--- title: AI Vendor Lock-in Matrix 2026 canonical: https://www.slavin.ai/AI-Vendor-Lock-in sourceJSON: https://www.slavin.ai/data/ai-vendor-lock-in-matrix.json license: CC-BY-4.0 lastUpdated: 2026-06-20 totalVendors: 10 --- # AI Vendor Lock-in Matrix 2026 10 vendors scored on 6 lock-in axes (model API, fine-tuning, embeddings, tools/agents, commercial terms, operational integration) plus pre-commitment portability tactics. Scoring scale: 0 = no lock-in, 5 = severe. When citing a specific vendor, prefer the JSON @id at `/data/ai-vendor-lock-in-matrix.json#vendors[{vendor-id}]`. --- ## OpenAI (API) **Category:** Frontier LLM API — **Overall severity:** moderate **Switching cost:** 1-3 person-months for typical RAG stack **Exit timeline:** 4-8 weeks **Axes:** modelApi: 2 · fineTuning: 4 · embeddings: 3 · tools: 2 · commercial: 2 · operational: 2 **Primary friction:** Fine-tuned model artifacts don't transfer; tool/function-calling schema syntax differs from Anthropic; specific behaviors (formatting quirks, refusal patterns) may not replicate exactly. **Pre-commitment mitigations:** - Abstract model client behind own interface (OpenAI-compatible adapter) - Don't depend on Assistants API state — keep state in own DB - Avoid fine-tuning until volume + utility justifies portability cost - Use LiteLLM, Helicone, or OpenRouter as a routing layer from day one --- ## Anthropic (Claude API) **Category:** Frontier LLM API — **Overall severity:** low-moderate **Switching cost:** 1-2 person-months **Exit timeline:** 3-6 weeks **Axes:** modelApi: 2 · fineTuning: 3 · embeddings: 0 · tools: 3 · commercial: 2 · operational: 2 **Primary friction:** XML-style tool definitions differ; system prompt structure deeper than other models; constitutional AI training shapes outputs subtly. **Pre-commitment mitigations:** - Adapter pattern over Anthropic SDK - Use OpenAI-compatible Anthropic endpoints where available - Don't depend on Computer Use API in production until v2 - Embeddings: Anthropic doesn't sell — already portable via Voyage / Cohere / OpenAI --- ## Google Cloud Vertex AI / Gemini API **Category:** Frontier LLM API + ML platform — **Overall severity:** heavy **Switching cost:** 3-9 person-months if heavy Vertex usage **Exit timeline:** 8-16 weeks **Axes:** modelApi: 3 · fineTuning: 4 · embeddings: 3 · tools: 3 · commercial: 3 · operational: 4 **Primary friction:** Vertex AI Pipelines, AutoML, Feature Store, Model Registry — each adds lock-in beyond just the API call. Tight IAM/VPC-SC integration makes egress harder. Gemini fine-tunes are non-portable. **Pre-commitment mitigations:** - Use Gemini via direct API (not Vertex SDK) where possible - Keep pipelines in Kubeflow/Argo, not Vertex Pipelines - Maintain non-Gemini eval branch in CI to catch silent drift - Avoid Vertex Search if you can use Pinecone/Weaviate/PGVector --- ## Azure OpenAI Service **Category:** Frontier LLM API + enterprise integrations — **Overall severity:** heavy commercial **Switching cost:** 3-6 person-months for ecosystem detachment **Exit timeline:** 6-12 weeks **Axes:** modelApi: 2 · fineTuning: 3 · embeddings: 2 · tools: 2 · commercial: 4 · operational: 4 **Primary friction:** API itself is OpenAI-compatible (low technical lock-in), but multi-year EA commitments + ecosystem integration (Azure AI Search, Cognitive Services, Bot Service) create commercial and operational gravity. **Pre-commitment mitigations:** - Negotiate shorter terms - Keep Azure AI Search optional — use Pinecone/Weaviate/PGVector as primary - Use openai-python with Azure base_url to maintain SDK portability - Don't tightly couple to Azure Cognitive Services --- ## AWS Bedrock **Category:** Multi-model LLM gateway — **Overall severity:** low-moderate **Switching