Comparing Cloud Platforms for Generative AI

How Amazon Bedrock, Microsoft Foundry and Google's Gemini Enterprise Agent Platform compare on models, agents, grounding, safety and cost, and how to choose.

Generative AI

AWS, Microsoft Azure and Google Cloud all sell the same basic thing for generative AI: hosted models behind an API, billed per token, with tools around them for grounding answers in your data, building agents and filtering unsafe content. The differences that matter are in which models each one hosts, how its identity and networking fit the rest of your estate, and which surrounding services you would otherwise have to build yourself. This article compares the three platforms on those points and ends with a way to choose.

The earlier articles in this series look at each provider on its own: AWS Offerings, Azure Offerings and GCP Offerings. Here the focus is on how they differ.

The three platforms in one paragraph each

Amazon Bedrock is AWS’s managed service for foundation models. You call models from many providers through one set of APIs, authenticate with AWS IAM, and keep traffic private with VPC endpoints. AWS’s documentation describes more than 100 models from providers including Amazon (the Nova family), Anthropic, Meta, Mistral AI, DeepSeek, OpenAI and xAI. Around the models sit Knowledge Bases for retrieval, Guardrails for content filtering, and AgentCore for running agents in production.

Microsoft Foundry is Azure’s platform for models and agents. It was called Azure AI Studio and then Azure AI Foundry before the current name. Its best-known feature is Azure OpenAI, which serves OpenAI’s GPT models from Azure regions, but the catalog also includes Anthropic’s Claude models (generally available since mid-2026), Meta Llama, Mistral, DeepSeek and xAI models. Foundry Agent Service runs agents, Foundry IQ (built on Azure AI Search) grounds them in your content, and guardrails based on Azure AI Content Safety filter inputs and outputs. Authentication is through Microsoft Entra ID.

Gemini Enterprise Agent Platform is Google Cloud’s AI platform. Until April 2026 it was called Vertex AI, and you’ll still see that name in API endpoints, client libraries and older guides. Google’s own Gemini models are the centre of it, and Model Garden adds partner models as managed APIs, including Anthropic’s Claude, Mistral, xAI’s Grok and OpenAI’s open-weight gpt-oss models. It includes RAG Engine and Vertex AI Search for grounding, a managed runtime for agents, and Model Armor for screening prompts and responses.

flowchart TB
    App["Your application or agent"] --> GW["Platform API, cloud IAM"]
    GW --> M["Hosted models"]
    GW --> R["Retrieval over your data"]
    GW --> S["Safety filters"]
    GW --> A["Agent runtime and tools"]
    M --> M1["Bedrock: Nova, Claude, Llama, GPT, others"]
    M --> M2["Foundry: GPT, Claude, Llama, others"]
    M --> M3["Agent Platform: Gemini, Claude, others"]

The shape is the same on all three. What changes is the contents of each box and how well it connects to services you already run.

Model choice

Model choice is the clearest difference, and it changes faster than anything else on these platforms.

Model familyAmazon BedrockMicrosoft FoundryGemini Enterprise Agent Platform
OpenAI GPT (proprietary)Current GPT-5.x models since June 2026, in selected regionsYes, the broadest selection outside OpenAI itselfAnnounced as a limited preview
OpenAI gpt-oss (open weight)YesYesYes
Anthropic ClaudeYesYesYes
Google GeminiNoNoYes, only here among the three
Provider’s own modelsAmazon NovaMicrosoft models such as PhiGemini, Imagen, Veo
Open models (Llama, Mistral, DeepSeek)YesYesYes

A few patterns hold even as individual versions change:

  • Gemini is exclusive to Google’s platform among the three clouds. If Gemini is the model you want, the choice is between Google Cloud and Google’s direct Gemini API.
  • Claude is on all three. That makes Anthropic’s models a reasonable default if you want to keep the option of moving clouds without changing models.
  • OpenAI’s models reach Azure first and most completely. Bedrock now carries recent GPT models, but regional availability is narrower. Exploring OpenAI’s GPT Models compares the routes to GPT in detail.

