Apple's announcement that it will integrate Google's Gemini models into its AI infrastructure rather than building proprietary large language models represents more than cost optimization. According to Financial Times reporting, this multibillion-dollar deal positions Apple as a kingmaker in enterprise AI while deliberately avoiding the infrastructure arms race. The decision reveals a structural pattern that organizational theory has yet to adequately address: how do enterprises develop adaptive expertise in algorithmic coordination when the underlying models themselves are treated as interchangeable commodities?

The simultaneous news that business spending on OpenAI models reached record levels in December 2025, with Ramp data showing OpenAI significantly outpacing Anthropic and Google in paid enterprise usage, creates an apparent paradox. If Apple's strategic calculus is correct and model providers are functionally substitutable, why do enterprises exhibit such pronounced concentration in their AI vendor selection? The answer lies in what I call the topography trap in platform coordination.

The Substitutability Illusion in Enterprise AI

Apple's approach treats foundation models as infrastructure layers where competitive advantage derives from integration and application rather than model ownership. This mirrors classical make-or-buy decisions in organizational economics. However, the concentrated spending patterns on OpenAI suggest enterprises are not actually treating models as substitutable inputs. They are developing what Hatano and Inagaki (1986) would classify as routine expertise: procedural knowledge optimized for a specific platform rather than adaptive expertise that transfers across model architectures.

This distinction matters because it reveals the coordination mechanism at work. When enterprises invest heavily in prompt engineering, fine-tuning, and workflow integration with a single model provider, they are building topographical knowledge (how to navigate this specific terrain) rather than topological understanding (the shape of constraints that govern all such systems). The organizational cost of switching providers is not merely contractual or financial. It is the accumulated procedural knowledge that does not transfer.

Why Schema Induction Fails in Enterprise Adoption

The concentration on OpenAI despite the theoretical substitutability of models suggests enterprises lack structural schemas for reasoning about algorithmic coordination across platforms. Kellogg, Valentine, and Christin (2020) document how algorithmic awareness among workers does not translate to improved outcomes. The enterprise pattern shows the same dynamic at the organizational level. CTOs understand that models are probabilistic systems with similar capabilities, but this awareness does not produce adaptive procurement or integration strategies.

Apple's infrastructure arbitrage works precisely because Apple is not developing deep procedural expertise with any single model. By positioning itself as an orchestration layer, Apple maintains the optionality that comes from topological rather than topographical knowledge. The company understands the structural constraints of AI-mediated interaction without binding itself to the specifics of any implementation.

The Organizational Learning Trap

This creates a troubling implication for organizational theory. The variance puzzle in platform coordination holds that workers with identical access show dramatically different outcomes, with power-law distributions emerging from algorithmic amplification of initial differences (Schor et al., 2020). At the enterprise level, we observe the inverse: organizations with vastly different resources converge on similar vendor concentration despite having the capital and technical capability to diversify.

The mechanism appears to be path dependence created by routine expertise accumulation. Early adoption of OpenAI's APIs creates organizational knowledge that is optimized for that specific platform. Subsequent investments deepen this specialization. The awareness that other models exist and may be superior does not overcome the coordination costs of switching when procedural knowledge is context-specific.

What Apple's Abstraction Strategy Reveals

Apple's deliberate positioning as a model-agnostic orchestration layer represents organizational design for adaptive rather than routine expertise. By not optimizing for any single model's specifics, Apple maintains what Gentner (1983) would describe as structural alignment: the capacity to map relationships across different implementations because the underlying schema focuses on invariant features rather than surface particulars.

The question for organizational theory is whether this abstraction strategy is available only to platform orchestrators or whether it represents a generalizable principle for enterprise AI adoption. If schema induction (teaching structural features of algorithmic coordination) can be systematically developed, enterprises should be able to maintain strategic optionality across model providers. If not, the current vendor concentration represents a stable equilibrium where switching costs continually increase as procedural knowledge deepens.

The divergence between Apple's infrastructure arbitrage and enterprise vendor concentration suggests we are observing two distinct coordination mechanisms operating simultaneously. One builds adaptive expertise through abstraction; the other accumulates routine expertise through specialization. Which mechanism dominates will determine whether AI capabilities become organizational core competencies or permanent sources of vendor lock-in.