モデルプロバイダへの支配権移転の罠

imageWhy cheaper compute won't mean cheaper AI, and what your stack is risking right now.

OpenAI projected operating losses of $74 billion in 2028 alone, before an expected pivot to profitability around 2030. Deutsche Bank research calculates that OpenAI's projected cumulative cash burn could exceed $200 billion by 2030 – making it the largest startup loss in business history.

Every token you've sent at subsidized rates has been financed by venture capital and hyperscaler infrastructure deals, with a specific expectation attached: by the time the bill comes due, your switching costs will exceed the cost of paying it.

That bill is arriving now.

The Playbook

Every major model provider has been running the same three-phase strategy.

Flood cheap access. Give away functionality to build dependency. Free access, all you can eat, unknown limits.

Embed deeply. Make the AI and the workflow inseparable. Copilot inside Microsoft Office. Claude inside your IDE. The goal is to make the boundary between your product and their infrastructure invisible before you think to look for it.

Upcharge at renewal. Once switching costs are measured in rewrites, retraining, and workflow reconstruction, extract value by exerting pricing power.

The evidence that Phase 3 has begun is already in the market. OpenAI removed GPT-4o from its free plan and gated it behind a paywall. Anthropic priced its large-context Claude 4 models with a tiered structure that charges a premium at exactly the token volumes enterprise workloads require. Both companies introduced rate limits that create upsell pressure precisely as your teams are ramping usage.

One layer down, the developer tools that charge for inference and build on these providers are absorbing cost increases until they can't. Cursor moved from a 500-request pricing plan to a credit pool billed for overage at API rates with little to no notice, issuing a public apology when developers ran out of requests after a handful of Claude 4 prompts. The reason was stated explicitly: newer models cost more, and the cost had to be passed through. Windsurf announced Wednesday that they were following suit, citing the same reason.

This is the inverted house of cards that is the revenue model of every developer tool that charges for and resells inference. Cost pressure at the model layer inevitably cascades to developer tools, then to your teams, with each layer having less leverage than the one above it.

Pricing Power and Lock-In Are the Same Problem

Model providers don't reprice arbitrarily. They reprice when your switching costs exceed the cost of absorbing the increase. That calculation gets more favorable to them with every month your teams spend building against their models. Every prompt optimized against one model's behavior, every evaluation pipeline calibrated to its output, every parsing layer built around its response structure transfers pricing leverage from your organization to your provider.

Token prices falling at the commodity end of the market is real, but it's happening where switching costs are lowest: new customers, early evaluations, proof-of-concepts. The moment a workload moves to production (with prompt libraries, evaluation harnesses, and institutional knowledge built against a specific model), the switching cost curve inflects sharply upward. That's when the rate limits appear, the context window premiums arrive, and the capability gating begins. Your costs of leaving now exceed the cost of absorbing the increase. The pricing follows the lock-in.

Cursor (and now, Windsurf) illustrate what happens one layer down. Both built products on top of expensive models. Both priced those products before understanding what sustained usage would actually cost. When inference prices rose, they couldn't absorb the increase and couldn't substitute cheaper models without breaking the workflows their users depended on. The only move left was to pass the cost through. Their users' lock-in became their lock-in.

The feedback loop runs in one direction: teams build, lock-in accumulates, pricing leverage transfers to providers, providers reprice, switching costs rise further. Each cycle makes the next more likely.

The Lock-In Nobody Is Accounting For

Cloud vendor lock-in is a known, bounded problem. Your workloads run on AWS. Moving them to GCP is expensive, but the scope is visible before you start.

AI inference lock-in has no visible perimeter. It accumulates silently across every team that touches a model, compounds at every layer of your stack, and reveals its true scope only when you try to leave. Switching providers is a rewrite that reveals its own cost as you go. Organizations egregiously underestimate that cost, because most of it never appears in a project tracker. It surfaces as unexplained engineering drag in the quarters that follow.

Windsurf users were in this position last year. When acquisition conversations with OpenAI fell through, competitive tensions with Anthropic surfaced immediately and developers were caught in the middle. Free tier users lost access to Claude 3.x models with five days' notice. Paid subscribers faced severe capacity constraints. The engineers affected had simply built their workflows on a tool whose model access was controlled by a third party whose incentives changed overnight. A corporate event, with no recourse.

Gartner projected that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs and inadequate risk controls. These failure modes become substantially worse with no fallback and no visibility into the reliability of the LLM layer your stack depends on.

The Architectural Response

Dependency that isn't visible, managed, or distributed doesn't stay neutral. It becomes leverage held by whoever sits above you in the stack. The architectural response isn't about anticipating every failure mode. It's about ensuring that when those failure modes materialize, your organization has options rather than obligations.

The answer is building the infrastructure layer underneath your AI usage with the same discipline you'd apply to any other critical dependency.

In traditional software engineering, you abstract your database connection rather than hardcode it. You maintain redundancy in critical services. You instrument the things that can fail quietly, not just the things that fail loudly. Applied to inference and model providers, those same assumptions produce four properties:

Substitutability. Prompt logic, evaluation pipelines, and output parsing built against a normalized interface. Switching providers should require a configuration change. Any architecture where swapping your inference provider touches more than the configuration layer has already accumulated debt.

Optionality. Active integrations with multiple providers so failover is operational. The cost of maintaining multiple integrations is lower than the cost of a single unplanned outage or price renegotiation at renewal.

Observability. Standard monitoring tells you when a provider is down. It doesn't tell you when output quality has dropped, when response structure has drifted, or when the model version serving your traffic changed without notice. Quality metrics against production traffic are the difference between knowing your inference layer is working and assuming it is.

Auditability. Visibility into what data crosses your inference boundary, what was sent to which provider, and what the model did with it. For security and compliance teams, this is the precondition for governance.

These are the principles Cline was designed around. Model agnostic, inference agnostic, deployment agnostic. Runnable directly on your machine, connecting to providers through your own API keys, with full visibility into tokens consumed and costs incurred at the task level.

The specific model or provider you're using today is just an implementation detail. The interface your architecture is built against is the thing that lasts. Build against the capability and the pricing cycle your competitors are trapped in becomes a problem you can navigate rather than absorb.

The question isn't whether your organization has accumulated inference lock-in. It has. The question is whether you find out on your own timeline, or someone else's. The thread running through all of this is a single architectural conviction: your engineering capability is too important to be mediated by someone else's incentives.