Availability also varies by region within each cloud. A model listed in the catalog may not be offered in the region where your data has to stay, so check the region before you design around a model.

Calling a model looks similar everywhere

The APIs differ in detail but follow the same pattern: authenticate with the cloud’s identity, name a model, send messages, read the reply. On Bedrock, the Converse API gives one request format across model providers:

import boto3

client = boto3.client("bedrock-runtime", region_name="us-east-1")

response = client.converse(
    modelId="MODEL_ID_FROM_THE_BEDROCK_CONSOLE",
    messages=[{"role": "user", "content": [{"text": "Explain VPC peering in two sentences."}]}],
    inferenceConfig={"maxTokens": 300},
)
print(response["output"]["message"]["content"][0]["text"])

Foundry accepts the standard OpenAI client for Azure OpenAI deployments, and Google’s Gen AI SDK (google-genai) calls Gemini on Google Cloud when you set it to use your project and location. Credentials come from the environment in all three cases: an IAM role, a managed identity or a service account, rather than a key in code.

The practical lesson is to keep the model ID and endpoint in configuration. Model versions are retired on a schedule on every platform, and code with a hard-coded model name breaks when that happens.

Your team has standardised on Gemini models. Which cloud platform can host them?

  • Any of the three, through their model catalogs
  • Amazon Bedrock and Google Cloud
  • Only Google Cloud's Gemini Enterprise Agent Platform

Correct! Gemini is available on Google Cloud and through Google's own Gemini API, but not on Bedrock or Foundry.

Grounding, agents and safety

A model alone rarely makes a useful application. Most production systems also need retrieval over private documents, a way to let the model call tools, and filters on what goes in and out. Each platform has a managed answer for each.

CapabilityAWSAzureGoogle Cloud
Retrieval (RAG)Bedrock Knowledge BasesFoundry IQ, Azure AI SearchRAG Engine, Vertex AI Search
Agent runtimeBedrock AgentCoreFoundry Agent ServiceAgent Platform’s managed agent runtime
Content filteringBedrock GuardrailsFoundry guardrails (Azure AI Content Safety)Model Armor, Gemini safety settings
IdentityAWS IAMMicrosoft Entra IDGoogle Cloud IAM
Monitoring and auditCloudWatch, CloudTrailAzure MonitorCloud Monitoring, Cloud Logging
Private networkingVPC endpoints (PrivateLink)Private endpointsPrivate Service Connect

Retrieval

All three follow the same retrieval-augmented generation pattern: split documents into chunks, store embeddings in a vector index, find the chunks closest to the question, and pass them to the model with an instruction to answer from them. The managed services take care of ingestion, chunking and the index, so the decision is mostly about where your documents already live. Bedrock Knowledge Bases reads from S3 and other connectors, Azure AI Search sits naturally next to SharePoint and Microsoft 365 content, and Google’s search and RAG services connect to Cloud Storage, BigQuery and Google Workspace. Moving a large document set across clouds just to reach a retrieval service adds egress cost and a second copy to govern.

sequenceDiagram
    participant U as User
    participant App as Application
    participant Ret as Retrieval service
    participant G as Safety filter
    participant LLM as Model
    U->>App: Question
    App->>G: Check input
    App->>Ret: Search private documents
    Ret-->>App: Relevant chunks
    App->>LLM: Instructions, chunks, question
    LLM-->>App: Draft answer
    App->>G: Check output
    App-->>U: Answer with sources

Agents

Agents add tool calls to that loop: the model decides to look something up, call an API or run code, then continues with the result. The managed runtimes handle sessions, memory, identity for tool calls and tracing. Bedrock AgentCore became generally available in October 2025 and works with agents written in open-source frameworks, not just AWS’s own. Foundry Agent Service can publish agents into Microsoft 365 Copilot and Teams, which matters if that’s where your users already work. Google’s platform pairs its agent runtime with the open-source Agent Development Kit.

