (SageMaker endpoints, Kendra indexes, provisioned Redshift or EMR clusters, Quick subscriptions) and unpartitioned CSV scanned repeatedly by Athena. Tag resources and set budgets before experimenting.

Every service here has Regional, adjustable and non-adjustable quotas; read them in the Service Quotas console or with aws service-quotas list-service-quotas --service-code <code>, and request increases before a launch rather than during one. The ones that most often bite are Bedrock requests-per-minute and tokens-per-minute (per model and per Region, and very low on new accounts), SageMaker instance-type quotas that default to zero for GPU families, Athena concurrent queries and per-query limits, Glue concurrent job runs, Kinesis shard throughput (1 MB/s or 1,000 records/s of writes per shard), Firehose delivery limits and Textract or Transcribe concurrent job limits. See Service Quotas and each service’s quotas page in the references.

Dates below were verified against the official lifecycle and availability pages on 2026-10-01. The authoritative source is the AWS service lifecycle page. "Closed to new customers" means existing accounts keep using the service.

Open model catalogues also change fast: Bedrock added Anthropic, OpenAI, xAI, DeepSeek and other families during 2026, and exposes newer models through the same bedrock-runtime calls, so list models at run time instead of hard-coding identifiers.

Older tutorials and certification material predate much of the above. Expect to correct these: Kinesis Data Firehose and Kinesis Data Analytics are no longer the names; QuickSight is now part of Amazon Quick; SageMaker is SageMaker AI and sits inside a larger platform; Bedrock is now a model, retrieval and guardrail platform with AgentCore rather than a short model list; Kendra is closed to new customers; Pinpoint is ending; and guides that start every setup with IAM-user access keys should be replaced with IAM Identity Center, aws login or CloudShell (see Getting started). AWS Glue DataBrew and the S3 plus Glue crawler plus Athena pattern are unchanged and still current.

Everything above is for illustration; remove anything you created, in dependency order, because several of these resources bill while idle:

Also delete the IAM roles you created for Glue and Firehose, the Kendra index (aws kendra delete-index) if you made one, and the Amazon Quick subscription (from Manage account → Account settings, as a Quick administrator), because Quick bills per user. Delete a versioned S3 bucket only after emptying all versions, as described in S3 object storage.