Home / Articles / Cloud
Cloud AI Industry

AWS, Azure and Google Cloud: The AI Platform War

Bedrock, Foundry, Vertex AI and the custom chips underneath — how the hyperscalers compete for AI workloads.

By SiliconMind EditorialNov 202613 min read
Key takeaways
  • The three hyperscalers compete at every layer: model marketplaces, ML platforms, GPU capacity and custom silicon.
  • AWS leads with Bedrock (multi-model, including Claude), SageMaker, and its own Trainium and Inferentia chips.
  • Microsoft built its AI business around its OpenAI partnership and has broadened Foundry to many model providers, including Anthropic.
  • Google Cloud combines Gemini, Vertex AI and a decade of TPU development.
  • Pick a platform based on where your data lives, the models you need, networking and compliance — not hype.

For most organizations, the path to frontier AI runs through a cloud provider. Rather than buying GPUs or signing contracts with every model lab, they call models through managed platforms on AWS, Microsoft Azure or Google Cloud. Competition among those three — and specialized "neoclouds" — has become one of the biggest capital spending races in corporate history.

What a cloud AI platform provides

Agents & appsagent runtimes · knowledge bases · guardrails Model catalogfirst-party + partner + open-weight models ML platformtraining, fine-tuning, eval, deployment, MLOps ComputeNVIDIA GPUs · custom accelerators · networking + IDENTITY · NETWORKING · DATA · COMPLIANCE
Fig. 1 — The hyperscaler AI stack. Enterprises value the bottom row they already trust (identity, networking, data governance) as much as the models on top.

Amazon Web Services

  • Amazon Bedrock — a fully managed service offering models from Anthropic (Claude), Meta, Mistral, Cohere, DeepSeek, OpenAI's open-weight models, Amazon's own Nova family and others through one API. It adds knowledge bases for RAG, guardrails, model evaluation, batch inference and agent tooling (AgentCore).
  • Amazon SageMaker AI — the end-to-end platform for building, training and deploying your own models, including large distributed training through SageMaker HyperPod.
  • Custom silicon — Trainium chips for training and inference and Inferentia for inference, programmed via the Neuron SDK. AWS built "Project Rainier," a very large Trainium2 cluster used by Anthropic to train Claude models.
  • GPU instances — EC2 P5 (H100/H200) and P6 (Blackwell) instances, with UltraClusters and EFA networking for scale-out.

AWS's pitch is choice and integration: models sit inside your VPC boundary with IAM, CloudWatch, PrivateLink and existing compliance programs.

Microsoft Azure

  • OpenAI partnership — Microsoft's multi-billion-dollar investment gave Azure early, deep access to GPT models, and Azure hosts OpenAI models for enterprise customers.
  • Microsoft Foundry (formerly Azure AI Foundry) — a unified platform with a catalog of thousands of models, including OpenAI, Anthropic's Claude, Meta, Mistral, xAI and open-weight models, plus agent services, evaluation and safety tooling.
  • Copilot — Microsoft's AI assistant layer spans Microsoft 365, GitHub Copilot and Windows, making Microsoft both a platform and a major consumer of AI compute.
  • Silicon — the Maia AI accelerator and Cobalt Arm CPU complement massive NVIDIA deployments.

Google Cloud

  • Gemini — Google DeepMind's model family, available through Vertex AI and integrated into Google Workspace.
  • Vertex AI — Google's ML platform, with Model Garden offering Gemini, Gemma, Claude, Llama and many others, plus agent-building tools.
  • TPUs — Google has designed Tensor Processing Units since the mid-2010s. Recent generations, including the inference-focused Ironwood, are deployed in very large pods with optical circuit switching. Anthropic is among the major external TPU customers.

Side-by-side

AWSAzureGoogle Cloud
Managed model serviceBedrockFoundry / Azure OpenAIVertex AI
Flagship first-party modelsAmazon NovaMicrosoft models (MAI, Phi)Gemini, Gemma
Claude availableYesYesYes
Custom AI chipsTrainium, InferentiaMaiaTPU
ML platformSageMaker AIAzure Machine LearningVertex AI

The neoclouds

Specialized GPU clouds — CoreWeave, Lambda, Crusoe, Nebius and others — grew rapidly by offering large NVIDIA clusters with less overhead and faster delivery of new hardware. Oracle Cloud Infrastructure also emerged as a major AI capacity provider through very large contracts with model labs. These providers mostly serve AI labs and large training workloads rather than general enterprise IT.

Why custom silicon matters

Each hyperscaler buys enormous quantities of NVIDIA GPUs and also designs its own chips. The reasons are economic and strategic:

  • Cost per token — chips tuned to a provider's own workloads and data centers can be cheaper to run at scale.
  • Supply security — less dependence on a single vendor's allocation.
  • Vertical integration — co-designing chips, servers, networking and cooling.

The catch is software: developers want PyTorch to "just work." Investment in compilers (XLA, Neuron) and frameworks like JAX and vLLM backends is what makes custom silicon usable.

Choosing a platform

  1. Start where your data is. Moving petabytes and re-doing compliance is expensive; most teams start with their primary cloud.
  2. Confirm model availability in the regions you need, including newest versions and features.
  3. Check quotas and throughput. Provisioned capacity and rate limits can matter more than list price.
  4. Plan for multi-model. Abstract model calls so you can switch providers as the landscape changes.
  5. Use private networking (PrivateLink / Private Service Connect) and least-privilege IAM from day one.
Hands-onSee Build an AI App with Claude on AWS for a working Lambda example that calls Claude through both the Anthropic API and Amazon Bedrock.