← Agent Frameworks 🕐 22 min read
Agent Frameworks

Enterprise AI Platform Vendor Comparison — 2026

Executive Summary · March 2026

> **Audience:** Enterprise architects and CIOs making strategic AI platform decisions.

Audience: Enterprise architects and CIOs making strategic AI platform decisions. This note covers AWS Bedrock, Microsoft Azure AI Foundry (now Microsoft Foundry), Google Vertex AI (now Gemini Enterprise Agent Platform), Databricks Mosaic AI, and Snowflake Cortex AI as full enterprise platforms — not just inference APIs.

Five vendors have consolidated the enterprise AI platform market in 2026. None of them is a clear winner across all workloads. The correct selection depends on three variables in priority order: (1) existing cloud/data platform commitment, (2) workload type — inference, training, or analytics AI, and (3) team ML maturity. Most Fortune 500 enterprises will operate two of these platforms simultaneously, not one.

Key 2026 platform rebrands that matter for vendor discussions:

  • Azure AI Foundry → renamed Microsoft Foundry (January 2026). Same product, new name signals it is a first-party Microsoft platform, not an Azure service.
  • Vertex AI → now surfaces under Gemini Enterprise Agent Platform (announced Google Cloud Next '26, April 2026). Vertex AI remains the underlying infrastructure brand in developer docs.
  • AWS Bedrock AgentCore → reached General Availability for Policy (March 2026), GovCloud availability (May 2026), and added optimization/evaluation capabilities (preview, May 2026).
  • Databricks Unity AI Gateway → launched April 2026, adding MCP server governance and on-behalf-of user execution to the existing Unity Catalog model registry.

1. Platform Capability Matrix

Ratings: ★★★★★ (best-in-class) → ★★☆☆☆ (limited) → ★☆☆☆☆ (absent or nascent)

