AI Vendor Concentration Creates a New Resource Dependency

I wrote recently about how GPU supply is a textbook resource dependency problem, with NVIDIA at the narrowest point of the AI compute chain. The more I think about that argument, the more I realize I only told half the story. The GPU market is the upstream dependency. The downstream dependency, sitting directly in front of every organization that builds on AI, is even more concentrated and harder to escape. I mean the model provider layer. OpenAI, Anthropic, Google, and Meta control the frontier models. That is four organizations, really three private companies and one that gives the technology away for strategic reasons. No critical organizational resource in IS history has been this concentrated.

Pfeffer and Salancik published The External Control of Organizations in 1978. Sutton and Staw (1995), in their theory quality paper that I have in my library, discuss how Pfeffer and Salancik showed that power becomes a stronger predictor of resource allocation under conditions of uncertainty. That is the core mechanism. When an organization depends on an external actor for a critical resource, and when that resource is controlled by few suppliers, the dependent organization loses discretion over its own strategy. Pfeffer and Salancik identified two dimensions that determine the strength of the dependency. Resource criticality: how essential the resource is to the organization's operations. Resource concentration: how many alternative sources exist. When both are high, as they are in the AI model market, the dependency is structural.

The thing that made this concrete for me was remembering the OpenAI governance crisis in November 2023. The board fired Sam Altman on a Thursday. By Monday, every company that built on OpenAI's API was in a state of emergency. I watched teams scramble to test alternative providers, switch their evals to Anthropic or open-source models, and renegotiate contracts. The panic was not about boardroom drama. It was about resource dependency being exposed in real time. Those organizations had built their products, their customer experiences, and their internal workflows around a single provider's API. When that provider suddenly looked unstable, the dependency became visible. It had always been there, they just had not felt it until the moment they could not assume access anymore.

This is why I think the multi-model strategies that have become standard practice are not about technical excellence. They are dependency-reduction strategies. Every architecture that routes prompts to a mix of GPT, Claude, Gemini, and open-weight models is doing exactly what Pfeffer and Salancik prescribed: reducing concentration by diversifying the resource base. Every organization that maintains redundant evaluation pipelines across providers is paying a switching cost insurance premium. The cost is real, maintaining multiple integrations, prompt templates, guardrails, and monitoring systems across providers is expensive and technically demanding. But the alternative is the risk that a governance crisis, a pricing change, a terms-of-service revision, or a model deprecation disables your product overnight. I wrote about this dynamic before in the context of vendor lock-in as a deliberate strategy. In AI models, the lock-in is not deliberate in the same way, the provider does not need to design exit barriers because the technology creates them naturally. From a resource dependency lens, the multi-model architecture looks less like a sophisticated technical choice and more like basic organizational survival.

The switching costs are structural, not incidental. Fine-tuning locks model-specific knowledge into the weights. Prompt engineering creates provider-specific syntax and behavior expectations. Evaluation pipelines measure against a specific model's output distribution. Every layer of integration increases the cost of leaving. Organizations that invested deeply in one provider's ecosystem are not staying because of quality. They are staying because the accumulated switching costs exceed the perceived risk of dependency. That is exactly the condition that Pfeffer and Salancik described: dependency persists when alternatives exist but are not practically accessible.

The concentration is increasing, not decreasing. The capital required to train a frontier model is now in the billions. Moore's law historically reduced the cost of computing over time. In AI training, the opposite is happening. Each generation of model requires more compute, not less. New entrants cannot raise the capital or secure the GPU allocation needed to compete at the frontier. The number of organizations that can train a GPT-4-class model is smaller today than it was two years ago, not larger. Wade and Hulland (2004) built on Pfeffer and Salancik to argue that environmental dimensions including munificence shape competitive dynamics. The AI model market is not munificent. It is resource-constrained at the point of production and concentrated at the point of access. Organizations that need frontier models face an environment where the resource they need is becoming scarcer and more controlled.

