Why Closed-Source LLM Business Models Are Under Pressure: From Model Rentals to Workflow Sovereignty

1. Starting Point: The Logic That Once Made Closed-Source Models Work

For the past several years, the commercial model for closed-source LLMs has rested on a core assumption: the capability advantage of leading models is a sustainable scarce resource. The classic narrative was roughly: “My model is months ahead; you cannot catch up. So if I charge by usage, I can keep collecting rent.”

That logic did hold when the technical gap was clear and the open-source catch-up cycle was long. But in the past year or two, two fundamental changes have taken place:

1. Open-source model capability curves are approaching quickly. On many everyday office tasks—text drafting, meeting-note summarization, spreadsheet cleaning, document polishing—open-source models can already reach most of the performance level of closed-source flagship models, while deployment costs may be an order of magnitude lower.
2. The enterprise-side “good enough” mindset is winning. For the vast majority of non-core business scenarios, users barely notice the difference between 90 points and 95 points, but the cost difference is real and shows up in financial statements.

As a result, a common industry phenomenon has emerged: dynamic routing has become standard in enterprise AI architectures.

2. First Narrowing: When Intelligence Becomes a Commodity, API Rentals Stop Working

A typical enterprise LLM gateway now looks like this:

def route(task):
    if task.complexity == "high":
        return call_frontier_model(task)   # Complex reasoning, long documents, agent orchestration
    else:
        return call_local_open_source(task)  # Email, meeting notes, spreadsheet processing

Behind these few lines of code is a harsh commercial reality: usage of closed-source flagship models is structurally compressed into a small number of high-complexity scenarios. Once “intelligence” itself becomes a commodity that can be purchased in bulk, charging a capability premium through a black-box API no longer adds up.

Closed-source vendors have only one direction of response: climb upward. Instead of being model peddlers, they aim to become an “office operating system,” packaging model capabilities into complete work scenarios and replacing pure capability premiums with productization premiums.

3. Second Narrowing: Building an AI Office Means Contesting Workflow Sovereignty

The upward path runs directly into the heartland of traditional office suites. The approach of a new generation of AI office products is quite clever: instead of directly claiming to “replace a certain software,” they first seize the entry point from discussion to delivery:

• After a whiteboard discussion, directly generate a structured table;
• After a table is edited, move into a data dashboard with one click;
• During a meeting, AI simultaneously outputs action items and to-dos.

The value of this entry point is far greater than that of any single software product. Whoever controls the flow from discussion to delivery controls workflow sovereignty.

But this path is also full of traps:

1. The closed-source camp is not a monolithic bloc. Models from different vendors are already embedded in one another, office ecosystems and model ecosystems cross-license, and competitive boundaries are blurring.
2. The stickiness of AI office is not in the model but in the Harness. What truly retains enterprises is process orchestration, permission systems, data connectors, and interfaces that fit enterprise habits—none of which are moats built on model capability.
3. Enterprises will deliberately adopt multi-model routing. Once the Harness is in place, smart enterprises will design an abstraction layer that makes underlying models pluggable: today model A, tomorrow model B, the day after tomorrow a self-deployed open-source model. Switching costs are deliberately kept low, and vendor lock-in is actively dismantled.

A pragmatic architecture recommendation is that when building AI workflows, enterprises should treat the model layer as a replaceable component and isolate differences through a unified interface abstraction:

workflow:
  steps:
    - name: summarization
      provider: local-open-source   # Default low-cost option
    - name: deep-analysis
      provider: frontier-api        # Call only for high-value tasks

4. Fundamental Conclusion: Models Are Replaceable; Workflows Are Valuable

Looking back at the two narrowings of the closed-source model, the essence is the same structural problem:

• There is no forced entry point in the AI era. In the past, OS-level suite binding relied on file formats, installation habits, and IT procurement inertia. AI workflows are naturally heterogeneous, and models are fluid.
• Workflow is the value anchor, but the open-source side can grow it too. Model capability will continue to commoditize, while the Harness around workflows—processes, permissions, and data connections—can be built by closed-source vendors, and can equally be built by open-source communities and in-house teams.

The implication for enterprises is clear: when building AI capabilities, prioritize investment in the workflow abstraction layer and data governance rather than binding to any single model vendor. The implication for model vendors is harsher: the window for pure capability leadership is getting shorter, and future moats must be built outside the model—scenario understanding, enterprise-grade delivery capability, and deep integration that is difficult for routing algorithms to replace.

The road for closed-source LLMs is not impassable, but it is shifting from “selling intelligence” to “building systems.” Building systems means facing an open battlefield where there is no longer a Windows-style forced entry point.