Google's $40B commitment to Anthropic and Amazon's $25B follow-on deposited a single frontier AI lab into two competing infrastructure stacks at once, each securing something more valuable than equity: lock-in. Not Anthropic's — yours. Every enterprise that builds on Claude via Vertex AI or Bedrock is wiring itself into a dependency chain the model vendor doesn't fully control.
This is not a hypothetical risk. It is the structural outcome $65B in hyperscaler capital was explicitly designed to create. This analysis traces where AI platform lock-in is built, why it persists, and what the real cost of switching looks like.
The Architecture of Lock-In — Four Stacking Mechanisms
Compute entanglement, proprietary API differentiation, data gravity, and organizational embedding stack on top of each other. Google's $40B and Amazon's $25B to Anthropic each came with infrastructure clauses — the frontier model's compute dependency becomes the customer's compute dependency, invisibly, at every API call.
The Hyperscaler Matrix — AWS, Google, Azure Compared
AWS Bedrock locks in at the workflow layer (Agents, Guardrails, Model Evaluation) regardless of which model runs underneath. Google Vertex AI locks in vertically through data connectors and TPU-optimized checkpoints. Azure AI Foundry locks in through Microsoft 365 and compliance guarantees unavailable elsewhere.
The Capital Lock-In Dynamic
The $65B combined Google-Amazon commitment to Anthropic is not a financial stake — it's a compute reservation binding future model training to specific hardware roadmaps. The model is Anthropic's; the infrastructure revenue, and the customer's dependency, belongs to the cloud.
The Middleware Layer — Where Lock-In Is Manufactured
Layer 3 in the 10-Layer Framework is where lock-in is operationalized: agent orchestration, fine-tuning pipelines, guardrails, and vector-store integrations each add a switching-cost increment that compounds with deployment depth, even when the abstraction layer claims to be provider-agnostic.
The Open-Weight Escape — Real Option or Marketing Fiction?
Self-hosted open-weight models are a genuine escape path, but not a free one — the operational maturity to run frontier-scale inference reliably is rarer than open-source advocacy suggests, and the frontier capability gap has not closed.
Switching Cost Anatomy
Re-integration engineering, data egress, re-fine-tuning, operational disruption, and organizational retraining sum to an estimated $2M–$5M for a mid-sized enterprise running three production AI applications on one hyperscaler — before management time and competitive risk during the transition.
Strategic Implications
The most defensible position is a maintained portability budget: reviewing every six months which production deployments are provider-specific versus portable, and capping that ratio before migration becomes economically unviable.
Lock-in is real, partially intentional, and harder to reverse than the product marketing suggests — but it is not inescapable. The organizations that treat portability as a standing budget line, not a one-time migration project, are the ones that keep optionality past 2027.