Back to Blog

Azure AI vs AWS AI vs Google Cloud AI - An Australian Comparison

April 26, 2026•10 min read•Michael Ridland

Every second discovery call I take starts with some version of the same question. "We're already on Azure, but our data science team likes Google. AWS has the cheapest GPUs. What do we actually do?"

There is no clean answer. There is a right answer for your specific situation, though, and after building production AI on all three clouds for Australian clients across health, financial services, and government, I have strong opinions about when each one is the right call.

This post is the version of the comparison I wish someone had written for me three years ago. Pricing in AUD where I can give it, real residency considerations, the politics that show up in procurement, and the bits the marketing pages skip.

The short version if you only read one section

If you are running Microsoft 365, Dynamics, or any meaningful .NET footprint, Azure AI is the default and the right one nine times out of ten. The integration savings alone pay for any per-token premium.

If your team is deeply on AWS already, your data sits in S3, and you want the broadest catalogue of foundation models (Claude, Llama, Mistral, Amazon's own) in one place with the same IAM you already use, AWS Bedrock is the answer.

If you are building a Gemini-first product, doing heavy multimodal work with video, or your team is research-led and wants the best raw model capability for certain tasks, Google Cloud Vertex AI is worth the friction.

For most Australian mid-market and enterprise customers we work with, the order is Azure first, AWS second, Google third. That is not a marketing position. It is what the actual procurement, integration, and skills picture in Australia looks like in 2026.

The model catalogues right now

This changes monthly so take exact lists with a pinch of salt, but the shape has been stable.

Azure AI Foundry gives you OpenAI's full lineup (GPT-5, GPT-4.1, the o-series reasoning models, embeddings, image and audio), plus a growing set of partner models including Mistral, Cohere, Meta Llama, and the Phi small models from Microsoft Research. The OpenAI exclusivity is the headline feature. If your business case requires GPT-5 or o1, Azure is the only enterprise-grade way to consume it under an EA.

AWS Bedrock leans on Anthropic Claude (Sonnet 4.6 and Opus 4.7 are both there), Meta Llama, Cohere, Mistral, AI21, Stability, and Amazon's own Titan and Nova families. Bedrock's main pitch is "pick a model, change one line of code, switch later." That flexibility is real and matters for teams who want to A/B test models against the same workload.

Google Cloud Vertex AI is Gemini-first. Gemini 2.5 Pro and Flash sit alongside Anthropic Claude, Llama, Mistral, and a strong open-source library through the Model Garden. Google's edge is multimodal, especially video understanding and long context windows, plus the integration with BigQuery for AI-on-your-warehouse patterns.

Data residency and the Australian reality

This is where the conversation usually gets serious, especially for federal government, state government, healthcare, and APRA-regulated financial services.

Azure has the strongest Australian footprint. Australia East (Sydney), Australia Southeast (Melbourne), and the Australia Central regions in Canberra including the IRAP PROTECTED zones. Azure OpenAI is available with Australia East residency for most of the popular models, though new model releases sometimes take a few months to land in-region. If you need PROTECTED data handling, Azure is essentially the only mature option.

AWS has Sydney and Melbourne regions. Bedrock is in Sydney, with most models available locally and the rest a quick API call to US-East if you can accept egress. AWS has IRAP assessments for many services but the AI services lag a little behind Azure for in-region certified availability.

Google Cloud has Sydney and Melbourne too. Vertex AI is in Sydney, though some of the newer Gemini features still ship US-first by a quarter or two. IRAP coverage is improving but if your compliance team has a list, Google often has more "assessed" than "certified" boxes ticked.

One client in the insurance space failed their first vendor review with us because we suggested Google Vertex for a customer-facing chatbot. The data was non-sensitive marketing content, but their security team had a blanket "Australian-region OpenAI-grade certification" requirement that effectively meant Azure. We moved the workload across in three weeks and the project shipped. The lesson: get your data classification and compliance team in the room before you pick the cloud, not after.

Pricing the way it actually shows up on invoices

Pricing AI services is a moving target and I won't pretend to give you precise per-token rates because they change every few months. Here are the shapes that matter.

Azure OpenAI is priced per 1K tokens, with separate input and output rates, and you can choose Pay-as-you-go or Provisioned Throughput Units (PTUs). PTUs are the enterprise pattern - you reserve capacity and get predictable latency at a flat monthly rate. PTU pricing starts roughly around AUD 3,000-4,000 per month per unit for the cheaper models and runs into tens of thousands for the larger ones. For most Australian SMEs, PAYG is fine until you hit consistent high volumes.

AWS Bedrock has on-demand pricing per token across all models, plus Provisioned Throughput for steady workloads. Claude Sonnet on Bedrock prices comparably to Azure OpenAI's GPT-4.1 class. Bedrock's hidden value is that you can move a workload to a cheaper model (Llama, Nova, Mistral) without rewriting your application.

Google Vertex AI prices Gemini Flash very aggressively. For high-volume, low-complexity tasks like classification, extraction, and summarisation, Flash is often the cheapest production option of the three by a meaningful margin. Gemini Pro is competitive with the top tier on Azure and AWS.

A rough rule we use when scoping: budget AUD 5,000-15,000 per month for a real production AI application doing low millions of tokens, before you optimise. The actual cloud bill is rarely the deciding cost factor. The team time to integrate, secure, observe, and improve the system is 5-10x bigger.

