Skip to content
Go back

Every Cloud Sells the Same AI Stack. Each One Anchors It Somewhere Else.

AWS, Microsoft Azure, and Google Cloud now sell matching parts for every layer of an agent. They separate at the point where each one converts anything foreign into one of its own systems.

Foreign means a login from another identity provider, data stored on another cloud, employees who work in another vendor’s apps, or a model from another lab. All three clouds convert those into a home system, and no two pick the same one.

Four companion posts go one layer below this map: agent platforms, agent identity, models, and cloud marketplaces.

The map

Names and status come from each vendor’s product pages and docs as of November 18, 2026, and preview rows are marked preview.

What people use

CapabilityAWSMicrosoft AzureGoogle Cloud
Employee AI assistantAWS QuickMicrosoft 365 CopilotGemini Enterprise
Research over your documentsQuick ResearchResearcher agent in CopilotGemini Notebook Enterprise
AI coding agentKiro (IDE and CLI)GitHub CopilotAntigravity and Antigravity CLI
Chat with business dataAWS Quick SightCopilot in Power BI, Fabric data agentsConversational Analytics in BigQuery and Looker
Low-code agent builderQuick agents and FlowsCopilot StudioAgent Designer, Agent Studio

Agents

CapabilityAWSMicrosoft AzureGoogle Cloud
Agent runtimeAmazon Bedrock AgentCore Runtime and harnessMicrosoft Foundry Agent Service, hosted agentsAgent Runtime on Gemini Enterprise Agent Platform
Agent frameworkStrands AgentsMicrosoft Agent FrameworkAgent Development Kit (ADK)
Agent memoryAgentCore MemoryAgent Service memory (preview)Memory Bank and Sessions
Agent identityAgentCore IdentityMicrosoft Entra Agent IDAgent Identity (SPIFFE)
Agent registry and policyAWS Agent Registry, AgentCore PolicyMicrosoft Agent 365, Foundry Control PlaneAgent Registry, Agent Gateway
Safety filtersBedrock GuardrailsFoundry Guardrails, Prompt ShieldsModel Armor
Agent evaluationAgentCore EvaluationsFoundry evaluationsAgent Platform Evals
Managed RAGBedrock Managed Knowledge BaseAzure AI Search, Foundry IQRAG Engine, Agent Search
Custom ML platformAmazon SageMaker AIAzure Machine LearningAgent Platform training and inference
Custom AI chipsTrainium3, Inferentia2Maia 200 (runs Microsoft’s own services; not yet rentable)Ironwood TPU

Data

CapabilityAWSMicrosoft AzureGoogle Cloud
WarehouseAmazon RedshiftFabric Data WarehouseBigQuery
Open-table lakehouseS3 Tables, SageMaker lakehouseOneLake (Delta, readable as Iceberg)Lakehouse for Apache Iceberg
Catalog and governanceSageMaker Catalog, Glue Data Catalog, Lake FormationOneLake catalog, PurviewKnowledge Catalog
SparkAmazon EMR, AWS GlueFabric Spark, Azure DatabricksManaged Service for Apache Spark
StreamingKinesis, MSK, Managed FlinkEvent Hubs, Fabric Real-Time IntelligencePub/Sub, Managed Kafka, Dataflow
Pipeline schedulingAmazon MWAAData Factory in FabricManaged Service for Apache Airflow
Sensitive-data discoveryAmazon MacieMicrosoft PurviewSensitive Data Protection
High-performance PostgresAmazon AuroraAzure HorizonDBAlloyDB
Distributed SQLAurora DSQLAzure SQL Hyperscale (closest match)Spanner
NoSQLDynamoDBCosmos DBBigtable, Firestore
GraphNeptuneGraph in FabricSpanner Graph, BigQuery Graph
Vector storageS3 Vectors, OpenSearchAzure AI Search, Cosmos DBVector Search, AlloyDB AI
CacheElastiCache for ValkeyAzure Managed RedisMemorystore for Valkey

Models: one exclusive left

Model access converged faster than anything else in this map. Claude Opus, Sonnet, and Haiku 5.5 run on all three clouds, and so does xAI’s Grok 4.7. OpenAI’s GPT-6.2 Astra and Sol are GA on Amazon Bedrock and Foundry, while Google lists them only in private preview. Gemini is the last frontier family on a single cloud: neither the Bedrock model list nor Foundry’s catalog includes it. The clouds diverge on open-weight models and on who runs inference, which the models post traces.

Price stopped being the tiebreaker

Managed agent compute lists in a narrow band: $0.085 per vCPU-hour on Agent Platform, $0.0994 on Foundry hosted agents, and $0.1276 on AgentCore’s new default runtime, which reclaims idle memory after 120 seconds (AgentCore pricing). Each cloud bills a different unit and treats idle time differently, which the agent platforms post works through. My read is that for an agent that spends most of its life waiting on a model, tokens outweigh runtime compute. The bigger cost gap is in seat licensing.

One agent on three clouds

A sales-operations lead wants one agent that answers quarterly sales questions and files a follow-up task for the account owner. Each cloud’s recommended setup answers eight questions about that agent differently.

