← VS.

Azure

37 services

Cloud reference — Microsoft AI stack

The Azure stack behind enterprise AI

Thirty-seven services across eight layers — from Azure OpenAI and AI Foundry down to Entra ID, Key Vault, and the KQL queries that keep it accountable. Each entry carries real quotas, pricing shape, integration points, and its closest AWS equivalent for anyone translating between clouds.

60+
Regions
200+
Services
192
Front Door PoPs
⊞
Microsoft global backbone
◈

Generative AI

5 services
◈
Generative AIAWS · Amazon Bedrock

Azure OpenAI

OpenAI frontier models inside Azure's security boundary

▸GPT-4o, GPT-4o mini, o1/o1-mini/o3-mini reasoning models, DALL-E 3, Whisper, text-embedding-3-large/small
▸Global, Data Zone, and Regional deployment types with different quota pools and data-residency guarantees
▸Provisioned Throughput Units (PTU) — reserved capacity with predictable latency for production workloads
Details →2023
⬡
Generative AIAWS · Amazon Bedrock + SageMaker Studio

AI Foundry

Unified portal and SDK for building, evaluating, and shipping AI apps

▸Model catalog: 1,600+ models — OpenAI, Llama 3.x, Mistral, Cohere, Phi, DeepSeek, Hugging Face
▸Serverless API (pay-per-token MaaS) or managed compute deployment for open models
▸Prompt flow: DAG-based LLM workflow authoring with Python/LLM/prompt nodes and variants
Details →2023
❖
Generative AIAWS · Amazon Q Business

Copilot Studio

Low-code platform for building copilots and autonomous agents

▸Generative answers grounded on SharePoint, Dataverse, websites, and uploaded documents
▸Generative orchestration — LLM-driven selection of topics, knowledge, and actions per turn
▸Topic authoring canvas with variables, conditions, and adaptive cards
Details →2023
◬
Generative AIAWS · Amazon Bedrock Guardrails

AI Content Safety

Harm detection, prompt shields, and groundedness checks for GenAI

▸Four harm categories — hate, violence, sexual, self-harm — with 0–7 severity scoring
▸Prompt Shields: jailbreak (direct) and indirect prompt-injection attack detection
▸Groundedness detection — verifies responses against source documents, with optional correction
Details →2023
✦
Generative AIAWS · Amazon Nova Micro / Titan Models

Phi Models

Small language models punching far above their parameter count

▸Phi-4 (14B): frontier-class math and reasoning performance at SLM size
▸Phi-3 family: mini (3.8B), small (7B), medium (14B) with 4K and 128K context variants
▸Phi-3 Vision / Phi-3.5 Vision for image + text multimodal understanding
Details →2023
⬢

ML Platform

1 service
⬢
ML PlatformAWS · Amazon SageMaker

Azure Machine Learning

End-to-end enterprise MLOps — train, track, deploy, govern

▸Compute clusters: auto-scale 0→N CPU/GPU nodes, spot/low-priority VMs, A100/H100 ND-series with InfiniBand for distributed training
▸Compute instances: managed dev boxes with JupyterLab, VS Code Remote, auto-shutdown schedules, and setup scripts
▸AutoML for classification, regression, forecasting, NLP, and computer vision with model explainability built in
Details →2018
◉

AI Applications

9 services
◉
AI ApplicationsAWS · Amazon Rekognition

AI Vision

Image analysis, OCR, faces, and spatial video understanding

▸Image Analysis 4.0 on the Florence foundation model — captioning, dense captions, tags, objects in one call
▸Read OCR: printed + handwritten text, 160+ languages, images and multi-page PDFs
▸Smart cropping and thumbnail generation aware of regions of interest
Details →2017
☷
AI ApplicationsAWS · Amazon Comprehend

AI Language

NLP suite — NER, sentiment, summarization, PII, custom QA

▸Named entity recognition across 100+ entity types, plus entity linking to Wikipedia
▸PII detection and redaction (SSNs, credit cards, names, addresses) for text and conversations
▸Sentiment analysis with aspect-based opinion mining (target + assessment pairs)
Details →2021
∿
AI ApplicationsAWS · Amazon Transcribe + Polly