A practical comparison table

Dimension Azure AI AWS Bedrock Google Vertex AI
Best foundation model GPT-5, o-series (OpenAI exclusive) Claude Opus 4.7, Llama Gemini 2.5 Pro
Cheapest high-volume model GPT-4.1 mini Nova Lite, Llama Gemini 2.5 Flash
Australia in-region availability Strongest (East, Southeast, Central) Strong (Sydney, Melbourne) Good (Sydney, Melbourne)
IRAP PROTECTED ready Yes (Australia Central) Partial Partial
Integration with M365 / Dynamics Native Add-on Add-on
Integration with S3 / Lambda / IAM Add-on Native Add-on
Integration with BigQuery / Workspace Add-on Add-on Native
Model swap flexibility Limited (within Foundry) Best in class Good
Agent framework maturity Microsoft Agent Framework, AutoGen Bedrock Agents, Strands Vertex Agent Builder, ADK
Local consulting depth in Australia Largest pool Large pool Smaller but growing

When Azure AI is the right call

You should default to Azure when:

  • You already pay for Microsoft 365 or Dynamics. Copilot connectors and Power Platform integration give you weeks of work for free.
  • You need IRAP PROTECTED hosting in Canberra. Few other options exist.
  • Your business case depends on GPT-5 or o-series reasoning models specifically.
  • Your team is .NET-heavy. The Microsoft Agent Framework SDK and Semantic Kernel are well-supported and well-documented.
  • You want a single Microsoft procurement and an EA discount applied to AI consumption.

We do most of our Azure AI consulting and Azure AI Foundry work for exactly these reasons. If you are building a Copilot extension, a Power Platform agent, or anything that touches SharePoint, Outlook, or Teams data, Azure is the path that does not fight you.

When AWS Bedrock is the right call

Pick AWS when:

  • Your data plane lives in S3, Aurora, or Redshift and you have no appetite to move it.
  • You want to A/B test Claude vs Llama vs Nova on the same workload using one API.
  • You have AWS-experienced engineers and you do not want to retrain them.
  • Cost optimisation through model selection matters more than picking a single best model.
  • You need Bedrock Guardrails or AWS-native PII detection because your compliance posture is built on AWS services.

I have seen Bedrock work very well for product companies who already shipped on AWS and want to add an AI feature without a cloud migration. Claude Sonnet 4.6 on Bedrock is one of the most pragmatic enterprise AI choices in the market right now.

When Google Cloud Vertex AI is the right call

Pick Google when:

  • You are building a multimodal product that does serious work with video or images.
  • Your data warehouse is BigQuery and you want AI close to it.
  • You need the longest context window and you have actually hit Azure or AWS context limits.
  • Your team is research-led and comfortable with newer tooling.
  • You want Gemini Flash specifically for high-volume cheap inference.

The catch in Australia is the smaller local consulting market for Google Cloud AI. Sydney has decent depth, but if you are in Brisbane, Perth, or Adelaide, finding senior Vertex engineers is harder than finding senior Azure or AWS ones. Plan for that in your hiring.

Common objections and where they hold up

"Azure is more expensive than the alternatives." Not really, once you factor in your existing EA discount and the integration time you save. The per-token rates are within 10-20% across the three providers for equivalent models.

"AWS has the biggest model library so we have the most options." True, but most teams pick one model and stick with it for 6-12 months. The optionality is real but rarely exercised in practice.

"Google's Gemini is the smartest model." Depends on the benchmark and the task. For some coding and long-context work, Gemini 2.5 Pro is excellent. For agentic workflows and tool use, Claude and OpenAI are still ahead in our testing.

"We should multi-cloud our AI." No. Pick one as primary and use a second only where it has a specific, measurable advantage for a specific workload. Multi-cloud AI architectures look elegant on whiteboards and cost twice as much to operate.

A simple decision framework

Ask these five questions in order:

  1. Do we have a hard residency or compliance requirement (PROTECTED, APRA CPS 234 strict interpretation, health data)? If yes, Azure is likely the only answer.
  2. What is our current cloud commitment in dollars? Whichever one is biggest is the starting point.
  3. Do we need a specific model (GPT-5, Claude Opus 4.7, Gemini 2.5 Pro) for our use case? Pick the cloud that hosts it natively.
  4. What does our engineering team know already? Skills determine velocity more than tools do.
  5. What is our cost ceiling at scale? If you are projecting tens of millions of tokens per day, run a real benchmark before you commit.

If those five answers do not converge on one cloud, you probably do not have a strong enough business case yet. Go back to the use case and the data first.

Where Team 400 fits in

We are a Microsoft-leaning shop because that is where most of our clients live, but we have built and run production workloads on all three clouds. Our Azure AI consulting service handles the bulk of our engagements, and our AI strategy consulting work is genuinely cloud-neutral. We will tell you when AWS or Google is the right answer.

If you are sitting on a cloud decision right now, the fastest way to get unstuck is a 90-minute working session. Bring your use cases, your compliance constraints, your current cloud spend, and your team skills. We will give you a written recommendation by the end of the week.

Book a discovery call or look at our case studies to see how this plays out in practice.

The right cloud is the one that gets your project to production. Everything else is noise.