This is also why the push for open-weight models like Llama and Mistral is a collective response to dependency. Open-weight models reduce concentration not by creating more suppliers in the traditional sense but by making the resource itself portable. An organization that deploys Llama on its own infrastructure still depends on Meta's initial training investment, but it does not depend on Meta's API availability, pricing, or governance stability. Tushman and Anderson (1986), who cite Pfeffer and Salancik in their argument about environmental uncertainty and resource constraints, would recognize this as a competence-destroying shift in the resource base. Open-weight models do not just compete with proprietary APIs. They change the dependency structure by moving the resource from an externally controlled service to an internally deployable asset. That is why every major corporation now runs its own model infrastructure. It is not because self-hosting is cheaper or technically superior. It is because the dependency cost of relying on a single API is higher than the infrastructure cost of running your own models.

The pattern I keep seeing is familiar from the GPU post. The same Pfeffer and Salancik logic appears at every layer. The company that trains models depends on GPU suppliers. The company that builds applications depends on model providers. The company that offers an AI feature to customers depends on both. The actor at the narrowest point in each chain holds the power. And the textbook recommendation is the same at every layer. Reduce dependency through vertical integration, which means building your own models or your own compute, which most organizations cannot afford. Or reduce dependency through multi-sourcing, which is what multi-model strategies and open-weight deployments achieve, partially and imperfectly. Or accept the dependency and manage the relationship, which is what most organizations are doing without admitting it.

My opinion is that most organizations building on AI today are not doing either strategy seriously enough. They run a single API with a fallback script and call it multi-model. They deploy one open-weight model for a non-critical task and call it diversified. The concentration at the model provider layer is the most extreme resource dependency the IS field has ever studied, and the organizational response looks like it would in any other context: denial, followed by incremental hedging, followed by crisis. Pfeffer and Salancik wrote their book almost fifty years ago. The advice still applies. Dependence on a concentrated supplier of a critical resource is not a technical problem. It is a structural vulnerability. And it will not resolve itself.---
verification:
claims_checked:
- "Pfeffer and Salancik (1978) argued that under uncertainty, power predicts resource allocation: supported by Sutton and Staw (1995, lines 453-455) in local library at Final Paper/ALL_PAPERS_MD/[6010] Sutton&Staw_1995_ASQ/paper.md"
- "Pfeffer and Salancik (1978) cited for environmental uncertainty definition: supported by Tushman and Anderson (1986, lines 514-516) in local library at Final Paper/ALL_PAPERS_MD/[6670] Tushman+&+Anderson+1986/paper.md"
- "Pfeffer and Salancik (1978) cited as foundation for environmental dimensions: supported by Wade and Hulland (2004, lines 1806-1813) in local library at Final Paper/ALL_PAPERS_MD/[6010] Wade and Hulland_2004_MISQ/paper.md"
- "Tushman and Anderson (1986) environmental uncertainty and resource constraint logic: supported by paper in local library (lines 514-529)"
- "Sutton and Staw (1995) discuss Pfeffer and Salancik as example of strong theory: supported by local library (lines 453-461)"
- "Van Osch et al. (2023) cite Pfeffer and Salancik (1978) for cooperation challenges: supported by local library (lines 726-728)"
- "Kudesia lists resource dependence as one of three influential 1970s organizational theories: supported by local library (lines 225-228)"
claims_unverified:
- "OpenAI governance crisis (November 2023): widely reported industry event, not from local source"
- "Multi-model strategies as dependency-reduction: author's theoretical interpretation, not from local source"
- "Open-weight model push (Llama, Mistral) as collective dependency response: author's interpretation"
- "Frontier model training costs in billions: industry reporting"
- "Model provider market concentration: industry observation"
- "Fine-tuning, prompt engineering, evaluation pipelines as switching costs: author's framework"
- "Number of organizations that can train frontier models decreasing: industry observation"
sources_used:
- "local: [6010] Sutton&Staw_1995_ASQ/paper.md (lines 448-462)"
- "local: [6670] Tushman+&+Anderson+1986/paper.md (lines 510-529)"
- "local: [6010] Wade and Hulland_2004_MISQ/paper.md (lines 1800-1819)"
- "local: [6010] Van Osch et al. (2023) JAIS/paper.md (lines 726-728)"
- "local: [6670] Organizational Sensemaking Ravi S Kudesia/paper.md (lines 225-228)"
internal_links:
- "/blog/gpu-dependency-resource-dependency-theory"
- "/blog/vendor-lock-in-is-a-strategy"
- "/blog/open-weight-ai-digital-commons"
word_count: 1191
---