The Differences That Matter

Lesson 3 of 3 in Bedrock, Azure AI Foundry, Vertex AI.

After the tour, the checkboxes all match — which means checkboxes can no longer make the decision. In practice, four deltas do.

Catalog composition. Not the counts — the exclusives and the escape hatches. Each catalog carries model families the others do not, and some families now appear in all three (the docs pages behind our matrix list Anthropic models in Bedrock’s catalog, in Foundry’s partner category, and in Model Garden’s partner models). The escape hatches differ more than the headliners: Bedrock’s Custom Model Import brings your own Open weights weights behind the on-demand API, while Model Garden’s self-deploy path runs open models on compute in your own project and VPC — with your own containers.

Cross-region and global serving. All three route globally; they hang the switch in different places. On Bedrock it is an inference profile you invoke; on Azure it is the deployment type itself; on Vertex it is the endpoint hostname. The sweeteners and constraints differ too: Bedrock documents its global profiles at approximately 10% savings; Azure documents a Global-first rollout order for new models; Vertex documents the global endpoint’s feature exclusions. Where the switch lives determines who in your org controls it — the caller, the deployer, or the client configuration.

Customization depth. The SFT baseline is universal. Past it, the menus fork: Bedrock adds reinforcement fine-tuning with Lambda-scored rewards, managed distillation, and weight import; Foundry adds DPO and RFT on documented model sets, with serverless the exclusive route for OpenAI models; Vertex adds preference and RL tuning on an adapter model with multimodal training data, checkpoints, and continuous tuning. If your roadmap includes anything beyond supervised fine-tuning, this row alone can decide the platform.

Guardrail configurability. Same slot, three architectures: a standalone versioned resource callable outside any model invocation (Bedrock, including Automated Reasoning checks and contextual grounding checks); filter configurations attached to individual deployments, with approval gates on weakening them (Azure); per-category thresholds over a non-configurable base layer, plus a citation filter (Vertex). Which architecture fits depends on whether your safety controls live with the platform team, the app team, or the request path.

Interactive sorting exercise: A vendor slide makes each of these claims about their cloud’s model platform. Sort each one: is it table stakes that all three platforms document, or a genuine differentiator that only one of the three documents?

Tool: Cloud Model Selection Matrix — Every claim in this module, cell by cell, with a per-cell source link and the verification date: open the Cloud Model Matrix and compare the three platforms facet by facet — then pressure-test a vendor slide against it.

Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.