The common risk is the same everywhere. An agent can only be trusted with the permissions you give its identity, and text it reads from a web page or a document can carry instructions that try to override yours. Grant each agent the narrowest role that works, and require a person to confirm actions that change or delete things.

Safety filters

Bedrock Guardrails, Foundry guardrails and Model Armor all screen prompts and responses for harmful content, prompt injection attempts and sensitive data such as personal information. They are configured separately from the models, so the same policy can apply whichever model you call. None of them is a complete defence. They reduce risk; they don’t replace least-privilege access, output validation and logging.

Your documents are in SharePoint and your models are on another cloud. What is usually the first thing to consider for retrieval?

  • Which retrieval service sits closest to the documents, to avoid copying them across clouds
  • Which provider's vector index has the most dimensions
  • Retrieval services only work with models from the same cloud, so move the documents

Correct! Keeping retrieval next to the data avoids egress charges and a second copy of sensitive documents to secure.

Developer and business assistants

The providers also sell finished assistants built on these platforms. They are bought per user rather than per token, and they are often the first generative AI service an organisation actually pays for.

UseAWSMicrosoftGoogle
Coding assistantKiro (Amazon Q Developer’s IDE plugins reach end of support on April 30, 2027)GitHub CopilotGemini Code Assist
Assistant in the cloud consoleAmazon Q Developer in the AWS consoleAzure CopilotGemini Cloud Assist
Assistant for business usersAmazon QuickMicrosoft 365 CopilotGemini Enterprise

Older articles, including an earlier version of this one, use names that no longer exist: CodeWhisperer became Amazon Q Developer, Duet AI became Gemini for Google Cloud, Bard and PaLM 2 gave way to Gemini, Google Agentspace became Gemini Enterprise, and Vertex AI became Gemini Enterprise Agent Platform. If a tutorial uses one of the old names, check the current documentation before following it.

Cost

Per-token prices change several times a year, differ by region and model version, and are usually lower for the newest small models than for last year’s flagships. A price table in an article goes out of date within months, so this one doesn’t include one. Use the official pricing pages: Amazon Bedrock, Azure OpenAI and Gemini Enterprise Agent Platform.

What you can compare is how each platform lets you buy capacity:

OptionWhat it isWhen it fits
On-demand, per tokenPay for input and output tokens as you goPrototypes, spiky or unpredictable traffic
BatchSubmit many requests and collect results later, at a discount on supported modelsSummarising archives, classification, evaluation runs
Discounted, lower-priority tiersBedrock’s Flex tier and similar options trade latency for priceBackground work that can wait
Reserved or provisioned throughputPay for dedicated capacity for a termSteady, high-volume production traffic with latency targets
Prompt cachingRepeated prompt prefixes billed at a reduced rateLong fixed instructions or reference text sent on every call

Tokens are rarely the only cost. Vector indexes, search units, agent runtime hours, logging, storage and data transfer all appear on the bill, and on a retrieval-heavy application the search service can cost as much as the model. Estimate the whole system, set budget alerts before launch, and rate-limit public endpoints per user so one script can’t spend the month’s budget in an afternoon.

Choosing a platform

Model quality matters less to the choice than it seems, because the leading model families are close and change places often. These questions usually decide it:

  1. Where do your data and identities already live? Retrieval, private networking and access control are simplest on the cloud that already holds your documents and users.
  2. Do you need a specific model family? Gemini means Google Cloud. The widest set of OpenAI models means Azure. Claude, Llama and Mistral are available everywhere.
  3. Where do your users work? Agents that must appear inside Microsoft 365 or Teams point to Foundry; Google Workspace users point to Gemini Enterprise.
  4. Which regions must you use? Check that the exact models you need are offered in the regions your compliance rules allow.
flowchart TD
    Start{"Must you use Gemini?"} -->|Yes| GCP["Gemini Enterprise Agent Platform"]
    Start -->|No| Data{"Where do your data and users live?"}
    Data -->|Microsoft 365 and Entra ID| AZ["Microsoft Foundry"]
    Data -->|Mostly AWS| AWS["Amazon Bedrock"]
    Data -->|Google Cloud and Workspace| GCP
    Data -->|Spread across clouds| Port{"Need to switch models or clouds later?"}
    Port -->|Yes| Claude["Model available on all three, behind your own abstraction"]
    Port -->|No| Home["Platform next to the largest data set"]