Capability AWS Bedrock Azure AI Foundry / MS Foundry Google Gemini Ent. Agent Platform Databricks Mosaic AI Snowflake Cortex AI
Proprietary model access Amazon Nova family GPT-4o, o-series, MAI family Gemini 3.1 Pro/Flash, Gemma 4 DBRX (open weights) Arctic (open weights)
Third-party model catalog ★★★★★ Anthropic, Meta Llama, Mistral, Cohere, AI21, Stability, DeepSeek ★★★★☆ Anthropic, Mistral, DeepSeek, Llama ★★★★☆ Claude Opus/Sonnet/Haiku, Llama, Mistral via Model Garden (200+ models) ★★★☆☆ External models via Model Serving; limited managed catalog ★★☆☆☆ Arctic, Llama, Mistral, Reka; no external API calls
Fine-tuning support ★★★☆☆ Supervised FT + Reinforcement FT on select models (Claude, Nova, Llama) ★★★☆☆ Fine-tuning on GPT-4o mini, base models; Azure ML for custom ★★★☆☆ Supervised + RLHF via Vertex AI Tuning; Gemini Flash tuning GA ★★★★★ First-class: SFT, RLHF, DPO, continual pre-training; any open model ★★☆☆☆ Arctic fine-tuning in preview; limited to Snowflake-hosted models
RAG / knowledge base (managed) ★★★★☆ Bedrock Knowledge Bases (OpenSearch Serverless or S3 Vectors); built-in chunking, embedding ★★★★☆ Azure AI Search; “on your data” for SharePoint/Blob; chunking/embedding tooling ★★★★☆ Vertex AI Search; Grounding API; enterprise data connectors ★★★★☆ Vector Search on Delta Lake; MLflow tracing for RAG evals ★★★★☆ Cortex Search (semantic search on Snowflake data); no external data movement needed
Agent runtime (managed orchestration) ★★★★★ AgentCore: Runtime (Firecracker microVM), Memory, Gateway (API/Lambda/MCP), Identity, Policy (Cedar), Evaluation ★★★★☆ Foundry Agents + Prompt Flow; Foundry Control Plane for observability ★★★★★ Agent Studio, Agent-to-Agent Orchestration, Agent Registry, Agent Gateway, Agent Simulation, Agent Identity ★★★☆☆ Mosaic AI Agent Framework (Python-native, LangChain/LlamaIndex); no fully managed runtime ★★☆☆☆ Cortex Agents (early stage); primarily analytics task automation
Observability / monitoring ★★★☆☆ CloudWatch + Bedrock model invocation logs; AgentCore Observability for agents; requires assembly ★★★★☆ Foundry Control Plane: invocations, tool calls, latency, error rates unified ★★★★☆ Agent Observability, Agent Evaluation dashboards; multi-step reasoning traces ★★★★★ Managed MLflow: experiment tracking, trace logging, model monitoring; Unity Catalog lineage ★★★☆☆ AI Observability (GA imminent 2026): no-code + pro-code monitoring; Cortex Guard
Data integration (native lakehouse) ★★★☆☆ S3, Glue, Athena integration; no native Delta Lake; must wire via Glue ★★★☆☆ Azure Data Lake, Fabric OneLake integration; tight with M365 data ★★★★★ Native BigQuery integration; direct SQL-to-LLM via BigQuery ML; no data movement ★★★★★ Native Delta Lake; Unity Catalog manages data + models in same catalog ★★★★★ Models run on Snowflake data in-place; zero data movement; Cortex AISQL (preview)
Governance (model cards, audit, RBAC) ★★★★☆ IAM, Guardrails (content filters, PII, grounding), Service Control Policies; model invocation logging ★★★★★ Azure AD/Entra RBAC, FedRAMP High, DoD IL5, content filtering, Responsible AI dashboard ★★★★☆ VPC Service Controls, Customer-managed keys, model version registry; Agent Identity ★★★★★ Unity Catalog: fine-grained RBAC on models + data; lineage tracking; audit logs; MLflow model cards ★★★★☆ Snowflake RBAC extends to models; Cortex Guard; row/column-level security on underlying data
Pricing model Per-token (on-demand) + Provisioned Throughput (hourly commitment) + Guardrails per text unit Per-token + PTU (Provisioned Throughput Units, monthly commitment); expensive at scale without commitment Per-token (on-demand); committed use discounts; Agent Platform seat-based licensing emerging DBU-based (hourly compute, $0.07–$0.75/DBU) + cloud VM passthrough (adds 50–200% to DBU spend) Snowflake credits ($2–$4/credit by tier); unified with data warehouse spend; no separate AI fee tier
Enterprise support tiers AWS Business/Enterprise Support; Bedrock-specific TAM available Microsoft Unified Support; dedicated AI FastTrack engineering Google Cloud Premium Support; Customer Engineer for Gemini migrations Databricks Premium + Platinum (24/7); dedicated Customer Success Engineer Snowflake Business Critical + Virtual Private Snowflake tiers; dedicated support

2. AWS Bedrock

Platform Character

Bedrock is the broadest inference marketplace in the enterprise cloud. It does not try to be the best model — it tries to be the platform where any model can run under enterprise security controls. AgentCore, reaching GA through early 2026, elevates Bedrock from an inference API to a managed agent operating system.

Key 2026 Developments

  • AgentCore GA timeline: AgentCore Policy (Cedar-based authorization) hit GA March 2026. GovCloud availability followed May 2026, unlocking regulated workloads (ITAR, FedRAMP High). AgentCore Optimization (batch evaluation + A/B testing for agents) entered preview May 2026.
  • AgentCore architecture: Runtime uses Firecracker microVMs on EC2 bare metal — each agent invocation is isolated at the hypervisor level. Memory service handles short- and long-term state across sessions. Gateway converts APIs, Lambda functions, and MCP servers into agent-callable tools without custom integration code. Identity service handles outbound authentication to enterprise APIs.
  • Model catalog depth: Remains unmatched. Anthropic (Claude Opus, Sonnet, Haiku), Meta (Llama 3.x series), Mistral, Amazon Nova, Cohere, AI21, Stability AI, DeepSeek. Switching models requires no infrastructure change.
  • Knowledge Bases pricing note: Default vector store (OpenSearch Serverless) has a $345/month floor at 2 OCUs regardless of traffic. Amazon S3 Vectors (launched late 2025) offers up to 90% lower cost for lighter workloads — now the smarter default for new RAG deployments.
  • Fine-tuning scope: Supervised fine-tuning and Reinforcement Fine-Tuning (RFT) on select models. Coverage is narrower than Databricks — not all catalog models support fine-tuning.

