Data, machine learning & technical architecture
AI cloud platform comparison 2026: Microsoft Foundry, Vertex AI, AWS, Databricks, and Claude Developer Platform
Cost, governance, edge support, and EU AI Act fit — a structured way to choose, instead of defaulting to whichever model is trending this quarter.
Picking an AI cloud platform in 2026 is no longer about picking a model
As of August 2026, choosing Microsoft Foundry, Google Vertex AI, AWS (Bedrock + SageMaker), Databricks, or the Claude Developer Platform means choosing a model catalog, an agent runtime, an MLOps stack, a governance layer, and a data residency posture — all at once. There is no universal winner. Microsoft Foundry suits organizations already standardized on Microsoft. Vertex AI is the cheapest option in the specific gpt-oss-120b scenario below and integrates tightly with BigQuery. AWS Bedrock and SageMaker offer a broad infrastructure toolbox. Databricks is especially relevant when governed data and AI operations need to stay together. The Claude Developer Platform provides first-party Claude access and a model-centric, MCP-native architecture; teams with a contractual or policy requirement for EU-only inference must instead evaluate a supported partner's EU regional deployment.
This guide breaks down what each platform actually delivers, what it costs, what it means for governance in Spain, Colombia, and El Salvador, and how to run a structured evaluation instead of defaulting to whichever LLM is trending this quarter.
Platform selection should sit inside a broader AI strategy for business, with the target workflow, operating model, governance, and success metrics defined before procurement.
The five platforms at a glance
Each platform groups a different set of products under one comparison, because there's no clean one-to-one match. Azure and AWS split classic ML from the newer generative layer; Databricks starts from a data/lakehouse platform and extends into ML and GenAI from there; and the Claude Developer Platform isn't a hyperscaler or a data layer at all — it's the model vendor's direct developer surface.
| Dimension | Microsoft Foundry + Azure ML | Vertex AI / Gemini Enterprise | AWS Bedrock + SageMaker AI | Databricks AI Platform | Claude Developer Platform |
|---|---|---|---|---|---|
| Positioning | Unified models, agents, and tools layer, plus Azure ML for the full ML lifecycle | Complete ML/GenAI stack; agent surface evolving into Gemini Enterprise Agent Platform | Bedrock for managed GenAI; SageMaker for training, MLOps, and custom ML | Multicloud data+AI lakehouse with MLflow, Unity Catalog, and AI Gateway | Cloud-agnostic developer surface; Anthropic models only, no data/ML layer of its own |
| Model catalog | 10,000+ reported catalog assets; deployable availability varies by model, region, and mode | 200+ validated options in Model Garden (Gemini, open, and partner models) | 100+ foundation models in Bedrock, plus Bedrock Marketplace | Foundation Model APIs plus proprietary/open models and external endpoints via AI Gateway | Anthropic only, with first-party API access and feature support — currently Opus 5, Sonnet 5, Haiku 4.5, and Fable 5 |
| Training / fine-tuning | Azure ML managed training, custom containers, GPU compute | Custom Training on GPU/TPU, KFP/TFX pipelines | SageMaker managed and custom training, HyperPod | AI Runtime with serverless A10/H100 GPUs (Public Preview; distributed API Beta) | None — development centers on prompting, Agent Skills, and the Agent SDK |
| On-prem / edge | Strong: Foundry Local on-device, Foundry Local over Azure Local/Arc | Strong for sovereign installs: Google Distributed Cloud, including air-gapped | Broad IoT/Greengrass/Outposts ecosystem, but Bedrock itself stays cloud-first | Strong multicloud, weak as a device/IoT runtime | Not offered directly; hybrid/sovereign scenarios route through Bedrock or Vertex |
| Governing standard | RBAC, private networking, Responsible AI Dashboard for classic ML | IAM, VPC Service Controls, CMEK | IAM, KMS, VPC, Bedrock Guardrails with Automated Reasoning checks | Unity Catalog for lineage and permissions across data and models | SOC 2 Type I/II, ISO 27001:2022, ISO/IEC 42001:2023; conditional HIPAA BAA and ZDR options for eligible organizations, models, and features |
What each platform actually offers
Microsoft Foundry (formerly Azure AI Foundry)
Microsoft renamed Azure AI Foundry to Microsoft Foundry in 2026, unifying models, agents, tools, evaluations, tracing, RBAC, and networking under one product. Microsoft reports more than 10,000 catalog assets and roughly 50 additions a month, spanning Microsoft/Azure OpenAI, Anthropic, Meta, Mistral, Cohere, DeepSeek, and Hugging Face; the number of models immediately deployable in a particular region or deployment mode is smaller. Azure Machine Learning continues to cover classic ML: AutoML, a visual designer, registries, endpoints, and data labeling, none of which Foundry itself replaces. Microsoft deployment overview.
