Every second week I sit in a room where someone senior asks the same question: which cloud should we build our AI on? Then someone name-drops a model, someone else name-drops a competitor, and forty minutes later nobody has decided anything. Here is how I actually cut through it.
For most Australian organisations, the best cloud AI platform is the one you already run production on: AWS, Azure and Google Cloud are now close enough on models and pricing that data gravity, not model bragging rights, should decide it.
Let me unpack why, because the reasoning matters more than the verdict.
What are we actually comparing here?
We are comparing the three platforms for building AI into your own applications, not the consumer chat apps sitting on top of them.
This is the first place these conversations go wrong. ChatGPT, Claude.ai and the Gemini app are consumer products. The cloud AI platforms are the engines you call from your own code to build features: a support agent that reads your knowledge base, a document extractor for finance, a summariser inside your CRM. The three that matter for a CIO or CTO are:
- Amazon Bedrock on AWS
- Azure AI Foundry (which absorbed the old Azure AI Studio and Azure OpenAI Service) on Microsoft Azure
- Google Vertex AI on Google Cloud
Each gives you a governed, billable, enterprise-grade way to call foundation models with your identity, networking and logging already wired in. That is the whole point. You are not buying a model. You are buying the plumbing around it.
What is Amazon Bedrock?
Amazon Bedrock is a fully managed service that lets you call more than a hundred foundation models through a single API without running any infrastructure.
Bedrock’s pitch is breadth and neutrality. As of mid-2026 it carries models from around eighteen providers, including Anthropic’s Claude family, Amazon’s own Nova models, Meta Llama, Mistral, Cohere, DeepSeek and OpenAI’s open-weight releases. In February 2026 AWS added the latest open-weight models to the Sydney region, which is a meaningful signal for anyone who needs inference to stay onshore.
Pricing is per token, quoted per million input and output tokens, with output charged higher than input. There are five billing modes worth knowing: on-demand pay-per-token, batch inference at a discount, provisioned throughput for reserved capacity, prompt caching for repeated inputs, and model customisation for fine-tuning.
One clarification I make constantly: Bedrock is not SageMaker. Bedrock is for calling models. SageMaker is the full platform for training and hosting your own. If you are doing genuine custom model work, that is a different tool, and I have written a full walkthrough of that path in end-to-end machine learning on AWS SageMaker with Claude Code.
Azure AI Foundry vs Google Vertex AI
Azure leads if your shop is Microsoft-native and wants first-class access to OpenAI’s GPT models; Vertex AI leads if you want Google’s Gemini models and strong data and analytics tooling underneath.
Azure AI Foundry is the rebrand and consolidation of what used to be Azure AI Studio, the model catalogue and Azure OpenAI Service, now under one roof. Its headline advantage is being the enterprise home for OpenAI’s GPT-5 family, sold directly by Microsoft alongside models from DeepSeek, xAI’s Grok, Meta, Mistral and Black Forest Labs. If your organisation already lives in Microsoft 365, Entra ID and Azure, Foundry is the path of least resistance. Pricing is consumption-based per token, published openly, with provisioned throughput units for sustained workloads.
Google Vertex AI is the counterweight. It is the enterprise front door to the Gemini family, and Google’s published Gemini pricing is competitive across the range, from cheap Flash-tier models up to the flagship Pro models. Vertex earns its keep when your data already sits in BigQuery, because the analytics-to-AI pipeline is genuinely tight. Vertex also pairs with Google AI Studio, the free browser-based prototyping environment. AI Studio is free to experiment in with daily limits, and the model price is identical whether you call from AI Studio or Vertex. What you pay Vertex for is the enterprise wrapper: SLAs, compliance, provisioned throughput and data residency controls, not the model itself.
The comparison at a glance
| Platform | Flagship AI service | Models offered | Best for | Australian region |
|---|---|---|---|---|
| AWS | Amazon Bedrock | Claude, Amazon Nova, Llama, Mistral, DeepSeek, OpenAI open-weight (100+ total) | Model choice and neutrality; teams already on AWS | Asia Pacific (Sydney) |
| Microsoft Azure | Azure AI Foundry (with Azure OpenAI) | GPT-5 family, Grok, Llama, DeepSeek, Mistral, FLUX | Microsoft-native shops wanting OpenAI models | Australia East (Sydney), Australia Southeast (Melbourne) |
| Google Cloud | Vertex AI (with Google AI Studio) | Gemini family, plus Claude and Llama via the model garden | Data-heavy teams on BigQuery; strong prototyping | Sydney, Melbourne |
Which cloud AI platform is best for enterprise?