Running generative AI on more than one cloud is possible, and it is common in organisations that already work that way. It also doubles the identity, networking, safety configuration and monitoring you have to maintain. Unless a specific model or data residency requirement forces it, start on one platform and keep your application portable: model IDs in configuration, a thin wrapper around the model call, and prompts stored as files you can test against another model. Choosing a Cloud Provider: One Cloud or Several covers that trade-off for cloud services in general, and AWS vs. GCP vs. Azure: Core Differences compares the three providers beyond AI.

You want the freedom to move clouds later without rewriting prompts for a new model. Which approach helps most?

  • Use each cloud's own first-party model family
  • Use a model family offered on all three clouds and keep the model ID in configuration
  • Hard-code the model call so it is easier to read

Correct! A model available everywhere, called through a thin wrapper with configurable IDs, keeps a later move cheap.

Further reading

Key takeaways

  • Bedrock, Microsoft Foundry and Gemini Enterprise Agent Platform offer the same building blocks: hosted models, retrieval, agent runtimes, safety filters and cloud identity.
  • Gemini runs only on Google Cloud among the three, Azure has the broadest OpenAI lineup, and Claude, Llama and Mistral are available on all of them.
  • Put retrieval and agents on the cloud that already holds your data and users; moving documents across clouds adds cost and governance work.
  • Don’t trust printed token prices, including old ones in tutorials. Check the pricing pages and budget for search, storage and agent runtime as well as tokens.
  • Keep model IDs in configuration and the model call behind a thin wrapper, because model versions are retired on every platform.

Part 28 of 30 · Introduction to Generative AI

  1. 01Introduction to Generative AI: What Is It?
  2. 02The History of Generative AI: From Concept to Reality
  3. 03Key Definitions and Concepts in Generative AI
  4. 04Pioneers of Generative AI: Meet the Inventors
  5. 05Understanding Neural Networks: The Backbone of AI
  6. 06Deep Learning Basics: Layers and Learning Rates
  7. 07Generative vs. Discriminative Models: What's the Difference?
  8. 08Overview of Generative Models: VAEs, GANs, and More
  9. 09Introduction to Variational Autoencoders (VAEs): From Basics to Advanced Applications
  10. 10Introduction to Generative Adversarial Networks (GANs)
  11. 11The Role of Transformers in Generative AI
  12. 12Applications of Generative AI in Everyday Life
  13. 13Ethical Considerations in Generative AI: Navigating the New Frontier
  14. 14Generative AI in Art and Music: The Creative Revolution
  15. 15Case Study: AI-Generated Artwork
  16. 16Generative AI in Text Generation
  17. 17Exploring OpenAI's GPT Models: Word Up! Bot
  18. 18Real-World Use Cases: Deepfakes and Their Impact
  19. 19Generative AI in Image and Video Synthesis
  20. 20Introduction to Reinforcement Learning in Generative AI
  21. 21The Future of Generative AI: Trends to Watch
  22. 22Generative AI in Healthcare: Opportunities and Challenges
  23. 23Generative AI in Finance: Risk Modeling and Beyond
  24. 24Generative AI in Marketing and Personalization
  25. 25Cloud Services for Generative AI: AWS Offerings
  26. 26Cloud Services for Generative AI: GCP Offerings
  27. 27Cloud Services for Generative AI: Azure Offerings
  28. 28Comparing Cloud Platforms for Generative AI
  29. 29Recap and Q&A: Introduction to Generative AI
  30. 30Building Your First Generative AI Model: A Practical Guide
Ready to check your understanding?Take the Word Up! Bot quiz
esc

Type to search all articles · ↑↓ to move · ↵ to open