Strengths

  • Deepest third-party model catalog; no vendor lock-in to a single model family
  • AgentCore provides the most complete managed agent runtime in the market: isolated execution, managed memory, tool gateway, authorization, evaluation in one service
  • Bedrock Guardrails: content filters, PII redaction, grounding checks, topic denial — applied uniformly regardless of which model is invoked
  • Tight integration with S3, Glue, Athena, Lambda, IAM — the entire AWS security posture extends to AI workloads
  • GovCloud availability makes Bedrock the default choice for federal and defense workloads

Weaknesses

  • No native data lakehouse. Enterprises that want to query Delta Lake or BigQuery tables must wire through S3 exports or custom Glue jobs.
  • Observability requires assembly: CloudWatch for infrastructure metrics, Bedrock invocation logging for model-level data, AgentCore Observability for agent traces. No unified AI observability dashboard out of the box.
  • Fine-tuning is model-constrained. Large-scale training workloads do not belong on Bedrock — Databricks Mosaic AI is the appropriate platform.
  • Per-token pricing without PTU commitments becomes expensive at high throughput. Provisioned Throughput requires capacity pre-commitment with limited flexibility.

Best For

AWS-native enterprises; Anthropic-heavy workloads (Bedrock offers the deepest Claude integration outside Anthropic’s own API); regulated industries requiring GovCloud; enterprises that need model optionality without infrastructure lock-in.

3. Azure AI Foundry (Microsoft Foundry)

Platform Character

Microsoft Foundry (rebranded from Azure AI Foundry, January 2026) is the tightest integration between AI capability and enterprise productivity data. No other platform has direct access to SharePoint, Teams, Exchange, and Dynamics data without ETL. That single fact defines the majority of Microsoft Foundry’s differentiation.

Key 2026 Developments

  • Rebrand significance: The rename from “Azure AI Foundry” to “Microsoft Foundry” is not cosmetic. It signals the product is treated as a first-party Microsoft platform like Office 365, not an Azure PaaS service. Sales motion, SKU bundling, and enterprise agreements are being restructured accordingly.
  • Multi-model expansion at Build 2026: Foundry now hosts Anthropic, Mistral, DeepSeek, and Microsoft’s MAI (Microsoft AI) model family alongside OpenAI. The Foundry Control Plane provides unified invocations, tool calls, model responses, latency, and error rates across all model providers.
  • Prompt Flow maturity: Prompt Flow for orchestrating retrieval-generation pipelines is production-hardened. The “on your data” feature — connecting GPT models directly to SharePoint document libraries, Azure Blob, and Dynamics 365 — remains exclusive to the Microsoft stack.
  • Compliance posture: FedRAMP High and DoD IL5 certifications remain the strongest compliance story in the market. For enterprises in defense, healthcare, and financial services under strict data residency requirements, this is often the deciding factor before any capability evaluation begins.
  • PTU complexity: Provisioned Throughput Units (PTUs) require monthly capacity commitments priced at the model level. Enterprises running multiple GPT model versions simultaneously face significant PTU management overhead. This is the primary TCO trap in Foundry deployments.

Strengths

  • GPT-4o exclusive first-mover availability before other clouds; o-series models available on parity with OpenAI direct
  • M365 Copilot integration: AI agents built in Foundry can access the same SharePoint/Exchange data as Copilot, with the same Entra RBAC controls
  • Strongest enterprise compliance: FedRAMP High, DoD IL5, HIPAA, PCI DSS, ISO 27001 — the most complete regulated-industry coverage of any AI platform
  • Azure AI Search for RAG: production-hardened at enterprise scale; hybrid search (vector + keyword) with built-in semantic reranking
  • Responsible AI Dashboard: model fairness analysis, error analysis, counterfactual analysis — unique among the five platforms

Weaknesses

  • Most expensive for non-Microsoft models. Running Anthropic Claude through Foundry costs more than running it through Bedrock at equivalent scale because Azure adds a platform margin.
  • PTU commitment complexity creates financial risk. Enterprises frequently over-provision PTUs for peak load and pay for idle capacity. The alternative (on-demand) becomes cost-prohibitive above 10M tokens/day.
  • Heavy Azure-first lock-in. The “on your data” M365 integration that drives most enterprise Foundry deployments is exclusively available on Azure. Multi-cloud or hybrid architectures erode the core value proposition.
  • Prompt Flow, while capable, is less flexible than Python-native frameworks (LangChain, LlamaIndex) favored by ML-mature teams. Databricks and Bedrock have both moved to support those frameworks natively.

