⚡The verdict up front
Cloud comparison articles love to declare a winner. For AI workloads specifically, the honest answer is that your model access strategy decides more than the cloud does, and your existing commercial relationships decide more than the technology does. A company with a Microsoft Enterprise Agreement, a data platform on AWS, and a preference for Gemini has three competing gravitational pulls — and the right answer is to follow the strongest one, not to split the difference.
The commitments: choose Azure when OpenAI models are your target and your organization already buys Microsoft — the procurement, identity and compliance story is genuinely smoother. Choose AWS when you want the broadest managed model catalog, the deepest general infrastructure, and you value optionality across model providers. Choose Google Cloud when training or large-scale inference economics dominate (TPUs are a real cost lever) or when your AI workload is really a data-analytics workload with a model attached — BigQuery to Vertex AI is the tightest data-to-model loop of the three.
What you should not do is run production AI across all three clouds because a slide deck said multi-cloud. The egress, identity and skills overhead of that choice exceeds any negotiating leverage it creates for most companies under a few hundred engineers. The sections below justify each of those claims in turn.
| Profile | Recommended home | Why |
|---|---|---|
| Microsoft enterprise, OpenAI-first | Azure (Azure OpenAI + Foundry) | Contract, identity and compliance alignment |
| Model-agnostic builder, broad catalog | AWS (Bedrock) | Widest managed model choice, mature blocks |
| Training-heavy or TPU-compatible | Google Cloud | TPU economics, Vertex AI depth |
| Data-analytics-centric AI | Google Cloud | BigQuery-to-Vertex loop |
| Startup with credits to burn | Wherever the best credits are | All three run aggressive programs |
| Regulated industry, existing cloud | Stay where you are | Migration cost beats model delta |
🤖Bedrock vs Azure OpenAI vs Vertex AI, honestly
Amazon Bedrock is a managed model marketplace and toolkit: Anthropic’s Claude family, Meta’s Llama models, Mistral, Amazon’s own Nova family and others, behind one API, with agents, knowledge bases (managed RAG), guardrails and evaluation built in. Its strength is breadth and neutrality — you can swap models without changing vendors, which matters when the best model for your task changes every two quarters. Its weakness is that the building blocks are capable but generic; you assemble more of the system yourself than the marketing suggests.
Azure OpenAI Service gives you OpenAI’s models — the GPT line and friends — deployed inside your Azure tenancy, with Microsoft’s enterprise identity, networking and compliance wrapped around them. Microsoft’s Foundry platform has broadened the catalog beyond OpenAI models, but the core draw remains the tightest enterprise packaging of the OpenAI ecosystem. The honest caveats: capacity and quota management on popular models has been a recurring operational complaint, and you are tying your model roadmap to one provider’s relationship with its primary cloud partner.
Vertex AI is the most complete ML platform of the three as a platform: first-party Gemini models, a Model Garden that includes third-party and open models, plus mature MLOps — pipelines, feature store, evaluation, model monitoring — that predate the generative wave. If your workload is classic ML plus generative AI (forecasting and ranking alongside summarization and agents), Vertex treats that as one discipline, which matches how real systems actually look. Its weakness is gravitational: the platform is at its best when your data already lives in Google Cloud.
Pick the platform by where your data and contracts already live, then pick models within it. Choosing a cloud for one exclusive model is how companies end up migrating two years later when the model leaderboard reshuffles.
🎛️GPU availability and pricing: the moving target
All three clouds sell the same headline NVIDIA hardware — H100 and B200-class instances, with GB200 systems rolling out — and all three will tell you availability is fine. The operational truth as of writing: on-demand flagship GPUs remain scarce at times in popular regions, committed-use reservations are how serious buyers secure capacity, and spot capacity for training is viable but demands checkpoint discipline. Treat any specific GPU price you read, including in this article, as stale on arrival — all three vendors adjust pricing and introduce new instance families quarterly.
The real pricing story is the custom silicon. AWS pushes Trainium for training and Inferentia for inference, Google has its TPU line (with the latest generations aimed squarely at large-scale training and inference), and Azure has its Maia accelerators entering the mix. These chips matter because they are where the clouds put their best price-performance — typically a meaningful discount per unit of work versus comparable NVIDIA instances, in exchange for framework compatibility work and ecosystem lock-in. For sustained large inference fleets, that discount is often the difference between a viable and an unviable unit cost.
For most companies, the practical guidance is boring: if you are fine-tuning occasionally and serving through managed endpoints, GPU instance shopping is a distraction — your cost driver is tokens through Bedrock, Azure OpenAI or Vertex, not instances. If you are training models weekly or running an inference fleet at scale, then instance economics, reserved capacity and custom silicon become worth a dedicated evaluation, and Google’s TPUs and AWS’s Trainium deserve benchmarking on your actual workload rather than on vendor slides.
| Need | AWS | Azure | Google Cloud |
|---|---|---|---|
| Flagship NVIDIA instances (H100/B200 class) | Yes — P-series; reserve for capacity | Yes — ND-series; quota constraints reported | Yes — A-series; reserve for capacity |
| Custom AI silicon | Trainium (training), Inferentia (inference) | Maia (newer, narrower deployment) | TPU (most mature, best-documented) |
| Custom-silicon trade-off | Price-performance for ecosystem work | Early — evaluate carefully | Strong economics for TPU-compatible stacks |
| Managed model endpoints (no GPU ops) | Bedrock | Azure OpenAI / Foundry | Vertex AI |
| Best fit for | Mixed fleets, broad infra | OpenAI-centric serving | Training-heavy, TPU-compatible |
🪤Egress cost traps and other invoice ambushes
Data egress remains the most reliably underestimated line in cloud AI budgets. The pattern is always the same: the pilot runs entirely inside one cloud and costs nothing to move data, then production arrives — data flows from your cloud object storage to a model endpoint, results stream to users on the public internet, logs replicate across regions — and the egress line becomes visible. Public internet egress and cross-region transfer are metered on all three clouds, and AI workloads move more data than the web apps that preceded them.