AI Speech

Speech-to-text, 400+ neural voices, translation, voice cloning

▸Real-time streaming STT with interim results, word-level timestamps, and punctuation
▸Batch transcription API for large audio archives (diarization, multi-channel)
▸Custom Speech: acoustic + language model adaptation for domain vocabulary and accents
Details →2018
⊞
AI ApplicationsAWS · Amazon Textract

Document Intelligence

Layout-aware document extraction — invoices, IDs, custom models

▸Layout model: paragraphs, tables, selection marks, barcodes, formulas, reading order — with Markdown output for RAG
▸Read model: high-accuracy print + handwriting OCR across 300+ languages
▸Prebuilt models: invoice, receipt, ID document, W-2, 1098/1099, health insurance card, contract, pay stub
Details →2020
⌖
AI ApplicationsAWS · Amazon Kendra + OpenSearch

AI Search

Hybrid vector + keyword retrieval with semantic reranking

▸Vector search with HNSW (approximate) and exhaustive KNN, up to 4,096-dimension embeddings
▸Hybrid retrieval: BM25 + vector results fused with Reciprocal Rank Fusion (RRF)
▸Semantic ranker: cross-encoder L2 reranking with extractive captions and answers
Details →2015
⇌
AI ApplicationsAWS · Amazon Translate

AI Translator

Neural machine translation across 100+ languages

▸Text translation across 100+ languages, multiple target languages per request
▸Document Translation: async batch for DOCX/PPTX/XLSX/PDF/HTML with layout preservation
▸Synchronous single-document translation endpoint (no Blob staging required)
Details →2016
⌬
AI ApplicationsAWS · Amazon Lex

Bot Service

Bot Framework runtime and omnichannel connectivity

▸Bot Framework SDK v4 for C#, JavaScript, and Python with dialog and state management
▸Omnichannel: Teams, web chat, Slack, Facebook, email, SMS, Direct Line custom clients
▸Direct Line and Direct Line Speech channels for embedded and voice experiences
Details →2017
⍟
AI ApplicationsAWS · Amazon Personalize

AI Personalizer

Contextual-bandit reinforcement learning for content ranking

▸Contextual bandit learning loop: Rank API chooses, Reward API teaches
▸Action features + context features as arbitrary JSON namespaces
▸Configurable exploration percentage to balance exploit vs discover
Details →2019
≋
AI ApplicationsAWS · Amazon Lookout for Metrics

Anomaly Detector

Univariate and multivariate time-series anomaly detection

▸Univariate detection: ensemble of spectral residual, ESD, and seasonal decomposition algorithms
▸Automatic seasonality detection and per-series algorithm selection
▸Streaming (last-point) and batch (entire-series) detection modes
Details →2021
⬠

Compute & Serverless

5 services
ƒ
Compute & ServerlessAWS · AWS Lambda

Functions

Event-driven serverless compute with Durable orchestration

▸Triggers: HTTP, Timer (CRON), Blob, Queue, Service Bus, Event Hubs, Event Grid, Cosmos DB change feed, Kafka, SignalR
▸Declarative input/output bindings — read/write queues, blobs, and databases without SDK plumbing
▸Durable Functions: stateful orchestrators with fan-out/fan-in, chaining, human interaction, and entity patterns
Details →2016
⬠
Compute & ServerlessAWS · AWS App Runner / ECS Fargate

Container Apps

Serverless containers with KEDA scaling, Dapr, and GPU profiles

▸Scale to zero and event-driven autoscaling via 50+ KEDA scalers (HTTP, queues, Kafka, CRON)
▸Integrated Dapr sidecars: service invocation, pub/sub, state stores, bindings, secrets
▸Revisions with traffic splitting for blue/green and canary deployment
Details →2022
⎈
Compute & ServerlessAWS · Amazon EKS

AKS

Managed Kubernetes with GPU pools and an AI toolchain operator

▸Managed control plane with uptime SLA (Standard), automatic upgrade channels, and maintenance windows
▸System + user node pools, spot node pools, and cluster autoscaler with KEDA event scaling
▸AKS Automatic: preset production clusters with node auto-provisioning (Karpenter-based)
Details →2018
▣
Compute & ServerlessAWS · AWS Elastic Beanstalk