Best For

Microsoft-shop enterprises (E3/E5 agreements, Azure Consumption Commitments); organizations where GPT-4o is a hard requirement; regulated industries (defense, healthcare, federal) requiring FedRAMP High or DoD IL5; enterprises where M365 data access is the primary RAG source.

4. Google Gemini Enterprise Agent Platform (Vertex AI)

Platform Character

Google’s April 2026 consolidation of Vertex AI under the Gemini Enterprise Agent Platform brand is a substantive architectural move, not just a rename. The platform now spans model access, agent building, agent deployment, observability, and employee-facing agent interfaces in a single product surface. The BigQuery integration remains Google’s sharpest competitive advantage: no other platform lets you run LLM inference directly against petabyte-scale SQL tables without data movement.

Key 2026 Developments

  • Gemini Enterprise Agent Platform GA (April 2026): Unifies Model Garden (200+ models), Agent Studio, Agent-to-Agent Orchestration, Agent Registry, Agent Gateway, Agent Identity, Agent Observability, and Agent Simulation under one product umbrella.
  • Model Garden expansion: Gemini 3.1 Pro (1M context), Gemini 3.1 Flash Image, Lyria 3 (audio generation), Gemma 4 (open weights), plus third-party access to Claude Opus/Sonnet/Haiku, Llama, Mistral. The context window advantage (1M tokens) remains unique among proprietary models.
  • Agent Simulation: Unique capability — tests agents against synthetic user prompts before production deployment. Closes the evaluation gap that has plagued enterprise agent deployments.
  • Employee-facing layer: Within the Gemini Enterprise app, non-technical employees can build and deploy agents using natural language and reusable skills. The Inbox dashboard provides a unified view of all agent activity. This brings AI agent management to operations and business teams, not just IT.
  • BigQuery ML direct inference: LLM functions callable directly from BigQuery SQL — ML.GENERATE_TEXT(MODEL 'gemini-1.5-flash', TABLE my_table, STRUCT(...)). No data export, no pipeline, no embedding service required for tabular AI.

Strengths

  • Largest context window in the market: Gemini 3.1 Pro at 1M tokens. Workloads requiring full document analysis, long-horizon reasoning, or large codebase comprehension have no practical alternative.
  • Native BigQuery integration: direct SQL-to-LLM without data movement is a structural cost and complexity advantage for BigQuery-native enterprises
  • Most complete agent orchestration stack post-April 2026 rebrand: Agent Studio → Agent Registry → Agent Gateway → Agent Identity → Agent Observability → Agent Simulation is a coherent production deployment pipeline
  • Multimodal leadership: Gemini 3.1 Flash Image, native video understanding, Lyria 3 audio — strongest multimodal stack of the five platforms
  • Agent Simulation for pre-production testing is the most differentiated capability in enterprise agent safety

Weaknesses

  • Third-party model catalog thinner than Bedrock. Google’s Model Garden is growing but lacks Bedrock’s breadth across enterprise model providers (no Cohere, AI21, or Stability AI equivalent).
  • Google Cloud market share (~12% enterprise cloud) means fewer enterprises have the pre-existing infrastructure commitment that justifies platform-level lock-in. AWS and Azure sales motions are stronger in enterprise accounts.
  • BigQuery requirement for the best data integration creates its own lock-in. Enterprises on Snowflake or Databricks for analytics do not get the seamless SQL-to-LLM story.
  • Agent Platform rebrand is recent (April 2026) — product documentation, ISV ecosystem, and enterprise support playbooks are still maturing relative to Bedrock AgentCore.

