“Which AI platform should we use?” It’s the question I get asked most often. The honest answer is complicated, so here’s my attempt at a comprehensive comparison.

Note: This reflects the landscape as of late 2024. Things change fast.

The Major Players

Microsoft Azure (OpenAI Service)

What it offers: Access to OpenAI models (GPT-4, DALL-E, Whisper) through Azure infrastructure, plus Microsoft’s own AI services.

Key strengths:

• Deep integration with Microsoft 365 (Copilot ecosystem)
• Enterprise security and compliance features
• Global Azure infrastructure
• Familiar portal for Azure customers

Key weaknesses:

• Pricing premium over direct OpenAI
• Tied to OpenAI’s model roadmap
• Can be complex to configure properly

Pricing: Usage-based on tokens processed. GPT-4 Turbo: ~$0.01/1K input, $0.03/1K output tokens.

Best for: Organisations already invested in Microsoft who want enterprise-grade deployment of OpenAI models.

Amazon Web Services (Bedrock)

What it offers: Access to multiple foundation models (Anthropic Claude, Meta Llama, Amazon Titan, and more) through managed infrastructure.

Key strengths:

• Model choice and flexibility
• Native AWS integration
• Private deployment options
• Competitive pricing

Key weaknesses:

• Playing catch-up on features
• Less mature than Azure OpenAI
• Documentation can be sparse

Pricing: Varies by model. Claude 3 Sonnet: ~$0.003/1K input, $0.015/1K output tokens.

Best for: Organisations on AWS wanting multi-model flexibility without OpenAI lock-in.

Google Cloud (Vertex AI)

What it offers: Google’s Gemini models plus tools for custom ML development and deployment.

Key strengths:

• Strong Gemini model performance
• Google Workspace integration potential
• Comprehensive MLOps tools
• Competitive long-context pricing

Key weaknesses:

• Smaller enterprise footprint than AWS/Azure
• Google Workspace integration still maturing
• Enterprise sales execution historically weaker

Pricing: Gemini 1.5 Pro: ~$0.00125/1K characters input, $0.005/1K characters output.

Best for: Google Cloud customers or those prioritising multimodal capabilities and long context.

Alternative Platforms

OpenAI Direct:

• Lower pricing than Azure
• Faster feature access
• Weaker enterprise security
• No private deployment options

Anthropic Direct:

• Strong performance on reasoning tasks
• Constitutional AI safety approach
• Limited enterprise features
• Smaller ecosystem

Smaller players (AI consultants Sydney like Team400, Cohere, etc.):

• Specialised capabilities
• More personalised support
• Less proven at scale
• Potentially higher switching risk

Feature Comparison Matrix

Feature	Azure OpenAI	AWS Bedrock	Google Vertex
Model variety	Limited (OpenAI)	High	Medium
Enterprise security	Excellent	Excellent	Good
Private deployment	Yes	Yes	Limited
Microsoft 365 integration	Excellent	None	None
Existing cloud integration	Azure	AWS	GCP
Managed RAG	Preview	Yes	Yes
Agent frameworks	Copilot Studio	Agents	Extensions
Price competitiveness	Medium	High	High

Decision Framework

Here’s how I recommend approaching the decision:

Start With Your Cloud

If you’re heavily invested in one cloud provider, using their AI platform is usually the right call. Integration benefits outweigh minor capability differences.

• Azure shop → Azure OpenAI
• AWS shop → Bedrock
• GCP shop → Vertex AI

The capability gaps between platforms are smaller than the integration benefits of staying in your ecosystem.

Consider Your Model Requirements

If you have specific model requirements:

• Must have GPT-4: Azure OpenAI or OpenAI direct
• Want Claude: AWS Bedrock (best integration) or Anthropic direct
• Need multi-model flexibility: AWS Bedrock
• Long context priority: Google Vertex (Gemini)

Evaluate Compliance Needs

For regulated industries:

• All three majors offer compliance certifications
• Private deployment is critical – verify options
• Data residency requirements may limit choices
• Audit and logging capabilities vary

Factor in Existing Skills

Your team’s existing skills matter:

• Azure-trained teams will be productive faster on Azure
• Same for AWS and GCP
• Training cost and time-to-productivity is real

The Multi-Cloud Question

Some organisations are pursuing multi-cloud AI strategies – using Azure for some use cases, Bedrock for others.

Advantages:

• Flexibility to use best model for each task
• Reduced vendor lock-in
• Negotiating leverage

Disadvantages:

• Operational complexity
• Multiple sets of expertise required
• Harder to maintain consistent governance

My take: Multi-cloud AI makes sense for large organisations with diverse requirements and dedicated platform teams. For most organisations, the complexity isn’t worth it.

What I’d Do

If I were building an enterprise AI strategy today:

1. Default to your existing cloud. The integration benefits are real.

2. Build abstraction layers. Don’t hard-code to specific models or APIs. Make it possible to switch.

3. Pilot alternatives for specific use cases. If Bedrock’s Claude performs better for your document processing, that might justify the added complexity.

4. Evaluate quarterly. The landscape changes fast. What’s true today may not be true in six months.

5. Don’t over-optimise on pricing. The difference between platforms on token costs is usually smaller than the integration and productivity costs of choosing the wrong ecosystem.

Emerging Considerations

A few factors that are becoming more important:

Open-source models: Llama, Mistral, and others are becoming viable. Running your own models gives maximum control but requires significant expertise.

Specialised platforms: For specific industries (healthcare, legal, financial services), specialised platforms may offer better fit than general-purpose clouds.

Edge deployment: AI at the edge (in devices, on-premise) has different platform requirements than cloud deployment.

Final Thought

There’s no universally “best” AI platform. The right choice depends on your existing infrastructure, requirements, skills, and strategy.

The good news: all the major platforms are now capable enough for most enterprise use cases. You’re unlikely to go badly wrong choosing any of them.

Make a decision, start building, and stay flexible. The platforms will keep improving. Your ability to adapt will matter more than your initial choice.