The honest default is to build AI on the cloud you already run, and only break that rule for a specific, defensible reason.
I know that sounds anticlimactic. It is also correct, and here is the reasoning I walk boards through.
Your data has gravity. If your customer records, transactions and documents already sit in AWS, then building your AI layer in AWS means no cross-cloud data movement, no second security review, no duplicated networking, and one identity model instead of two. The moment you pick a different cloud for AI than the one your data lives in, you sign up for an egress bill, a latency penalty and a governance headache that recurs forever. That tax almost always outweighs the marginal quality difference between this month’s frontier models.
The models converge; the plumbing does not. All three platforms now offer excellent frontier models at similar per-token prices. The differences that persist are the boring, durable ones: how identity works, how networking is configured, how billing rolls up, how your existing team’s skills transfer. Those are the things that decide whether a project ships in six weeks or six months.
So when should you break the default and go multi-cloud or switch? A few legitimate reasons:
- A specific model you cannot get elsewhere. If your use case genuinely depends on one provider’s model and it is only first-class on one cloud, that is a real reason.
- A pricing cliff at scale. At very high volume, a meaningful per-token difference on your specific workload can justify the switch. Benchmark on your real traffic, not a vendor slide.
- Sovereignty or compliance constraints that one platform meets and another does not.
- Avoiding lock-in deliberately as a strategic hedge, accepting the operational cost with eyes open.
If none of those apply, resist the urge to be clever. Building an AI capability that ships repeatedly beats a perfectly optimised architecture that never leaves the whiteboard. I have argued this at length in why AI factories beat AI projects: the organisations that win are the ones that industrialise delivery, not the ones that agonise over vendor selection.
The Australian data sovereignty angle
All three platforms can keep your inference inside Australia, but Australian data residency is not the same thing as data sovereignty, and in 2026 that distinction has teeth.
The good news first. Every one of the big three now has genuine Australian regions. Amazon Bedrock runs in Asia Pacific (Sydney) and supports cross-region inference within Australia. Azure AI Foundry and Azure OpenAI run in Australia East (Sydney) and Australia Southeast (Melbourne). Google Vertex AI runs in both Sydney and Melbourne, with per-region data residency commitments in the Data Processing Addendum. So for most regulated workloads, you can pin storage and inference to Australian soil. Watch the fine print on region parity, though: some high-demand services still land in Sydney before Melbourne, so do not assume the two regions match feature for feature.
Now the part that catches people out. Residency means your data is stored and processed here. Sovereignty means it is also beyond the reach of foreign legal compulsion. Those are different guarantees. A US hyperscaler can still be compelled under the US CLOUD Act to hand over data regardless of where it physically sits. With Privacy Act reform tightening expectations through 2026, that gap is exactly the sort of thing an auditor or a regulator in financial services, health or government will probe. If you operate under those constraints, ask your provider the sovereignty question explicitly, in writing, and do not let a residency answer stand in for it.
Key takeaways
- The three cloud AI platforms have converged on models and pricing. All bill per token, all carry frontier models, all run in Australian regions. The model wars are close to a draw at the point of purchase.
- Data gravity should drive the decision. Build on the cloud your data already lives on unless you have a specific, defensible reason not to.
- Bedrock, Foundry and Vertex are for building AI into your applications, not for consumer chat. Bedrock is not SageMaker, and Vertex is not the free AI Studio prototype tier.
- Residency is available on all three; sovereignty is a separate conversation. In regulated Australian industries, ask about the CLOUD Act explicitly.
- Shipping beats optimising. The winning organisations industrialise delivery rather than perfecting vendor choice.
If you are still torn on the model layer specifically, my rundown of the best AI agent platforms and frameworks for 2026 goes a level deeper on what actually works in production.
I help Australian organisations choose where to build their AI and then actually ship it. If you are staring down this decision, get in touch.