App Service

Managed PaaS web hosting with slots and zero-downtime swaps

▸Runtimes: .NET, Node, Python, Java/Tomcat, PHP, plus custom containers on Linux/Windows
▸Deployment slots with zero-downtime swap, slot-specific config, and % traffic routing
▸Autoscale: schedule + metric rules (Premium v3 adds automatic scaling with pre-warmed instances)
Details →2015
⊡
Compute & ServerlessAWS · AWS Fargate

Container Instances

Single containers on demand, billed per second

▸Containers start in seconds with no cluster or plan to provision
▸Per-second billing — pay only while the container group runs
▸Container groups: multi-container sidecars with shared localhost network and volumes
Details →2018
⛁

Data & Storage

7 services
⛁
Data & StorageAWS · Amazon S3

Blob Storage

Massively scalable object storage with a data-lake mode

▸Hot / Cool / Cold / Archive access tiers with lifecycle management policies
▸Azure Data Lake Storage Gen2 mode: hierarchical namespace, POSIX ACLs, Hadoop ABFS driver
▸Block blobs up to 190.7 TiB; append blobs for logs; page blobs for random I/O
Details →2010
◍
Data & StorageAWS · Amazon DynamoDB

Cosmos DB

Planet-scale NoSQL with multi-region writes and DiskANN vectors

▸99.999% availability SLA with multi-region writes and automatic failover
▸Five consistency levels: strong, bounded staleness, session, consistent prefix, eventual
▸APIs: NoSQL, MongoDB, Cassandra, Gremlin (graph), Table, PostgreSQL (Citus)
Details →2017
⛃
Data & StorageAWS · Amazon Aurora / RDS

SQL Database

Managed SQL Server with serverless and 100 TB Hyperscale tiers

▸Serverless tier: per-second vCore billing with auto-pause/resume on idle
▸Hyperscale: up to 100 TB, page-server storage architecture, near-instant backups, fast named replicas
▸Business Critical: in-memory OLTP, local SSD, built-in readable secondary replica
Details →2010
⌗
Data & StorageAWS · Amazon ElastiCache

Cache for Redis

Managed in-memory cache with RediSearch vector similarity

▸Sub-millisecond latency for cache, session, counter, and queue patterns
▸Tiers: Basic, Standard (replicated), Premium (cluster + persistence + VNet), Enterprise (Redis Stack)
▸RediSearch vector similarity (HNSW/FLAT) for embeddings and semantic caching
Details →2014
⧉
Data & StorageAWS · AWS Glue

Data Factory

Serverless data integration with 100+ connectors

▸100+ connectors: SaaS (Salesforce, SAP, Dynamics), databases, files, REST/OData
▸Copy Activity with parallel DIU-scaled movement, staging, and fault-tolerant skip/log
▸Mapping Data Flows: visual Spark transformations on managed clusters
Details →2015
⫸
Data & StorageAWS · Amazon Kinesis Data Streams

Event Hubs

Kafka-compatible event streaming at millions of events per second

▸Millions of events/second with partitioned consumers and consumer groups
▸Apache Kafka protocol endpoint — existing Kafka clients work unchanged
▸Capture: automatic archival of streams to Blob/ADLS Gen2 in Avro
Details →2014
⇶
Data & StorageAWS · Amazon SQS + SNS

Service Bus

Enterprise message broker — queues, topics, sessions, DLQs

▸Queues with peek-lock, at-least-once or at-most-once receive, and competing consumers
▸Topics + subscriptions with SQL filter and correlation filter rules per subscription
▸Dead-letter queues on every entity with automatic poison-message routing
Details →2011
⟁

Networking & APIs

4 services
⊕
Networking & APIsAWS · Amazon API Gateway

API Management

Full API lifecycle gateway — now the AI gateway for LLM traffic

▸Policy engine: inbound/backend/outbound XML policies at global/product/API/operation scope
▸AI gateway: llm-token-limit (TPM per subscription), token quotas, and token-metric emitting policies
▸Semantic caching of LLM responses backed by Azure Cache for Redis
Details →2014
⟁
Networking & APIsAWS · Amazon CloudFront

Front Door

Global anycast entry point — CDN, WAF, and L7 load balancing

▸192+ global PoPs with anycast TLS termination and split TCP for faster handshakes
▸Traffic rides the Microsoft global backbone from edge to origin
▸Global load balancing: latency/priority/weighted routing with health probes
Details →2019
◫
Networking & APIsAWS · Application Load Balancer (ALB)