Best For

BigQuery-native enterprises; multimodal workloads (video, image, audio); long-context applications (contract analysis, research synthesis, large codebase agents); enterprises already invested in Google Workspace who want to connect Gmail/Docs to AI agents.

5. Databricks Mosaic AI

Platform Character

Databricks is the only platform in this comparison that is model-training-native rather than inference-native. Every other platform is fundamentally an API to run models that exist. Databricks is where enterprises build the models that will run elsewhere. Unity Catalog as a governance layer spanning data and models — not just one or the other — is a structural architecture advantage that no other platform replicates.

Key 2026 Developments

  • Unity AI Gateway (April 2026): Expands Unity Catalog into AI agent governance. Adds MCP server governance (Databricks can now enforce access policies on MCP tool calls from agents), on-behalf-of user execution (agents inherit the identity of the requesting user for downstream data access), and end-to-end MLflow tracing across agent invocations. This is the most significant Databricks release for enterprise AI governance in 2026.
  • MLflow 3: Introduces Logged Models and Deployment Jobs as first-class concepts. Model lifecycle (evaluation → approval → deployment) is managed as a workflow governed by Unity Catalog, with full activity logs on every model version page.
  • Mosaic AI Agent Framework: Python-native, integrates with LangChain and LlamaIndex. Agent traces go directly into managed MLflow for debugging and evaluation. Not a managed runtime in the AgentCore/Agent Studio sense — engineers write Python, not YAML configs.
  • AI/BI Genie: Natural language to SQL built into the Databricks workspace. Targeted at business users who want to query Delta Lake tables without writing Spark.

Strengths

  • Best-in-class fine-tuning and custom model training: SFT, RLHF, DPO, continual pre-training, model distillation on any open-weight model. This is what Bedrock fine-tuning cannot match and what enterprises with proprietary domain data require.
  • Unity Catalog is the most sophisticated data + model governance layer in the market. Fine-grained RBAC, column-level masking, row-level filtering, and model version control live in the same catalog with unified audit logs and lineage tracking.
  • Managed MLflow: the de facto standard for ML experiment tracking is deeply integrated. Every training run, evaluation, and serving endpoint is traced automatically.
  • Native Delta Lake integration: AI workloads run on the same transactional data store as data engineering pipelines. No ETL, no copy, no lag. This is structurally similar to Snowflake Cortex but for open-format data.
  • Unity AI Gateway’s MCP governance is the first enterprise-grade solution for controlling what external tools AI agents can call — a critical gap in regulated industries.

Weaknesses

  • Not a managed inference platform. Model serving requires DBU compute allocation, cloud VM passthrough (adding 50–200% to DBU cost), autoscaling configuration, and ongoing operations. For enterprises that want to call an API without managing infrastructure, Databricks requires significantly more engineering.
  • Cost structure is opaque for AI workloads. DBU charges range from $0.07 (model serving) to $0.75 (all-purpose compute), but cloud VM pass-through (EC2, Azure VMs, GCE) is billed separately and often exceeds DBU cost. TCO modeling requires Databricks-specific expertise.
  • Primarily appropriate for ML-mature teams. Enterprises without data science or ML engineering capability will struggle with the Python-native, infrastructure-aware model. Bedrock and Foundry are operationally simpler.
  • No proprietary frontier model. DBRX and fine-tuned models require training investment. Enterprises needing frontier model access without training budget are better served by Bedrock or Foundry.

Best For

Enterprises with large proprietary training workloads; financial services, healthcare, and life sciences that require domain-specific fine-tuned models; teams with existing Databricks investment for data engineering; organizations where model governance and lineage are regulatory requirements (SR 11-7, EU AI Act high-risk AI systems).

6. Snowflake Cortex AI

Platform Character

Snowflake Cortex occupies a deliberately narrow position: AI on your analytics data, governed by your existing Snowflake controls, priced on your existing Snowflake contract. It is not trying to compete with Bedrock on model breadth or with Databricks on training capability. Its value proposition is operational simplicity for the enterprise SQL stack — analysts and data engineers already in Snowflake can add AI capabilities without procuring a separate AI platform.