The OpenAI-compatible endpoint pattern is Foundry's clearest anti-lock-in feature: one endpoint and credential set works across multiple models, with SDKs for Python, JavaScript, C#, Java, and REST. That mitigates lock-in at the API layer, though it doesn't remove the dependency on Entra ID, Azure AI Search, Agent Service, and the surrounding Azure network. For edge and offline inference, Foundry Local — running on-device or over Azure Local/Arc — is one of the more mature options in the market, though parts of it remain in preview.
Vertex AI / Gemini Enterprise Agent Platform
Google is repositioning the agentic layer that grew out of Vertex AI under Gemini Enterprise Agent Platform, while Vertex AI's core services — training, Pipelines, Model Registry, Model Garden — remain the operative product for ML and serving. Model Garden passed 200 validated models in 2025, spanning Gemini, open models, and partner models. Pipelines supports Kubeflow Pipelines v2 and TensorFlow Extended, which matters for teams that value portable standards, and the native link to BigQuery is a real advantage for any organization whose feature and analytics data already lives in Google's warehouse.
Two product changes are worth flagging before an RFP assumes historical continuity: Google retired Vertex AI Data Labeling in 2024, and in 2026 it marks Explainable AI as deprecated while Model Monitoring v2 remains in preview. Vertex doesn't stop being a strong platform because of this — but a procurement document shouldn't assume every legacy component still ships.
Google cites Etsy, which combines Vertex AI with BigQuery and Dataflow in its machine-learning and personalization stack. Google Cloud customer story.
AWS: Amazon Bedrock + SageMaker AI
AWS splits the stack into two complementary products. Bedrock is the managed GenAI layer — 100+ foundation models, the Converse API, Agents, Knowledge Bases, Guardrails, Flows, and a Marketplace. SageMaker AI covers the lower level: custom containers, training, HyperPod, feature engineering, Model Registry, Pipelines, and Ground Truth for human labeling. That split gives an experienced AWS team a lot of architectural freedom — an organization can keep proprietary weights and containers in SageMaker while using Bedrock only for the models worth consuming as a service. For a team new to AWS, the tradeoff is that the "AI product" is spread across more services than Vertex or Foundry present as a single surface.
DoorDash documents cutting generative AI application development time in half using Amazon Bedrock, a 50x increase in testing throughput, and response latencies of 2.5 seconds or less using Anthropic's Claude Haiku model, serving hundreds of thousands of daily calls. Clearwater Analytics reports increasing assets under management by more than 20% without increasing operational headcount in an AWS case study centered on SageMaker JumpStart. DoorDash case study; Clearwater case study.
Databricks AI Platform
Databricks is the platform most built around keeping data, features, experiments, models, and governance within one integrated platform. Unity Catalog governs data and AI objects; MLflow handles experiment tracking, evaluation, and registry; Lakeflow Jobs orchestrates; AI Gateway centralizes policy and observability for both hosted and external models. AI Runtime adds on-demand A10 and H100 compute, but as of August 2026 it remains in Public Preview, while its distributed multi-GPU training API remains in Beta. Databricks AI Runtime documentation.
