← VS.

AWS

50 services

Cloud reference — AI deployment stack

Every AWS service an AI app actually needs

Fifty services across eight layers — from Bedrock and SageMaker down to the IAM policies and X-Ray traces underneath them. Each entry carries the real limits, pricing shape, and integration points you hit in production, not the marketing page.

200+
Services
33
Regions
105
Availability zones
450+
CloudFront PoPs
◈

Generative AI

6 services
◈Generative AI

Bedrock

Fully managed foundation model service

▸Single API across Claude, Llama 3/3.1, Mistral, Titan, Command R+, and Stable Diffusion
▸Converse API — unified request shape with streaming, tool use, and multimodal input
▸Knowledge Bases — managed RAG with chunking, embedding, and vector store sync
Details →2023
❖Generative AI

Q Business

Enterprise GenAI assistant over company data

▸40+ managed connectors: S3, SharePoint, Confluence, Salesforce, Jira, Gmail, ServiceNow
▸Permission-aware answers that respect source-system document ACLs
▸IAM Identity Center integration for workforce identity and SSO
Details →2024
⌘Generative AI

Q Developer

AI coding assistant across the SDLC

▸Inline multi-line code completion in VS Code, JetBrains, Visual Studio, and CLI
▸/dev agent — implements features across multiple files from a task description
▸/transform — automated Java 8/11 → 17 codebase upgrades
Details →2024
⌬Generative AI

Trainium2

Custom silicon for large-model training

▸Up to 4x training performance over first-generation Trainium
▸trn2.48xlarge: 16 Trainium2 chips, 1.5 TB HBM, 6 TB/s memory bandwidth aggregate
▸Trn2 UltraServers: 64 chips in one NeuronLink domain
Details →2024
⏣Generative AI

Inferentia2

Custom silicon for low-cost model inference

▸Up to 4x throughput and 10x lower latency vs Inferentia1
▸32 GB HBM per chip; inf2.48xlarge aggregates 384 GB across 12 chips
▸NeuronLink inter-chip interconnect enables tensor-parallel LLM sharding
Details →2023
✦Generative AI

Titan Models

AWS first-party foundation model family

▸Titan Text Lite / Express / Premier — tiered capability and cost
▸Titan Text Embeddings V2 with flexible 256/512/1024 dimensions
▸Titan Multimodal Embeddings — unified image+text vector space
Details →2023
⬡

ML Platform

11 services
⬡ML Platform

SageMaker Studio

Unified web IDE for the full ML lifecycle

▸Managed JupyterLab and VS Code-based Code Editor environments
▸Swap instance types (CPU↔GPU) without losing notebook state
▸Experiment tracking with managed MLflow
Details →2020
▣ML Platform

SageMaker Training

Managed, distributed model training jobs

▸Ephemeral managed clusters — provision, train, tear down automatically
▸100+ pre-built framework containers (PyTorch, TF, Hugging Face, XGBoost)
▸Bring-your-own training container from ECR
Details →2017
◎ML Platform

SageMaker Inference

Four endpoint types for any serving pattern

▸Real-time endpoints with auto-scaling and multi-AZ placement
▸Serverless Inference — scales to zero, 1–6 GB memory tiers
▸Asynchronous Inference — 1 GB payloads, queued, SNS completion notifications
Details →2017
⊞ML Platform

Feature Store

Central repository for ML features

▸Online store with single-digit-ms GetRecord latency
▸Offline store on S3 with Glue Data Catalog and Apache Iceberg format
▸Point-in-time (time-travel) queries for leakage-free training sets
Details →2020
⟐ML Platform

Pipelines

CI/CD workflow orchestration for ML

▸Python SDK DAG definitions with typed step dependencies
▸Step types: Processing, Training, Tuning, Transform, Lambda, Callback, EMR, Condition, Fail
▸Step caching — reuse prior results when inputs are unchanged
Details →2020
⊟ML Platform

Model Registry

Versioned model catalog with approval gates

▸Model packages versioned within model package groups
▸Approval statuses (Pending/Approved/Rejected) as deployment gates
▸EventBridge events on status change for CI/CD-triggered deployment
Details →2020
◐ML Platform

Clarify

Bias detection and model explainability

▸20+ pre-training bias metrics on raw datasets
▸Post-training bias metrics across sensitive facets (disparate impact, DPPL)
▸SHAP-based global and per-prediction feature attribution
Details →2020
▤ML Platform

Data Wrangler

Visual data preparation for ML

▸300+ built-in transforms with visual preview
▸Connectors: S3, Athena, Redshift, Snowflake, Databricks, EMR
▸Custom transforms in PySpark, pandas, and SQL
Details →2020
▥ML Platform

Ground Truth

Data labeling with human workforces

▸Three workforce options: Mechanical Turk, vendors, private teams
▸Task types: bounding box, segmentation, video tracking, NER, 3D point cloud
▸Active-learning automated labeling — up to 70% cost reduction
Details →2018
⬢ML Platform

HyperPod

Resilient clusters for foundation model training

▸Persistent clusters purpose-built for long FM training runs
▸Automated deep health checks on GPU/Trainium nodes
▸Auto-replacement of failed nodes with checkpoint-based resume
Details →2023
✧ML Platform

JumpStart

Model hub with one-click deployment

▸350+ models: Llama, Falcon, Bloom, Mistral, Stable Diffusion, embeddings
▸One-click deploy to a SageMaker real-time endpoint
▸Pre-validated serving containers and instance recommendations
Details →2020
◬

AI Applications

10 services
◬AI Applications

Rekognition

Computer vision for images and video

▸Object and scene detection across thousands of labels
▸Face detection, comparison, and search over 20M-face collections
▸Facial analysis: landmarks, emotions, age range, accessories
Details →2016
▦AI Applications

Textract

Document AI beyond OCR

▸OCR with layout-aware reading order, including handwriting
▸Forms extraction — key-value pairs with confidence scores
▸Table extraction with merged-cell and header recognition
Details →2019
▧AI Applications

Comprehend

NLP: entities, sentiment, PII, topics

▸Entity recognition: people, orgs, locations, dates, quantities
▸Document and targeted sentiment (per-entity polarity)
▸PII detection and redaction across 20+ identifier types
Details →2017
◭AI Applications

Transcribe

Speech-to-text, streaming and batch

▸Streaming (real-time) and batch transcription modes
▸100+ languages with automatic language identification
▸Speaker diarization (up to 30 speakers) and channel identification
Details →2018
◮AI Applications

Polly

Lifelike text-to-speech

▸60+ voices across 29 languages
▸Four engines: standard, neural, long-form, generative
▸SSML support: prosody, pauses, phonemes, say-as rules
Details →2016
⇄AI Applications

Translate

Neural machine translation at scale

▸75 languages, 5,500+ translation pair combinations
▸Real-time translation with automatic source-language detection
▸Batch translation of S3 document sets (HTML, DOCX, XLSX, PPTX)
Details →2018
◔AI Applications

Lex

Conversational bots for voice and text

▸Intent + slot conversation modeling with multi-turn dialogs
▸Built-in ASR and NLU — same technology family as Alexa
▸Lambda hooks for validation and fulfillment logic
Details →2017
⌕AI Applications

Kendra

ML-powered intelligent enterprise search

▸Natural-language question answering with extracted answer passages
▸40+ managed connectors with incremental sync
▸ACL-aware results honoring source-system permissions
Details →2020
⊛AI Applications

Personalize

Real-time recommendations as a service

▸User-personalization, similar-items, and trending-now recipes
▸Personalized ranking for re-ordering candidate lists
▸Next-best-action recommendations
Details →2019
⋰AI Applications

Forecast

Managed time-series forecasting

▸AutoML across DeepAR+, CNN-QR, Prophet, NPTS, ARIMA, ETS
▸Probabilistic forecasts at arbitrary quantiles (P1–P99)
▸Related time series covariates (price, promotion, traffic)
Details →2019
⬢

Compute & Serverless

6 services
⬢Compute & Serverless

Lambda

Event-driven serverless functions

▸Zero server management with automatic scaling from zero
▸Up to 10 GB memory / 6 vCPUs and 10 GB ephemeral storage per function
▸Container image deployment up to 10 GB
Details →2014
▣Compute & Serverless

ECS

AWS-native container orchestration

▸Task definitions with per-task IAM roles and resource limits
▸Fargate and EC2 launch types in the same cluster
▸Capacity providers with managed On-Demand/Spot scaling
Details →2015
⎈Compute & Serverless

EKS

Managed Kubernetes control plane

▸Managed, multi-AZ, upstream-conformant Kubernetes control plane
▸EKS Auto Mode — fully managed nodes, upgrades, and networking
▸Karpenter just-in-time node autoscaling with Spot awareness
Details →2018
⬟Compute & Serverless

Fargate

Serverless compute for containers

▸Zero host management — no AMIs, patching, or capacity planning
▸Firecracker micro-VM isolation per task/pod
▸Up to 16 vCPU / 120 GB memory per task
Details →2017
⟜Compute & Serverless

Step Functions

Visual workflow orchestration

▸Standard workflows: up to 1-year duration, exactly-once execution
▸Express workflows: high-throughput, at-least-once, 5-minute max
▸220+ service integrations including direct Bedrock invocation
Details →2016
⊶Compute & Serverless

EventBridge

Serverless event bus and scheduler

▸Default, custom, and partner event buses
▸Content-based rule matching on any event field
▸90+ AWS service event sources, SaaS partner integrations
Details →2019
☷

Data & Storage

7 services
☷Data & Storage

S3

Object storage with 11 nines durability

▸99.999999999% (11 nines) designed durability
▸Objects up to 5 TB; unlimited bucket capacity
▸Storage classes: Standard, Intelligent-Tiering, Express One Zone, Glacier
Details →2006
◫Data & Storage