Key 2026 Developments

  • Cortex AISQL (public preview): Multimodal data processing using AI callable from SQL. Enables complex AI workflows (text extraction, classification, summarization, structured output) through standard SQL syntax without Python or external API calls.
  • Document AI (approaching GA): Extracts structured content from unstructured documents (invoices, contracts, forms) using Snowflake Arctic-TILT, a multimodal LLM optimized for document understanding. Arctic-TILT benchmarks above GPT-4 on document extraction tasks according to Snowflake.
  • Cortex Guard: Safety layer for Cortex AI outputs. Content filtering and responsible AI guardrails applied at the SQL function level — no separate safety service to configure.
  • AI Observability (approaching GA): No-code and pro-code monitoring of generative AI applications built on Cortex. Integrates with existing Snowflake monitoring dashboards.
  • Cortex Analyst: Natural language to SQL. Business users ask questions in English, Cortex Analyst generates and executes the query, returns results with explanation. Productized for BI teams, not data scientists.

Strengths

  • Zero data movement: Cortex models run directly on Snowflake data. For enterprises with data residency requirements or strict egress cost controls, this eliminates a compliance and cost problem that every other platform creates.
  • Simplest governance story: Snowflake RBAC, row-level security, and column-level masking extend automatically to AI capabilities. No separate identity system, no additional access control layer to configure. An analyst who cannot see a column cannot ask a question that would expose it through Cortex.
  • Unified pricing: Cortex AI capabilities bill in Snowflake credits alongside data storage and compute. For enterprises with Snowflake Enterprise or Business Critical contracts, AI is an incremental spend within an existing budget line — not a new vendor procurement.
  • Cortex Analyst and Cortex AISQL lower the skill floor for AI adoption. SQL-fluent analysts can deploy AI features without learning Python, LangChain, or cloud ML services.
  • Arctic LLM is optimized for enterprise SQL generation and instruction following — measurably better at enterprise data tasks than general-purpose models at equivalent parameter counts.

Weaknesses

  • Narrow model catalog. Arctic, Llama, Mistral, and a small set of curated models. No Anthropic Claude, no GPT-4o, no Google Gemini through Cortex. Enterprises with model-specific requirements must go elsewhere.
  • No external API calls from Cortex agents. Agent-to-external-service integration (calling Salesforce, Jira, or custom APIs) is not supported in Cortex. This limits Cortex AI to analytics and data transformation use cases; enterprise workflow automation requires a different platform.
  • Limited agent runtime. Cortex Agents are early-stage compared to AgentCore, Foundry Agents, or Gemini Agent Studio. Complex multi-step agent workflows are not a Cortex strength.
  • Primarily an analytics-AI platform. Real-time inference workloads, customer-facing AI applications, and generative content production are outside Cortex’s designed scope.

Best For

Snowflake-native enterprises; analytics and BI use cases (text-to-SQL, document extraction, automated reporting); enterprises where data governance and residency concerns make external API calls unacceptable; data teams that want AI without a separate AI platform procurement; finance, legal, and compliance analytics workflows.

7. Decision Framework

Step 1 — Start With Existing Cloud and Data Commitment

The single highest-value decision input is existing infrastructure commitment. Ecosystem integration compounds across data connectors, identity, compliance posture, and commercial pricing. If the enterprise is 70%+ on one cloud, start with that cloud’s AI platform and only deviate when a capability gap is structural.

Existing commitment Default starting point Deviate if…
AWS AWS Bedrock Training workload requires Databricks; analytics AI requires Snowflake
Azure Microsoft Foundry Non-Microsoft model exclusive requirement; training workload
GCP / BigQuery Gemini Enterprise Agent Platform Thin model catalog; need Anthropic-exclusive capabilities
Snowflake (cloud-agnostic) Snowflake Cortex AI Workload exceeds analytics/BI; agent runtime needed
Databricks (cloud-agnostic) Databricks Mosaic AI Inference-only workload; no training budget
No dominant commitment Bedrock (model breadth) or Foundry (compliance) —

