# Cloud AI migrations: Azure OpenAI / AI Foundry, AWS Bedrock, Google Vertex AI, Direct model APIs, Self-hosted open models compared

> Pick where your AI workloads run today and where you want to go. We’ll show why teams make that move, what to watch out for and how we de-risk it.

Source: https://aibyos.com/platforms/migrations

⇆ Platforms & Cloud

# Cloud AI migrations

Moving AI workloads between Azure OpenAI, AWS Bedrock, Google Vertex AI, direct APIs and self-hosted models.

[Talk to an expert](https://aibyos.com/contact)

Pick where your AI workloads run today and where you want to go. We’ll show why teams make that move, what to watch out for and how we de-risk it.

[Open the interactive planner →](https://aibyos.com/platforms#guide-migrations)

### Azure OpenAI / AI Foundry

##### Why teams move here

-   Microsoft enterprise agreement or Azure commitment to consume
-   OpenAI models with Microsoft enterprise, compliance and EU data-boundary terms
-   Native fit with Entra ID, Microsoft 365, Copilot and Fabric

##### When leaving it, watch for

-   Replace Azure-specific content-filter configurations with equivalent guardrails
-   Provisioned throughput (PTU) commitments and quotas to wind down
-   Features tied to Azure AI Search or Foundry agents need an equivalent on the target

### AWS Bedrock

##### Why teams move here

-   Consolidate on AWS and use existing commitments and security controls
-   Wide model choice in one API: Anthropic Claude, Meta Llama, Mistral, Amazon Nova and more
-   VPC endpoints, IAM and Bedrock Guardrails in the same account model as your apps

##### When leaving it, watch for

-   Bedrock Knowledge Bases and Agents are managed features; rebuild or map them
-   Provisioned throughput contracts to plan around
-   IAM-based auth patterns change to the target’s identity model

### Google Vertex AI

##### Why teams move here

-   Gemini models with long context and strong multimodal capabilities
-   Data gravity: analytics already live in BigQuery
-   Partner models (including Claude) with Google Cloud commitments and VPC Service Controls

##### When leaving it, watch for

-   Vertex AI Search / Agent Builder features need an equivalent on the target
-   Gemini-specific capabilities (very long context, native multimodal) must be re-validated
-   Service-account-based auth and perimeter rules to re-create

### Direct model APIs (OpenAI, Anthropic, Mistral…)

##### Why teams move here

-   Newest models and features usually arrive first
-   Simplest setup for startups and fast-moving product teams
-   Fewer cloud quota requests and regional constraints

##### When leaving it, watch for

-   Rate limits and data-processing terms differ from the cloud-hosted versions
-   Provider-specific APIs (e.g. built-in tools, file search) need equivalents
-   Billing and usage tracking move into your cloud bill and gateway

### Self-hosted open models

##### Why teams move here

-   Full data sovereignty: nothing leaves your environment
-   Lower unit cost at high, steady volume
-   Deep customisation through fine-tuning and distillation

##### When leaving it, watch for

-   You give up operating GPUs; capacity planning becomes the provider’s job
-   Fine-tuned weights may not be portable; managed fine-tuning may be needed
-   Check data-processing terms meet the same requirements as self-hosting

## What always changes

-   Model mapping and prompt re-tuning: the same prompt behaves differently on another model
-   Embeddings are not portable: vector stores must be re-embedded and re-indexed
-   SDKs, auth, networking (private endpoints) and IAM differ per provider
-   Quotas, throughput reservations and regional availability
-   Safety filters and guardrail settings
-   Fine-tuned models must be re-trained on the target
-   Logging, monitoring and cost attribution

## How we de-risk it

1.  Assess: inventory every model call, prompt, embedding and provider-specific feature
2.  Decouple: route traffic through an AI gateway / abstraction layer
3.  Evaluate: run old and new side by side on golden sets and shadow traffic
4.  Cut over: phased by feature or tenant, with instant rollback
5.  Optimise & decommission: tune cost and latency, close old resources and contracts