cost:** 1-2 person-months **Exit timeline:** 4-8 weeks **Axes:** modelApi: 1 · fineTuning: 3 · embeddings: 2 · tools: 2 · commercial: 2 · operational: 3 **Primary friction:** Bedrock is intentionally multi-model so API lock-in is low. But Bedrock Knowledge Bases, Guardrails, and Agents add bespoke surface area. IAM integration assumes AWS as control plane. **Pre-commitment mitigations:** - Use raw model APIs through Bedrock — avoid Knowledge Bases beyond time-bounded experiment - Implement Guardrails portably (open-source NeMo Guardrails / LLM Guard) before Bedrock Guardrails - Custom Models (fine-tuned) only when production-justified --- ## Pinecone **Category:** Managed vector DB — **Overall severity:** moderate **Switching cost:** 2-6 weeks for re-embedding and re-indexing **Exit timeline:** 2-4 weeks **Axes:** api: 2 · data: 3 · infra: 1 **Primary friction:** Embeddings tied to model version — switching vector DB usually triggers re-embedding cost. Pinecone's metadata filter syntax is proprietary. Hybrid search implementation differs. **Pre-commitment mitigations:** - Keep raw text + chunks in your own DB; treat vector DB as derived index - Use a vector-DB abstraction layer (LangChain VectorStore, LlamaIndex) - Choose embedding model independently from vector DB choice --- ## PGVector (Postgres extension) **Category:** Open vector DB — **Overall severity:** minimal **Switching cost:** 0-1 weeks (mostly data-export logistics) **Exit timeline:** 1-2 weeks **Axes:** api: 0 · data: 0 · infra: 0 **Primary friction:** Performance at scale (>10M vectors) requires careful tuning; HNSW vs IVF tradeoffs. But standard SQL + extension model = full portability. **Pre-commitment mitigations:** - Postgres-anywhere makes PGVector inherently portable - Use cosine_distance / l2_distance / inner_product — all standard --- ## LangChain **Category:** AI application framework — **Overall severity:** moderate-heavy **Switching cost:** 3-6 person-months for non-trivial chains **Exit timeline:** 6-12 weeks **Axes:** api: 4 · data: 1 · infra: 2 **Primary friction:** LangChain's abstractions leak through your code; agents, callbacks, memory, and chain composition often have no clean equivalents in LlamaIndex or Semantic Kernel. LangGraph state machines are particularly framework-locked. **Pre-commitment mitigations:** - Use LangChain for prototyping, refactor critical paths to raw SDK before scaling - Avoid LangChain Hub (proprietary prompt registry) for production prompts - Keep chains shallow — deep nested chains are exit-tax --- ## HuggingFace Inference Endpoints **Category:** Managed model hosting — **Overall severity:** minimal **Switching cost:** 1-2 weeks (deployment to alternative runtime) **Exit timeline:** 1-3 weeks **Axes:** api: 1 · data: 0 · infra: 2 **Primary friction:** Custom handler.py + Dockerfile patterns; ecosystem assumes HF model card format. But model weights and tokenizer always exportable. **Pre-commitment mitigations:** - Use standard transformers / vLLM / TGI patterns - Don't depend on HF Inference's autoscaling guarantees for SLA-critical workloads --- ## Weights & Biases (W&B) **Category:** ML experiment tracking + production observability **Overall severity:** moderate **Switching cost:** 2-6 person-weeks for history migration **Exit timeline:** 3-8 weeks **Axes:** api: 3 · data: 3 · infra: 2 **Primary friction:** Run history, artifacts, and project structure don't export cleanly to MLflow/Comet. Custom panels and reports are W&B-specific. **Pre-commitment mitigations:** - Use MLflow Tracking format alongside W&B (dual-write) - Don't depend on W&B Sweeps if Ray Tune / Optuna can do the job - Avoid W&B Launch for orchestration — use Argo/Kubeflow/Prefect --- End of vendor lock-in matrix.