Step 2 — Workload Type Routing

Workload Best platform Rationale
Multi-model inference at scale AWS Bedrock Deepest catalog, no infrastructure management, AgentCore for agents
GPT-4o-required workloads Microsoft Foundry Exclusive access depth; M365 data integration
Long-context (>100K tokens) Google Gemini Agent Platform 1M token context; no practical alternative
Multimodal (video/audio/image) Google Gemini Agent Platform Lyria 3, Gemini Flash Image, native video understanding
Domain-specific fine-tuning Databricks Mosaic AI Only platform with full SFT/RLHF/DPO pipelines on arbitrary open models
Analytics AI / text-to-SQL Snowflake Cortex AI In-place compute on warehouse data; Cortex Analyst
Regulated/federal workloads Microsoft Foundry (FedRAMP High/DoD IL5) or Bedrock (GovCloud) Both have appropriate certifications; pick by cloud commitment
Document extraction at scale Snowflake Cortex AI (Arctic-TILT) or Bedrock (Claude) Snowflake for Snowflake-resident data; Bedrock for general document AI
Agent orchestration (production) AWS Bedrock AgentCore or Google Gemini Agent Platform Most complete managed agent runtimes in 2026
Model governance / lineage Databricks Mosaic AI Unity Catalog; MLflow; Unity AI Gateway for agent governance

Step 3 — Team Maturity Assessment

Team profile Appropriate platforms
ML mature (data scientists, MLEs) Databricks Mosaic AI, Bedrock (custom integrations), Vertex AI
Cloud architects (no ML background) Bedrock (managed services), Foundry (Prompt Flow), Snowflake Cortex
Data engineering / analytics Snowflake Cortex, Databricks (AI/BI Genie), BigQuery ML
No dedicated AI team Microsoft Foundry (M365 Copilot integration), Snowflake Cortex (SQL-native)

Multi-Platform Strategy vs. Single-Platform Standardization

The case for multi-platform: Most large enterprises will operate at minimum two of these platforms — one for inference/agents (Bedrock, Foundry, or Vertex AI) and one for training/governance (Databricks) or analytics AI (Snowflake). This is not an architecture failure; it reflects that no single platform is best-in-class across inference, training, and analytics simultaneously.

The common practical pattern:

  • Bedrock + Databricks: Bedrock for inference and agent runtime; Databricks for training, fine-tuning, and model governance. Models trained on Databricks, deployed to Bedrock for serving.
  • Bedrock + Snowflake: Bedrock for customer-facing AI applications; Snowflake Cortex for internal analytics AI. Two separate governance regimes but clean workload separation.
  • Databricks + Snowflake: For enterprises on both platforms (common in enterprise data stacks), use Databricks for ML training and Snowflake Cortex for analytics AI. Interoperability via Delta Sharing.

The case for single-platform: Operational simplicity, unified billing, single vendor support relationship, and unified governance layer. The premium is real: enterprises that standardize on a single platform typically pay 10–20% more per workload than a best-of-breed multi-platform architecture. The offset is 30–50% lower DevOps overhead for AI platform management.

Recommendation: Standardize on the cloud-native AI platform for inference and agents. Add Databricks if and only if training/fine-tuning is a current or near-term requirement — not speculatively.

TCO Comparison — 50M Tokens/Day at Scale

The following estimates are directional at 50M tokens/day of mixed inference workload (equivalent to a mid-large enterprise AI program). Assumes 50% input / 50% output token split, Claude Sonnet-class model, no fine-tuning, standard support tier. Excludes data storage, egress, and DevOps labor.

Platform Inference model Monthly token cost (est.) Required commitments Key hidden costs
AWS Bedrock Claude Sonnet 4.x (on-demand) $45K–$65K None (on-demand); PTU for predictability OpenSearch Serverless floor ($345/month per RAG KB); Guardrails per text unit ($0.15/1K chars)
Azure AI Foundry GPT-4o (on-demand) $60K–$90K PTU commitment to avoid on-demand premium PTU over-provisioning waste; Azure egress costs; Entra P2 license for advanced RBAC
Azure AI Foundry GPT-4o (PTU committed) $35K–$50K 730-hour monthly PTU commitment Commitment waste at <70% utilization; model version upgrade friction
Google Gemini Agent Platform Gemini 3.1 Pro (on-demand) $40K–$60K None (on-demand); committed use for discounts Agent Platform seat licensing (emerging); BigQuery ML query costs
Databricks Mosaic AI Llama 3.x (Model Serving) $25K–$40K DBU commitment recommended Cloud VM passthrough (+50–200% DBU cost); GPU node reservation for training; MLe labor
Snowflake Cortex AI Llama / Mistral via Cortex $20K–$35K Snowflake Enterprise contract Credit consumption varies with Cortex function complexity; no external model options