Databricks doesn't eliminate cloud lock-in so much as relocate it. Instead of an application depending simultaneously on S3/Bedrock/SageMaker or ADLS/Foundry/Azure ML, Databricks becomes the stable data and ML plane while the underlying cloud or model provider changes underneath it. Unity Catalog and Databricks' serverless services then become the new strategic dependency.
Claude Developer Platform
Unlike the four platforms above, the Claude Developer Platform (api.anthropic.com) isn't a hyperscaler or an ML platform — it's a model-centric developer surface. Anthropic created and maintains the Model Context Protocol (MCP), the open standard the other four platforms have also adopted to connect agents to tools and data. The catalog is Anthropic-only — the current lineup is Opus 5, Sonnet 5, Haiku 4.5, and Fable 5 — and the first-party platform offers Anthropic's native API features, including extended thinking, prompt caching, the Files API, and Agent Skills. Availability can still vary by model, account, and feature.
The Claude Agent SDK (released September 29, 2025, renamed from "Claude Code SDK") exposes the same production infrastructure behind Claude Code as a Python/TypeScript library: file editing, bash execution, web search and reading, a tool-use loop with human-in-the-loop checkpoints, subagents, and persistent sessions. Agent Skills shipped separately as an open, cross-platform standard. Integration is deliberately cloud-agnostic — REST API and MCP toward any stack, with Workload Identity Federation for AWS IAM, Google Cloud, GitHub Actions, Kubernetes, Microsoft Entra ID, and Okta — making the platform a fit for cloud-agnostic teams, coding/agentic workloads, and organizations that prioritize first-party Claude features over a multi-model catalog or proprietary training infrastructure. Claude is also available through Bedrock, Vertex, and Microsoft Foundry for teams that prefer to consume it inside a cloud they already run.
The first-party API only offers "us" or "global" inference geographies via the inference_geo parameter — there's no EU-only option — and "us"-only inference on Claude 4.6+ models carries a 1.1x price multiplier. GDPR does not generally require EU-only hosting: organizations can use lawful international-transfer mechanisms. If a contract, internal policy, or transfer assessment requires EU-only inference, evaluate Claude through Amazon Bedrock or Google Vertex AI in a supported EU region. Anthropic data-residency documentation.
Inference and training costs: a 2026 snapshot
Token and GPU prices are only part of total cost. Use an AI ROI measurement framework to include implementation, integration, evaluation, governance, adoption, and ongoing operations in the business case.
Prices change monthly across all five platforms, so treat every figure below as an August 2026 snapshot, not a permanent number.
A standardized workload — 1 million monthly requests, 1,000 input tokens and 250 output tokens per request, served on gpt-oss-120b where a comparable rate exists — produces these approximate monthly costs:
| Platform | Input / 1M tokens | Output / 1M tokens | Approx. monthly cost |
|---|---|---|---|
| Microsoft Foundry | $0.15 | $0.60 | ~$300 |
| Vertex AI | $0.09 | $0.36 | ~$180 |
| Amazon Bedrock | ~$0.15 | ~$0.60 | ~$300 (regional; e.g. ~$309 in Sydney) |
| Databricks Foundation Model Serving | 2.143 DBU | 8.571 DBU | ~$300 at $0.07/DBU (contract-dependent) |
Calculation: 1 billion input tokens plus 250 million output tokens. Rates are August 2026 snapshots and can vary by region and service tier. Databricks' approximately $300 result assumes a contract-dependent $0.07/DBU and is therefore illustrative rather than a universal public list price. Google pricing; AWS pricing; Microsoft pricing.
The real takeaway isn't that Google looks ~40% cheaper on this one model — it's that the same open-weight model can carry different prices depending on who serves it, which makes the ability to switch endpoints a genuine FinOps lever. That advantage can disappear entirely with a different model, region, or modality. AWS states that prompt caching can cut costs by up to 90% and latency by up to 85% in eligible cases; Microsoft has added model routing to pick models by prompt; Google publishes meaningful batch discounts. These are vendor claims, not guaranteed savings for any specific workload.
Claude doesn't serve gpt-oss-120b, so it sits outside this direct comparison. Anthropic's own per-token pricing (August 2026): Haiku 4.5 at $1/$5 per million input/output tokens; Sonnet 5 at an introductory $2/$10 (moving to $3/$15); Opus 5 at $5/$25; Fable 5 at $10/$50. Batch API carries the same 50% discount pattern as the other platforms, and prompt caching cuts cached-input cost by roughly 90%, with context windows up to 1M tokens on the newest models at standard rates.
GPU training: 8×H100 for 24 hours
| Platform | Reference config | Rate | 24h cost |
|---|---|---|---|
| AWS | p5.48xlarge, 8×H100 |
$55.04/h | $1,320.96 |
| Google Cloud | a3-highgpu-8g, 8×H100 |
$88.49/h | $2,123.76 |
| Azure | Standard_ND96isr_H100_v5, 8×H100 |
$98.32/h | $2,359.68 |
| Databricks | AI Runtime on 8×H100 | $56.00/h ($7/GPU-h) | $1,344.00 |
Rates are on-demand snapshots for the cited configurations; regions and included services differ. The Databricks rate is its published US East Public Preview price. Storage, data transfer, taxes, commitments, and discounts are excluded. Databricks pricing; Google Cloud pricing.
Azure's higher hourly rate buys dedicated NVIDIA Quantum-2 CX7 InfiniBand networking at 400 Gbit/s per GPU, with documented scaling to thousands of GPUs. ClusterMAX currently places Azure in its overall Gold tier and AWS and Google in Silver; that rating combines multiple infrastructure criteria rather than measuring networking alone. GPU pricing is one of the most volatile variables here: AWS Capacity Blocks and other commitments can undercut on-demand pricing. Azure ND H100 v5 documentation; ClusterMAX review.
Security, governance, and data residency
Microsoft offers a tight fit for an organization already standardized on Entra ID, RBAC, VNet/Private Link, and CMK. For models sold by Azure, Microsoft states that prompts, outputs, embeddings, and training data aren't made available to other customers or model vendors and aren't used to train foundation models without permission. Residency depends on deployment mode: Global deployments can process data across geographies; DataZone deployments keep processing within the selected US, EU, or APAC zone, potentially across multiple regions; Regional deployments process in the deployment region. Azure's Responsible AI Dashboard combines error analysis, fairness, feature importance, and counterfactual analysis for supported classic-ML scenarios. Microsoft deployment types.
Google offers IAM, CMEK, VPC Service Controls, and Access Transparency, with VPC-SC particularly useful for building perimeters around data and AI resources. As noted above, Explainable AI's deprecation matters for regulated sectors that require feature-level explanation as a formal requirement.
AWS lets you select the storage region, supports encryption in transit and at rest, and offers AWS-managed or customer-managed keys with BYOK options depending on the service. Bedrock Guardrails' Automated Reasoning checks — introduced in 2025 to verify certain responses against formalized policies — claim up to 99% verification accuracy in AWS materials, a figure that applies to that specific mechanism, not overall model accuracy.
Databricks centers governance on Unity Catalog, with TLS in transit and AES-256 at rest, CMK support, and residency controls via Databricks Geos. Its compliance security profile supports configurations for PHI, PCI, and FedRAMP-relevant workloads where the underlying service and region support them.
Anthropic holds SOC 2 Type I and II, ISO 27001:2022, and ISO/IEC 42001:2023 certifications. A HIPAA BAA is available only after activation for eligible organizations, models, APIs, and features; ZDR is limited to qualified accounts and supported features. Default API retention is 30 days with automatic backend deletion; content flagged by moderation can be retained for longer periods under Anthropic's policy. The first-party API offers only "us" or "global" inference geographies, with no EU-only option, and Covered Models such as Claude Fable 5 and Mythos 5 require 30-day retention and aren't eligible for ZDR. GDPR does not itself require EU hosting, but an organization with an EU-only processing requirement should evaluate Claude through a supported partner deployment in an EU region. Anthropic BAA coverage; data residency.
Across every platform, "compliant" is never a property of the whole platform — it's a property of a specific region, service, model, and configuration. A regulated buyer should turn residency, key management, private networking, logging, and subprocessor terms into verifiable contract and architecture requirements, not marketing claims.
What this means in Spain, Colombia, El Salvador, and the United States
| Market | Current position | What enterprise buyers should do | Key sources |
|---|---|---|---|
| Spain / EU | The EU AI Act reached general application on August 2, 2026. Article 99 sets tiered ceilings of up to €35 million or 7% of worldwide annual turnover for prohibited practices, €15 million or 3% for other obligations, and €7.5 million or 1% for incorrect information. GDPR adds separate exposure of up to €20 million or 4%. Regulation (EU) 2026/1744 moved Annex III high-risk obligations to December 2, 2027 and Annex I product-safety obligations to August 2, 2028. | Document the transfer mechanism, deployment geography, subprocessors, and governance controls. EU residency is not the same as legal sovereignty; GDPR permits transfers through mechanisms including adequacy decisions, SCCs, and BCRs. | Regulation (EU) 2026/1744 EU transfer rules |
| Colombia | Statutory Law 1581 of 2012 anchors the framework, with the SIC as enforcement authority. Colombia has no general data-localization mandate. Transfers to countries without adequate protection require a statutory exception or authorized safeguard; transmissions to processors require a data-transmission agreement. Reform Bills 214/2025 and 274/2025 were archived and are not pending as of August 2026. | Apply privacy by design and by default under SIC Circular 002 of 2024. Distinguish transfers from processor transmissions, document the appropriate safeguard, and report qualifying incidents to the SIC/RNBD within 15 business days. | Colombian Chamber record SIC incident guidance |
| El Salvador | The Law for the Protection of Personal Data, Decree No. 144, became effective in November 2024 and broadly applies to public and private entities, subject to statutory exclusions. It regulates international transfers and sets fines at 1–40 monthly commerce-sector minimum wages—approximately $408.80–$16,352 using the rate effective since June 2025. | Appoint a data protection officer where obligated, document equivalent-protection or another statutory basis for international transfers, and confirm the current minimum wage before estimating exposure. Oversight sits with the State Cybersecurity Agency (ACE). | Decree No. 144 Minimum-wage announcement |
| United States | The United States has no single cross-sector AI statute equivalent to the EU AI Act. Federal obligations remain sector-specific and rely partly on existing agency authorities; the NIST AI Risk Management Framework is voluntary. State law drives much of the operational work: Texas's AI governance law took effect January 1, 2026, while Colorado's revised automated-decision law takes effect January 1, 2027. | Map each deployment by state, sector, affected person, and decision type. Maintain an AI inventory, clear notices, impact testing, human-review paths, and privacy and discrimination controls. A US cloud region does not establish compliance by itself. | CRS overview NIST AI RMF Texas HB 149 Colorado AG guidance |
These figures are an editorial starting point, not legal advice. Verify the current legislation, regulator guidance, contracts, and deployment configuration before making compliance commitments.
A 6-step framework to choose your AI cloud platform
Most platform evaluations start with "which model is best right now" — the wrong first question, since GPT, Claude, Gemini, and the leading open models all move faster than a company's architecture does. Use this sequence instead:
Before comparing any model, list what's already standardized: identity provider (Entra ID, Google Workspace, Okta), primary data warehouse, and existing cloud spend commitments. This alone often eliminates two of the five platforms on integration cost grounds.
For any workload you expect to run continuously, model Provisioned Throughput (Azure), Capacity Blocks (AWS), and committed-use pricing (Google) against on-demand rates. The gap is frequently large enough to change the platform recommendation outright.
"We use Azure/GCP/AWS" isn't a compliance answer on its own — Global vs. DataZone, region, model, endpoint, and modality all change which commitments actually apply. If Claude is in scope, confirm the inference geography and international-transfer mechanism instead of assuming that GDPR requires—or that a vendor automatically provides—EU-only processing.
Build an internal model-gateway interface and keep retrieval-augmented generation (RAG) and business logic out of the most proprietary SDKs. Databricks AI Gateway is one commercial option for this pattern; a custom abstraction layer works too. This single decision determines how expensive it is to switch models later.
Measure TTFT, tokens/second, p95/p99 latency, throttling behavior, output quality, and cost per resolved task — for the exact model, region, and concurrency you'll run in production. Independent measurements show meaningful differences for the same model across providers.
A platform that's excellent for agents doesn't automatically have the best human labeling, explainability, or custom training tooling. Vertex AI retired Data Labeling and is deprecating Explainable AI; Azure keeps a more specialized Responsible AI Dashboard for classic models. Evaluate these as two different purchasing decisions, even if you end up on one platform for both.
Which platform fits which use case
| Use case | First choice | Why |
|---|---|---|
| R&D across many models quickly | Microsoft Foundry | 10,000+ reported catalog assets plus integrated comparison/evaluation; deployable availability varies |
| Gemini/multimodal research | Vertex AI | Most direct access to Google's image/video/audio model portfolio |
| Large-scale custom model training | AWS SageMaker/HyperPod or Azure ML | AWS currently leads on H100 pricing; Azure leads on cluster networking — benchmark with the real model |
| High-volume serverless inference | Vertex AI or AWS Bedrock | Vertex wins on the reference model's price; Bedrock adds tiering, caching, and batch |
| Microsoft-standardized enterprise | Microsoft Foundry | Minimizes identity/network/data integration overhead |
| BigQuery-centric organization | Vertex AI | BigQuery, Vertex Pipelines, and Model Garden form the most direct data-to-model path |
| Existing AWS enterprise | Bedrock + SageMaker | Reuses IAM, S3, VPC, and KMS without adding an operational plane |
| Regulated industries (finance, insurance, healthcare) | Azure or AWS | Azure leads on identity and classic-ML explainability; AWS leads on regional controls and Guardrails |
| Sovereign / air-gapped cloud | Google Distributed Cloud or Azure Local | GDC air-gapped can host Vertex AI services; Azure offers Foundry Local/Azure Local and Arc |
| Lakehouse + GenAI on the same data | Databricks | Unity Catalog, MLflow, and Lakeflow cut duplicated data pipelines and governance work |
| Multicloud, model lock-in reduction | Databricks | AI Gateway governs hosted and external models across AWS/Azure/GCP |
| Cloud-agnostic team, first-party Claude API, coding/agentic workloads | Claude Developer Platform | First-party Claude features, native MCP, and Agent SDK — but no EU-only inference on the direct API |
| Claude required where policy or contract mandates EU-only inference | Claude via Amazon Bedrock or Google Vertex AI (EU region) | The direct Anthropic API offers only "us"/"global" geographies |
What's changed in the last 18 months
-
Microsoft adds the Responses API
The Responses API landed in Azure AI Foundry, expanding the platform's agent-building surface before the broader Foundry consolidation.
-
Catalogs, guardrails, and agent tooling expand
Google pushed Model Garden past 200 options. AWS launched Automated Reasoning in Bedrock Guardrails. Databricks launched Agent Bricks and added GPT-5 natively.
-
Open-weight models reach the managed clouds
Microsoft added
gpt-ossopen-weight models, AWS added OpenAI's open-weight models, and Google addedgpt-ossas a managed offering. -
Anthropic launches the Claude Agent SDK
The SDK exposed the production infrastructure behind Claude Code as a Python and TypeScript developer surface for building tool-using agents.