DynamoDB

Serverless NoSQL at any scale

▸Single-digit-millisecond latency at virtually any throughput
▸On-demand and provisioned capacity modes with auto-scaling
▸Global tables: multi-region, active-active replication
Details →2012
◍Data & Storage

Aurora

Cloud-native relational + pgvector

▸PostgreSQL and MySQL compatibility with up to 5x throughput
▸Six-way storage replication across three AZs
▸Auto-scaling storage to 128 TiB
Details →2015
◌Data & Storage

OpenSearch

Search, analytics, and k-NN vectors

▸k-NN vector search: HNSW/IVF via Faiss and Lucene engines
▸Hybrid search — BM25 + vector scores with normalization processors
▸Neural search plugin with ingest-time embedding generation
Details →2015
◧Data & Storage

ElastiCache

Managed Valkey/Redis in-memory layer

▸Valkey, Redis OSS, and Memcached engine support
▸Microsecond read latency at millions of ops/second
▸ElastiCache Serverless — no node management, 99.99% SLA
Details →2011
⊠Data & Storage

Glue

Serverless ETL and data catalog

▸Serverless Spark and Ray job execution
▸Glue Data Catalog — shared metastore for Athena/EMR/Redshift
▸Crawlers with automatic schema discovery and partition detection
Details →2017
≋Data & Storage

Kinesis

Real-time data streaming backbone

▸Sub-second latency from producer to consumer
▸Provisioned (shard-based) and on-demand capacity modes
▸Each shard: 1 MB/s in, 2 MB/s out, 1,000 records/s
Details →2013
⟁

Networking & APIs

4 services
⟁Networking & APIs

API Gateway

Managed front door for APIs

▸REST, HTTP, and WebSocket API types
▸Lambda proxy integration — the canonical serverless pattern
▸IAM, Cognito, Lambda authorizers, and mutual TLS auth
Details →2015
◍Networking & APIs

CloudFront

Global CDN with edge compute

▸450+ global points of presence
▸Origin Access Control for locked-down S3 origins
▸Origin Shield regional caching layer
Details →2008
▷Networking & APIs

App Runner

Source-to-URL managed app hosting

▸Deploy from ECR image or GitHub source with managed builds
▸Automatic HTTPS endpoint with managed TLS certificates
▸Request-concurrency-based auto-scaling
Details →2021
⫛Networking & APIs

ELB

ALB, NLB, and GWLB traffic distribution

▸ALB: layer-7 routing on path, host, headers, query strings
▸ALB: Lambda targets and built-in OIDC/Cognito authentication
▸ALB: WebSocket, HTTP/2, gRPC support with health checks
Details →2009
◉

Security & Identity

4 services
◉Security & Identity

IAM

Who can do what, on which resource

▸Identity-based and resource-based JSON policies
▸Roles with temporary credentials via STS
▸IAM Roles for service compute: Lambda, EC2, ECS tasks, EKS IRSA
Details →2011
◒Security & Identity

Cognito

Customer identity and app authentication

▸User Pools: managed sign-up/sign-in with email/phone verification
▸MFA: TOTP, SMS, and WebAuthn passkeys
▸Hosted UI and managed login pages with OAuth 2.0/OIDC flows
Details →2014
◈Security & Identity

Secrets Manager

Secrets lifecycle with automatic rotation

▸Managed rotation for RDS, Aurora, Redshift, DocumentDB
▸Custom rotation via Lambda for any credential type
▸Versioning with staging labels for zero-downtime rotation
Details →2018
⊘Security & Identity

KMS

Managed encryption keys and envelope crypto

▸FIPS-validated HSM-backed key storage — material never leaves
▸Symmetric and asymmetric CMKs, plus HMAC keys
▸Envelope encryption with GenerateDataKey
Details →2014
⊕

Monitoring & Ops

2 services
⊕Monitoring & Ops

CloudWatch

Metrics, logs, alarms, and dashboards

▸Automatic metrics from 100+ AWS services
▸Custom and high-resolution (1-second) metrics
▸Metric math, anomaly detection bands, and composite alarms
Details →2009
⌖Monitoring & Ops

X-Ray

Distributed tracing across services

▸End-to-end request tracing with segments and subsegments
▸Live service map with latency and error-rate overlays
▸Centralized, dynamic sampling rules
Details →2017

Reference architecture — serverless GenAI app

The request path, end to end

◎

User

Browser / mobile client

⊚

CloudFront

Global edge — TLS, caching, WAF

⟁

API Gateway

Auth, throttling, routing

⬢

Lambda

Orchestration & business logic

◈

Bedrock

Foundation model inference

◫

DynamoDB

Session & conversation state

☷

S3

Documents, embeddings source

A request enters at the CloudFront edge, where static assets are served from cache and dynamic calls are forwarded over the AWS backbone. API Gateway validates the JWT, applies per-key throttling, and proxies to Lambda, which assembles context — conversation history from DynamoDB, retrieved chunks from a knowledge base backed by S3 — and calls Bedrock's Converse API with streaming enabled. Tokens stream back through Lambda response streaming to the client while the completed turn is written back to DynamoDB.