Application Gateway

Regional L7 load balancer with WAF and end-to-end TLS

▸URL path-based and multi-site (host header) routing to backend pools
▸v2 autoscaling capacity units with zone redundancy and static VIP
▸WAF: OWASP CRS 3.x managed rules, per-site policies, exclusions, custom rules, bot protection
Details →2015
◳
Networking & APIsAWS · AWS Amplify Hosting

Static Web Apps

Git-driven Jamstack hosting with managed Functions APIs

▸Git-driven CI/CD: push to build and deploy via GitHub Actions/Azure Pipelines automatically
▸Globally distributed static content with free managed SSL
▸Integrated /api backend: managed Azure Functions or linked BYO backends (App Service, Container Apps, APIM)
Details →2021
⚿

Security & Identity

3 services
❂
Security & IdentityAWS · AWS IAM + Cognito

Entra ID

Cloud identity platform — OAuth, Conditional Access, managed identities

▸OAuth 2.0 / OpenID Connect / SAML identity provider with MSAL libraries for every stack
▸Managed identities (system- and user-assigned) — credential-free auth between Azure resources
▸Conditional Access: policy-driven MFA, device compliance, location, and risk gating per app
Details →2012
⚿
Security & IdentityAWS · AWS Secrets Manager + KMS

Key Vault

Centralized secrets, keys, and certificates with HSM backing

▸Secrets management with versioning, expiry dates, and activation windows
▸Keys: RSA 2048–4096, EC, software or FIPS 140-2 validated HSM protection (Premium)
▸Certificate lifecycle: issuance and auto-renewal via DigiCert/GlobalSign integration
Details →2015
⛨
Security & IdentityAWS · AWS Security Hub + GuardDuty

Defender for Cloud

CSPM + workload protection across Azure, AWS, and GCP

▸Secure Score: continuous posture assessment against Microsoft Cloud Security Benchmark
▸Attack-path analysis chaining misconfigs into exploitable routes (Defender CSPM)
▸Multicloud coverage: native connectors for AWS accounts and GCP projects
Details →2016
◭

Monitoring & Ops

3 services
◭
Monitoring & OpsAWS · Amazon CloudWatch

Azure Monitor

Unified metrics, logs, alerts, and autoscale for all Azure resources

▸Platform metrics: 1-min granularity, 93-day retention, free for all Azure resources
▸Diagnostic settings routing resource logs to Log Analytics, Event Hubs, or Storage
▸Metric alerts with dynamic thresholds (ML-trained baselines) and multi-resource scoping
Details →2017
⊙
Monitoring & OpsAWS · AWS X-Ray

Application Insights

APM — distributed tracing, dependency maps, live metrics

▸Auto-instrumentation for App Service, Functions, and AKS — no code changes
▸OpenTelemetry distro for .NET, Java, Node.js, and Python with Azure exporters
▸Distributed tracing with W3C trace context across service and queue hops
Details →2016
⋈
Monitoring & OpsAWS · CloudWatch Logs Insights

Log Analytics

KQL query workspace over all operational telemetry

▸KQL: pipe-based query language with summarize, join, parse, ML operators (autocluster, series_decompose_anomalies), and render
▸Table plans: Analytics, Basic, and Auxiliary tiers balancing cost vs query capability
▸Long-term archive up to 12 years with search jobs and restore for rehydration
Details →2015

Reference architecture — Azure GenAI app

The request path, end to end

◎

User

Browser / mobile client

⊚

Azure Front Door

Global edge — TLS, WAF, caching

⟁

API Management

AI gateway — auth, token limits

⬠

Container Apps / Functions

Orchestration & business logic

◈

Azure OpenAI

GPT-4o / o-series inference

⌕

AI Search

Hybrid vector retrieval (RAG)

◫

Cosmos DB

Session & conversation state

Requests land on Azure Front Door's 192 edge PoPs, ride the Microsoft global backbone, and hit API Management acting as an AI gateway — validating Entra ID tokens, enforcing per-subscriber token-rate limits, and load-balancing across Azure OpenAI deployments with semantic caching. The orchestration tier in Container Apps or Durable Functions grounds the prompt with hybrid vector + semantic results from AI Search, calls Azure OpenAI with streaming, and persists each turn to Cosmos DB — with managed identities securing every hop, so no connection string ever leaves Key Vault.