TCO reality check (sourced from multi-platform deployment data): Most enterprises discover actual costs are 40–60% above modeled costs. The platforms with the lowest per-token rate are rarely the platforms with the lowest twelve-month TCO once data transfer, storage, engineering overhead, and observability tooling are included. At twelve months, most enterprises cluster within 15% of each other on total spend across the three cloud-native platforms (Bedrock, Foundry, Vertex AI).

Lock-In Risk and Exit Strategy

Platform Lock-in vectors Exit difficulty Mitigation
AWS Bedrock AgentCore agent state (memory, policy), proprietary Knowledge Bases format, Guardrails configuration Medium Use LangChain/LlamaIndex abstractions for agent logic; export Knowledge Base documents; Guardrails logic is replicable
Azure AI Foundry M365 data connectors, PTU financial commitments, Prompt Flow DAGs, Azure AD RBAC Medium–High OpenAI models available via API elsewhere; Prompt Flow exports to Python; PTU commitments are the primary exit barrier
Google Gemini Agent Platform Agent Registry configurations, BigQuery ML function syntax, 1M-context Gemini dependency Medium Agent logic is Python-based; BigQuery SQL is portable at query level; Gemini context advantage is unique and hard to replicate elsewhere at same cost
Databricks Mosaic AI Unity Catalog metadata, Delta Lake format (open), MLflow models (open standard), DBU pricing Low–Medium Delta Lake and MLflow are open standards; model artifacts are portable; Unity Catalog metadata migration is the primary friction
Snowflake Cortex AI Cortex functions embedded in SQL, Snowflake-specific SQL syntax, credit contracts Medium SQL transformation logic is portable; Cortex function calls require rewrite to external API equivalents; contract exit timing is the primary barrier

Lowest lock-in: Databricks Mosaic AI (open standards: Delta Lake, MLflow, open model weights). Highest lock-in: Azure AI Foundry (PTU financial commitments + M365 data access exclusivity + GPT-4o availability premium create compounding exit barriers).

8. Summary — Which Platform Wins Where

Criterion Winner Runner-up
Broadest model catalog AWS Bedrock Google Gemini Agent Platform
Best managed agent runtime AWS Bedrock AgentCore Google Gemini Agent Platform
Strongest enterprise compliance Azure AI Foundry / MS Foundry AWS Bedrock (GovCloud)
Best M365 / productivity data integration Azure AI Foundry / MS Foundry —
Largest context window Google Gemini Agent Platform (1M) —
Best multimodal capabilities Google Gemini Agent Platform AWS Bedrock (Stability AI)
Best SQL-to-LLM / analytics AI Snowflake Cortex AI (in-place) Google Gemini Agent Platform (BigQuery ML)
Best fine-tuning and training Databricks Mosaic AI AWS Bedrock (limited)
Strongest model governance / lineage Databricks Mosaic AI (Unity Catalog) Azure AI Foundry (Responsible AI Dashboard)
Lowest operational complexity Snowflake Cortex AI AWS Bedrock
Best for ML-mature teams Databricks Mosaic AI Google Gemini Agent Platform
Lowest exit risk Databricks Mosaic AI Google Gemini Agent Platform
Best GovCloud / FedRAMP AWS Bedrock (GovCloud) Azure AI Foundry (FedRAMP High, DoD IL5)

Research Notes

This synthesis draws on platform documentation, vendor announcements, and third-party analysis current through June 2026. Key 2026 events incorporated: AWS AgentCore GA (March–May 2026), Microsoft Foundry rebrand (January 2026), Google Gemini Enterprise Agent Platform launch (April 2026), Databricks Unity AI Gateway (April 2026), Snowflake Cortex AISQL public preview. TCO figures are directional estimates based on published pricing; actual enterprise pricing will reflect negotiated discounts (typically 20–40% off list for >$1M annual commitments).

Sources consulted: