Last reviewed: August 2026
For foundation models (FMs) and which model to choose, see Getting Started with AI and AI Platform and Model Comparison first. This document is not about choosing a model — it focuses on where you consume the same FM : directly from the model provider (1P) vs. via a cloud platform (3P).
When adopting FMs, there are two independent decisions :
Decision Axis
Question
Options
① Usage Type Who uses it and how?
Seat (chat UI) / API (code integration) / Self-hosting
② Channel Where do you buy and operate?
Direct (1P) / Cloud-hosted (3P) / Provider + Cloud billing
Example: "Claude via API on Bedrock"
→ Channel: Cloud-hosted (3P Inference)
Example: "ChatGPT Enterprise for all employees"
→ Usage type: Seat (Enterprise)
Aspect
Seat Plans
API Plans
User Individual employees (incl. developers)
Applications, services, automation systems
How Use provider-built UI/tools directly (chat, Claude Code, Copilot)
Embed model programmatically in your product/service
Billing Per-user subscription
Per-token/request usage-based
Best for Personal productivity — writing, analysis, coding, research
Customer-facing features, large-scale automation, batch processing
Control User/group/admin policies
API keys, per-project budgets, rate limits
Cost predictability Seats × price = fixed
Variable with usage
Reason for 3P
Explanation
Consume existing cloud commits AWS EDP, Azure EA/MACC can offset LLM costs
Network isolation VPC/PrivateLink — traffic stays off public internet
Compliance leverage Reuse cloud vendor’s existing certifications (HIPAA, SOC 2, etc.)
Unified billing/governance Single invoice, cost tags for team tracking, IAM access control
Single contract No procurement effort for new vendor
Multi-model access Switch between models via same platform API
Reason to stay 1P
Explanation
Latest models/features New models and beta features ship to 1P first
Full functionality Fine-tuning, Realtime API, etc. may lag or be absent on 3P
Simple start Register card, start immediately, no cloud setup
Cloud-agnostic No dependency on specific cloud
Common enterprise pattern: Production on 3P (governance, commit consumption, isolation) + Experimentation on 1P (latest features, fast start).
Pattern
Analogy
Operator
Billing
Feature Scope
A. Direct (1P) Brand store
Model provider
Provider direct
Full (latest first)
B. Provider service + Cloud billing Department store brand shop
Model provider
Cloud marketplace (commit-eligible)
Full (same as 1P)
C. Cloud-hosted (3P Inference) Private-label/select shop
Cloud vendor
Cloud invoice
Cloud API scope (some features delayed)
D. Self-hosting (open-weight) Cook it yourself
Customer
Infrastructure costs (GPU/server)
Full customization within license
Provider
A. Direct
B. Provider + Cloud billing
C. Cloud-hosted (3P)
D. Self-hosting
OpenAI api.openai.com (Enterprise)
—
Azure Foundry, Bedrock
✗ (closed weights)
Anthropic api.anthropic.com (Enterprise)
Claude Platform on AWS (CCU billing)
Bedrock Claude, Vertex AI Claude
✗ (closed weights)
Meta Llama —
—
Bedrock, Vertex, Azure (hosted)
✅ open-weight (primary path)
Mistral api.mistral.ai
—
Bedrock, Azure, Vertex
✅ open-weight (partial)
Aspect
A. Direct
B. Provider + Cloud billing
C. Cloud-hosted
D. Self-hosting
New model availability
First
Near-simultaneous
Days to weeks delay
When open-weight released
Feature scope
Full
Full
Within cloud API
None (build yourself)
Fine-tuning
Full options
Full options
Limited or delayed
Complete freedom (own GPU)
VPC/Network isolation
Enterprise only
Within marketplace
Supported (check region/service)
Full isolation (own infra)
Commit consumption
No
Yes
Yes
No (infra costs only)
Operational burden
None
None
Low
High (GPU, model serving, updates)
Requirement
Suitable Pattern
Latest Claude Code + fast updates
A. Direct
Claude Code + consuming AWS EDP + AWS invoice
B. Provider service + Cloud billing (Claude Platform on AWS)
Claude API only + VPC isolation + existing AWS governance
C. Cloud-hosted (Bedrock Claude)
For the list of models and available channels by provider, see AI Platforms and Model Comparison . For Korean FMs (Upstage, EXAONE, etc.), see FM Provider Comparison (Korea) . For the pattern comparison table, use the Pattern Comparison section above.
Don’t fix your 1P/3P choice based on monthly spend alone. Work through the criteria below first, then choose a channel.
Criteria
1P (Direct) favored when
3P (Cloud-hosted) favored when
Features & release speed Latest models/agent features are needed immediately
Platform API scope is sufficient; a days-to-weeks delay is acceptable
Network & compliance Enterprise-grade isolation is sufficient
VPC/regional governance or inheriting existing certifications is required
Procurement & commits Little to no existing cloud commit
Significant EDP/EA/CUD commit consumption
Operations Comfortable operating the provider’s own console
Needs integration with existing cloud IAM, observability, and budget systems
Scale & quota PoC/small scale, quota headroom
Production traffic, reserved capacity, private offer negotiation
At large scale (roughly tens of thousands of dollars per month or more), procurement structure (commit consumption, reserved capacity, private offers) is often a bigger cost factor than per-token pricing. Dollar thresholds vary by organization, so confirm with official quotes and contract terms.
For channel choice driven by country-specific regulation and sovereignty requirements, see each country guide.
Each provider’s license tiers (Seat plans, API tiers), 3P reserved capacity (PTU, Provisioned Throughput), and cost management tools and patterns are covered in LLM Licensing and Cost Management .
Assuming “3P is expensive” — Published per-token rates are usually similar. With existing cloud commits, 3P may be cheaper in effective cost.All-in on a single channel — 1P only = weak cloud governance integration; 3P only = delayed access to latest features. Evaluate a hybrid approach.Not considering quotas upfront — Production traffic without pre-secured quota/rate limits leads to service outages.