The Platform Tour

Lesson 2 of 3 in Bedrock, Azure AI Foundry, Vertex AI.

The tour visits each platform once, walking the same slots in the same order. Every claim is the documentation’s own, read on 2026-09-16, hedges intact — where the docs say “the overview states,” so do we.

Amazon Bedrock

Catalog. The Model Catalog in the Bedrock console lists Bedrock serverless models and Amazon Bedrock Marketplace models in one place. The User Guide overview states Bedrock supports 100+ foundation models from providers including Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI, and xAI. The Marketplace separately documents over 100 additional popular, emerging, and specialized models — those deploy to endpoints managed by SageMaker AI (with a documented bring-your-own-endpoint option). And Custom Model Import serves imported Open weights models through the standard InvokeModel APIs with on-demand throughput — your own weights behind the same Managed endpoint surface.

Consumption. Shared models are served on-demand, pay-per-token, with documented request-level service tiers: Standard (the default), Priority (fastest responses at a price premium), Flex (lower cost for non-time-sensitive work), plus an account-level Reserved tier — dedicated throughput on a term commitment. For dedicated capacity you purchase Provisioned Throughput in Model Units: each MU defines a per-minute input- and output-token processing level, billed hourly with no-commitment, 1-month, or 6-month terms. One constraint sentence to memorize: a model customized in Bedrock must be run on Provisioned Throughput.

Customization. Three documented methods: supervised Fine-tuning on labeled data; reinforcement fine-tuning, where you define reward functions using AWS Lambda to score responses (training data comes from uploaded prompt datasets or your existing Bedrock invocation logs); and distillation, where Bedrock generates responses from a chosen teacher model and uses them to fine-tune a smaller, faster student. Training is charged by tokens processed — training corpus × epochs — plus monthly model storage.

Grounding. Knowledge Bases searches your connected data to ground responses, with citations back to the source, in two types: Managed, where Bedrock runs ingestion, indexing, storage, and Retrieval (including agentic multi-hop retrieval, Smart Parsing, multimodal image search, and reranking models), and Customer-managed, where you run your own Retrieval-augmented generation (RAG) pipeline over a vector store such as OpenSearch Serverless, Aurora, or Neptune.

Guardrails. Configurable safeguards evaluate inputs and responses across models: content filters over predefined categories (Hate, Insults, Sexual, Violence, Misconduct, Prompt Attack — in Classic and Standard tiers), denied topics, word filters, sensitive-information filters (Personally identifiable information (PII) blocking or masking plus custom regex), contextual grounding checks that catch ungrounded RAG answers, and Automated Reasoning checks that validate responses against logical rules. Guardrails are versioned resources, applied at inference by ID and version — or independently of any model call via the ApplyGuardrail API.

Network and placement. PrivateLink interface VPC endpoints keep inference traffic off the public internet. Cross-region serving is explicit: geographic inference profiles for residency, global profiles for availability and the documented ~10% saving — routing adds no cost, stays on the AWS network with encryption in transit, and logs the processing Region in CloudTrail. The trade: inference profiles do not currently support Provisioned Throughput.

Azure AI Foundry

Catalog. The model catalog in the Foundry portal is branded Microsoft Foundry Models — the docs state over 10,000 models, with roughly 50 new models published per month, spanning foundation, reasoning, small language, multimodal, domain-specific, and industry models. It splits into two categories that matter operationally: Foundry Models sold by Azure (Microsoft-hosted and supported, including Azure OpenAI models) and Foundry Models from partners and community (for example Anthropic’s Claude family, and open models from the Hugging Face hub). Providers the page names include Microsoft, Azure OpenAI, Anthropic, DeepSeek, Meta, Mistral, Cohere, NVIDIA, and Hugging Face.

Consumption. Foundry documents two deployment options: managed compute, where model weights run on dedicated virtual machines billed by VM core-hours, and serverless deployments, where Microsoft hosts the model and bills for inputs and outputs, typically in tokens. Serverless deployments fall into standard (pay-per-token), provisioned (reserved capacity), and batch categories — nine documented deployment types in all — plus an instant access preview that lets supported models be called by name with no deployment created at all. Some models sold by Azure also offer fungible provisioned throughput: quota and reservations usable flexibly across those models. Serverless Quota is assigned per subscription, per region, per model, per deployment type, in tokens-per-minute, with proportional requests-per-minute; provisioned types are purchased as PTUs; a dynamic quota preview lets standard deployments burst above configured TPM when spare capacity exists.

Customization. Two fine-tuning modalities: serverless (Microsoft-managed capacity, consumption-priced — and the exclusive route for OpenAI models) and managed compute (a wider range of non-OpenAI models through AzureML, but you bring your own VM quota for training and hosting). Three techniques are documented: supervised fine-tuning for the broadest model set, Direct Preference Optimization for GPT-4o and GPT-4.1 models, and Reinforcement Fine-Tuning for o4-mini; vision + text fine-tuning is supported on GPT-4o and GPT-4.1.