-
Agent Skills becomes an open standard
Anthropic released Agent Skills as an open, cross-platform format, extending its focus from models into reusable agent capabilities.
-
The major platforms consolidate around agents
Microsoft unified the stack under the Microsoft Foundry brand, adding model routing and expanding Foundry Local and Azure Local for edge inference. Google began shifting the agent experience toward Gemini Enterprise Agent Platform while Explainable AI moved toward deprecation. Claude reached general availability inside Microsoft Foundry, and Anthropic's Workload Identity Federation reached general availability.
-
Several operational capabilities are still maturing
AWS AgentCore Policy documentation includes temporal, session-aware controls, while service quotas define throttling limits. Databricks AI Runtime provides serverless GPUs but remains in Public Preview, with distributed multi-GPU training in Beta. Anthropic's first-party lineup includes Opus 5, Sonnet 5, Haiku 4.5, and Fable 5, while earlier 4.x models remain available on some surfaces. AgentCore Policy; AI Runtime status.
Microsoft was named a Leader in Gartner's 2025 Magic Quadrant for AI Application Development Platforms, while AWS and Google were Leaders in the 2024 Magic Quadrant for Cloud AI Developer Services. Synergy Research Group put AWS at 29% of global cloud-infrastructure spending in Q3 2025, with Microsoft at 20% and Google at 13%. Grand View Research values the global AI-agent market at $7.6 billion in 2025 and projects $182.9 billion by 2033 at a 49.6% CAGR. Gartner's often-cited 80% figure is a forecast published in 2023: it predicted that by 2026 more than 80% of enterprises would have used generative-AI APIs or models, or deployed GenAI-enabled applications—not that measured adoption had already reached 80%. Synergy Research; Grand View Research; Gartner forecast.
Frequently asked questions
Which AI cloud platform is cheapest for inference?
On the standardized gpt-oss-120b benchmark used in this article, Vertex AI comes out roughly 40% cheaper than Foundry or Bedrock (~$180/month vs. ~$300/month for 1 million monthly requests). Databricks is also approximately $300 only under the article's illustrative $0.07/DBU contract assumption. The gap can shrink or reverse with a different model, region, tier, caching, or batch discount.
Does Claude offer EU data residency?
Not through Anthropic's first-party API, which only supports "us" or "global" inference geographies. GDPR does not generally require EU-only hosting. If your contract, policy, or transfer assessment requires EU-only inference, evaluate Claude through Amazon Bedrock or Google Vertex AI in a supported EU region.
What is the Model Context Protocol (MCP), and why does it matter for platform choice?
MCP is the open standard Anthropic created for connecting AI agents to external tools and data sources. All five platforms in this comparison support it, which can reduce integration lock-in. Portability is not automatic: authentication, tool schemas, policy, hosting, and runtime behavior can still require platform-specific work. See the practical guide to AI agents for business for deployment patterns and guardrails.
Which platform is best for a regulated industry like finance or healthcare?
There is no platform-wide compliance winner. Azure is often shortlisted for Microsoft identity integration and classic-ML explainability; AWS for regional controls, key management, and Guardrails; and Google Distributed Cloud for supported disconnected deployments. Treat compliance as a property of the exact region, service, model, contract, and configuration.
Can I avoid being locked into one AI vendor?
Partially. Databricks and a well-designed internal model gateway both reduce lock-in by keeping data, evaluation, and routing logic independent of any single model provider. But model-level lock-in is different from infrastructure lock-in: if an application depends specifically on a Claude, Gemini, or GPT capability, only that vendor can serve it — regardless of which cloud sits underneath.
Choosing without guessing
The right platform depends on what's already standardized, what data residency actually requires, and how a team plans to build agents — not on which model tops a benchmark this month. Liorant helps CIOs, CTOs, and ML teams in Spain, Colombia, and El Salvador run this evaluation against real workloads instead of vendor marketing, and builds the working system on top of whichever platform fits.
We identify your highest-value automation opportunity and explain exactly how Liorant can help — no slides, no pitch.