One genuine improvement: under regulatory and competitive pressure, all three providers introduced waivers for egress fees when a customer is leaving the platform entirely. That reduces exit risk. It does nothing for the steady-state case — serving responses to users, feeding models from another cloud’s storage, or syncing data between regions — which is where the real money goes. Model the steady state, not the exit.
The other ambushes, briefly: managed model endpoints bill per token and the bill scales with success, which is fine if your per-task economics are measured and terrifying if they are not. Cross-AZ and cross-region replication of vector stores and training data accumulates silently. And premium support tiers, committed-use discounts and private pricing all interact — a discount on compute does nothing for an egress-heavy architecture. An architecture review before signing a commit is worth more than the discount negotiation itself.
📄Enterprise agreements and startup credits: the real pricing
List prices are for people without negotiating leverage, and if you are spending seriously on AI, you have leverage. AWS structures this through its Enterprise Discount Program, Azure through Enterprise Agreements with Azure consumption commitments, and Google Cloud through committed-use discounts and private pricing. In all three cases the mechanics are similar: commit to an annual spend, receive a discount across services, and accept that over-committing wastes money while under-committing leaves discounts unearned. The commit-versus-flexibility trade is the actual decision, and it should be made with a forecast you trust, not with the vendor’s forecast.
For startups, all three run aggressive credit programs, and the numbers move often enough that you must verify current terms before planning around them. As of writing, the shape is: AWS Activate offers credits up to roughly the low six figures for funded startups, Microsoft for Startups offers Azure credits on a tiered ladder into the low six figures, and Google for Startups offers credits up to the low-to-mid six figures over two years, with enhanced amounts for AI-focused startups. These credits are genuinely valuable — they can fund an entire pre-seed infrastructure budget — and they are also a customer-acquisition expense for the vendor, designed to make your second decision (where to stay) before your first invoice arrives.
Two disciplines protect you. First, architect so credits are spent on portable infrastructure — managed Kubernetes, standard databases, object storage — rather than on the deepest proprietary services, so the end of credits is a pricing event and not a migration crisis. Second, track credit burn-down as a board metric: the companies hurt by credits are not the ones whose credits expire, but the ones whose architecture assumed the pricing would never change.
🌐The multi-cloud truth
Multi-cloud for AI comes in three flavors, and only one of them is usually worth it. The first flavor — running the same workload redundantly across two clouds for resilience — is almost never worth it below true enterprise scale: you double your identity, networking, observability and skills burden to survive a failure mode (full-cloud outage) that is rarer than the outages your own complexity will cause.
The second flavor — best-of-breed by workload, such as data in BigQuery with inference on Bedrock — is legitimate but must pay an egress and latency tax at every boundary. It works when each side is strongly justified and the boundary is thin and well-defined, for example a nightly batch flow rather than a chatty real-time one. The third flavor — negotiating leverage and exit optionality — is real but is better served by portable architecture (containers, standard Postgres, object storage abstractions, model-provider-agnostic code) than by actually operating two clouds.
Our default recommendation: one primary cloud for production AI, architecture that keeps your exit option alive, and contractual attention at renewal time. That posture captures most of the leverage multi-cloud promises with a fraction of its permanent overhead. If a vendor knows you can leave, you own most of what multi-cloud was supposed to buy you.
🎯The recommendation by company profile
If you are a Microsoft enterprise — Office, Entra ID, an Enterprise Agreement, a compliance team that already trusts Azure — choose Azure and Azure OpenAI, unless your workload is training-heavy. The integration dividend (identity, networking, procurement, audit) is real money, and fighting it to save a theoretical few percent on tokens is a bad trade. If you are a product company building model-agnostic AI features and your infrastructure already runs on AWS, choose Bedrock: the catalog breadth protects you from model-leaderboard churn, and the surrounding primitives are the most battle-tested.
If you are training models as a core activity, or your AI feature is really a data product, give Google Cloud a serious evaluation even if it is not your incumbent — TPU economics and the BigQuery-to-Vertex loop are structural advantages, not marketing. If you are a startup with no gravity in any direction, take the best credits available, build on portable primitives, and revisit the question when the credits run out.
And if you are already somewhere with a working system: stay. Migration between clouds costs more in engineering and risk than any plausible model-price delta will repay within your planning horizon. The exception is a hard capability gap — a model you contractually need, a compliance requirement your current cloud cannot meet — in which case migrate deliberately, workload by workload, with the egress math done first. A scoped architecture review is the cheapest way to find out which situation you are actually in.
Plan your cloud AI architecture with usBook an architecture review
| Your company | Primary pick | Watch out for |
|---|---|---|
| Microsoft-heavy enterprise | Azure | Model-provider concentration; quota management |
| AWS-incumbent product company | AWS Bedrock | Assembling too much yourself; token cost visibility |
| Training-heavy AI company | Google Cloud | TPU compatibility work; data gravity shift |
| Data-analytics-led AI | Google Cloud | Serving-layer maturity versus your needs |
| Early-stage startup | Best current credits | Lock-in that outlives the credits |
| Regulated industry with incumbent cloud | Stay put | Paying migration cost for marginal model gains |