Grounding. The model-service-level feature is Azure OpenAI On Your Data: files are ingested — cracked, chunked, embedded — into an Azure AI Search index, and at inference the service performs intent generation, retrieval, filtration and reranking, then grounded response generation. Supported sources include existing AI Search indexes, Blob Storage, local upload, URL/web content, and the integrated vector database in vCore-based Azure Cosmos DB for MongoDB. Know the status: the docs mark On Your Data deprecated, retiring October 14, 2026 (only specific GPT-4o and GPT-4o-mini versions remain supported) and recommend migrating to Foundry Agent Service with Foundry IQ knowledge bases — an agent-platform destination, which is our sister AI Agent Academy’s territory.

Content safety. A built-in content filtering system powered by Azure AI Content Safety runs both prompts and completions through classification models covering four harm categories — hate, sexual, violence, self-harm — at safe/low/medium/high severity. Optional filters add Prompt Shields for user prompt attacks and indirect Prompt injection attacks, protected-material detection for text and code, groundedness detection, Personally identifiable information (PII) detection, and blocklists. Filter configurations are created on the portal’s Guardrails + controls page and attached to individual model deployments; severity thresholds are adjustable, while disabling filters or annotate-only mode for completions requires approval for modified content filtering. Blocked prompts return HTTP 400; filtered completions signal a content_filter finish reason.

Network. Inbound isolation comes from setting public network access to Disabled with an Azure Private Link Private endpoint into your VNet (or Selected networks for specific IPs and VNets). Clients keep the same connection string — DNS remaps to a privatelink subdomain so in-VNet resolution returns the private IP. Private endpoints must sit in the same region and subscription as the VNet, and endpoints to dependent resources (AI Search, Storage, Cosmos DB) are created separately.

Vertex AI

Catalog. Model Garden is described in the docs as “an AI/ML model library that helps you discover, test, customize, and deploy models and assets from Google and Google partners,” organized into foundation models, fine-tunable models, and task-specific solutions. The groupings: Google models (Gemini, Veo, Lyria, Virtual Try-On), partner models (Anthropic Claude, Grok, Mistral AI), and open models (Gemma, Llama, DeepSeek, Qwen, Kimi, MiniMax, OpenAI gpt-oss, AlphaFold 3 among them). Open models can be consumed as managed serverless MaaS APIs or self-deployed with prebuilt or custom containers (a documented example: a custom vLLM container) running on compute inside your own project and VPC; Hugging Face Hub models can be deployed after security scanning, and an organization policy can restrict which models are accessible. The page states no total model count — an honest contrast with Azure’s 10,000+ claim.

Consumption. The docs list five options for generative models: Provisioned Throughput — a fixed-cost, fixed-term subscription reserving throughput per model and location, sold as GSUs on 1-week, 1-month, 3-month, and 1-year plans, with overage billed as standard pay-as-you-go and controllable per request, plus a Single Zone variant — three pay-per-token PayGo tiers (Standard, Priority at a premium with priority processing, Flex discounted for latency-tolerant work), and Batch for high-volume asynchronous jobs. Reserved capacity is documented as the baseline-coverage pattern: cover steady traffic with Provisioned Throughput and let spillover fall back to PayGo.

Customization. Supervised fine-tuning is the primary documented method — the docs contrast parameter-efficient (adapter) tuning with full fine-tuning — joined by preference tuning (human-feedback data on top of SFT), tuning checkpoints, continuous tuning of an already-tuned model, and reinforcement learning fine-tuning for Gemini with self-defined reward functions. Supervised tuning spans text, image, audio, video, and document data; it produces an adapter combined with the base model, is excluded from any SLA, and tuned-model inference shares the base model’s quota.

Grounding. RAG Engine — “a configurable managed RAG service” in the grounding docs — covers the full lifecycle: ingestion from local files, Cloud Storage, and Google Drive; chunking; embedding; indexing into a searchable corpus; retrieval; and generation, available in a documented list of regions at varying launch stages. Distinctively, it is one pluggable Grounding source among several — Grounding with Google Search, Google Maps, Agent Search, Elasticsearch, or your own search API can back model responses instead.

Placement and quotas. The locations page documents three serving surfaces: specific regional endpoints; us/eu multi-region endpoints that keep ML processing inside a jurisdictional boundary; and a global endpoint that “covers the entire world,” improves availability, and reduces resource-exhausted 429 errors — with documented limitations (no tuning, no batch prediction for Anthropic/OpenMaaS models, no RAG corpus, and Provisioned Throughput only for an enumerated set of Gemini models) and an explicit warning against it when you have ML-processing residency requirements. Online quotas are enforced per project, per region, and per base model; batch inference has no predefined quota limits, drawing on a shared pool allocated by real-time availability and demand.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.