QuestionAWSMicrosoft AzureGoogle Cloud
Where does the agent’s identity come from?Created automatically in the account’s agent identity directory, one per AWS accountAn Entra Agent ID in the same directory as employeesA SPIFFE ID tied to the hosting resource, for agents on Google runtimes
What happens to an outside identity?Tokens from Entra or any OAuth-compatible provider are accepted as JWTsAWS and Google credentials are exchanged for Entra tokensGoogle issues an OIDC token that AWS and Azure trust (preview); Workload Identity Federation maps AWS and Azure workloads to Google principals
How does it reach the sales data?A structured knowledge base that generates SQL against RedshiftA Fabric data agent over OneLake; needs F2 capacity or higher, up to five sourcesA BigQuery data agent, grounded by Knowledge Catalog
What if the data sits on another cloud?Athena federated query connectors, including BigQuery and SnowflakeOneLake shortcuts to S3 and Google Cloud StorageThe cross-cloud lakehouse reads Iceberg tables in S3 through Glue or Unity catalogs
How do employees reach it?AWS Quick, with extensions for Teams, Outlook, and SlackFoundry agents publish to “Teams and Microsoft 365 Copilot, governed by Entra Agent ID”Gemini Enterprise, which Google calls “the new front door for AI in the workplace”
Who governs agents built elsewhere?The AWS Agent Registry indexes agents “regardless of where they’re built or hosted”; AgentCore Policy applies only to traffic through AgentCore GatewayBedrock and Google agents sync into the Microsoft Agent 365 registry through a read-only connection, with delete as the only optional actionAgent Gateway for Kubernetes outside Google Cloud is in limited preview, with the proxy self-managed
How is it billed?Per component (runtime, gateway, memory), plus Quick Professional at $20 or Enterprise at $40 per user, with a $250 per-account monthly feeM365 Copilot at $30 per user; Agent 365 at $15 per userGemini Enterprise seats; seat credits cover Gemini models only, Claude is extra
Can the agent use another vendor’s model?Yes: the AgentCore harness takes Bedrock, OpenAI, Gemini, or LiteLLM-compatible models and can switch mid-sessionFoundry: broad catalog. Copilot Studio: a curated OpenAI and Anthropic listAgent Platform: Gemini, Claude, and others. Gemini Enterprise app: Gemini by default, Claude 5.5 opt-in

What “anchored somewhere else” means

Read the table by one question: when something arrives from outside, which way does it get converted?

Azure converts it into Microsoft 365. An agent identity from AWS or Google gets exchanged for an Entra token, so Entra stays the authority. Data on S3 stays where it is, but the agent sees it only after it becomes a OneLake item queried on Fabric capacity. Agent governance is licensed per human user, and employees meet the agent in Teams and Copilot, so on Azure the agent belongs to a Microsoft 365 tenant.

On Google, the destination is BigQuery and Gemini Enterprise. Identity crosses in both directions: Google issues an agent credential that AWS or Azure can trust, and Workload Identity Federation maps AWS or Azure workloads onto Google principals. Remote data stays in S3 while BigQuery runs the query through a Google-hosted catalog. Employees reach the agent through Gemini Enterprise.

AWS anchors on the account and the gateway. Outside identities arrive as tokens; the agent’s own identity lives in a per-account directory, and rules apply only where traffic crosses AgentCore Gateway. The harness can call OpenAI or Gemini APIs directly and switch providers between turns, while structured-data questions still run on Redshift.

A third-party agent bought on each cloud’s marketplace lands in the same anchor: AgentCore Runtime or Gateway on AWS, the Microsoft 365 admin center on Azure, and Gemini Enterprise on Google, as the marketplaces post traces.

Each anchor leaves one weak row. Azure’s agent memory is still preview, and Copilot Studio won’t take an arbitrary model in its main model picker. Google’s Agent Gateway is GA only for agents on Agent Platform’s own runtimes; GKE coverage is preview and other clouds are limited preview. AWS’s builders, Quick agents and Flows plus the AgentCore harness, are configuration-driven, with no Copilot Studio-style visual canvas, and Quick lets users choose a model only in its Automate custom agents.

The anchor labels are my interpretation, and each vendor would describe its own more generously. The conversions themselves are in each vendor’s docs.

What’s missing: the agent record stays home

The anchors meet hardest at agent identity. In the trace, credentials cross clouds in every direction through federation, while the agent itself stays put: each federation maps the foreign agent onto a local stand-in, configured by hand, while the agent’s owner, approvals, and off switch stay in the cloud that hosts it. The agent identity post counts what that costs on the day one agent has to be stopped.

So what

If you are choosing a cloud for AI and data work, three questions decide it.

  1. Where do your people sign in, and where does your structured data live? A Microsoft 365 tenant pulls toward Azure. A BigQuery estate pulls toward Google. A data estate already in S3 and Redshift pulls toward AWS. Your anchor decides how much conversion you pay for.
  2. Do you need Gemini? It is the one frontier family you cannot get elsewhere. For every other model, the deciding question is who runs inference and where the prompts go.
  3. Which registry becomes the system of record for your agents? Pick where the record of who owns an agent and what it may do lives on purpose, because every other registry will hold only a partial copy, synced by hand or read-only.

This map is current as of November 18, 2026. Microsoft Ignite is running this week and AWS re:Invent starts November 30, so rows will change within weeks.

The question I have not answered: are the anchors loosening? Each cloud is now building outward from its own. Microsoft’s Fabric IQ connectors for ChatGPT and Claude, in preview, bring Power BI semantic-model insights into those apps. Google’s lakehouse reads Iceberg tables that sit in S3. AWS Quick runs inside Teams and Outlook. I can’t tell yet whether that reach makes the anchor matter less, or whether it extends each cloud’s surface while the record that governs the agent stays home.


Share this post on:


Previous Post
AgentCore and Foundry Wall Off the Session. Google Walls Off the Code.
Next Post
Three Cloud Marketplaces, Three Commit Rules, Three Front Doors