We have been assuming that frontier AI requires frontier investment. Massive compute. Billions in training costs. But the most disruptive force may not be what AI can do but how cheaply it can do it. In September 2025, DeepSeek published a peer-reviewed paper in Nature reporting that its R1 reasoning model was trained for $294,000 on 512 Nvidia H800 chips in 80 hours, with the full foundational model costing approximately $5.87 million (Liang et al., 2025). That model has been downloaded over 10 million times as open-weight software. Meanwhile, Microsoft's Phi-4 Mini achieves competitive reasoning performance with 3.8 billion parameters, a fraction of frontier model scale. The compute moat is dissolving. The question is what replaces it.
The hypothesis: as AI training and inference costs drop exponentially and small language models close the performance gap with frontier systems, competitive advantage is migrating from access to compute toward organizational learning speed, integration depth, and human-AI workflow design.
Three Takeaways
First, the cost disruption follows the economics of general-purpose technologies, but the compression timeline is unprecedented. Bresnahan and Trajtenberg (1995) defined general-purpose technologies as innovations characterized by pervasiveness, improvement over time, and the spawning of complementary innovations. Electricity, semiconductors, and the internet all followed similar trajectories: initial high costs limited access, rapid cost declines democratized the technology, and competitive advantage shifted from access to application. AI is compressing this trajectory from decades into months. DeepSeek's R1 demonstrated frontier-competitive reasoning at a cost two orders of magnitude below the prevailing assumption. The small language model category is accelerating the compression further: purpose-built models with 3–7 billion parameters now match or exceed the performance of general-purpose models ten times their size on domain-specific tasks, while running on edge devices without cloud dependency.
Second, cost democratization shifts the strategic question from whether you can afford frontier AI to whether you can use it well, and open source is the mechanism accelerating that shift. Penrose's (1959) theory of the growth of the firm argued that the binding constraint on growth is managerial capacity to productively deploy resources, not resource availability itself. When AI becomes cheap, the scarce resource is organizational intelligence. Open-source and open-weight models are the vehicle through which cost democratization reaches organizations: DeepSeek R1 is open-weight, Meta's Llama 4 is open-source with mixture-of-experts architecture, and Microsoft's Phi family is openly available. The strategic implication is that the model layer is becoming infrastructure, and organizations that built competitive strategy around proprietary model access are exposed. If the binding constraint is organizational, then the organizations that design their human-AI systems deliberately will outperform those that simply procure better technology.
Third, cost compression accelerates commoditization of the model layer, and value migrates toward the layers where organizational capability determines outcomes. Porter's (1985) Five Forces framework predicts that when barriers to entry fall and products become substitutable, value migrates to adjacent layers. In AI, value is migrating from the model layer toward integration, orchestration, and workflow design. The small language model trend reinforces this: when a 3.8 billion parameter model fine-tuned on domain-specific data outperforms a general-purpose frontier model on the task that matters, the competitive question is no longer which model you have access to but how well your organization can select, fine-tune, compose, and govern an ecosystem of fit-for-purpose models. That is an organizational capability, not a procurement decision.
The Longer View
The semiconductor industry provides the clearest historical analog through Moore's Law. Gordon Moore (1965) observed that transistor density doubled approximately every two years, driving exponential cost declines in computing. Each wave of cost reduction unlocked new categories of application and shifted competitive advantage from hardware access to software and integration capability. AI is following this pattern with an additional dynamic: open-source release accelerates diffusion by removing the licensing constraint entirely. When DeepSeek published R1 as open-weight, it did not just reduce cost. It eliminated the access barrier for anyone with the organizational capability to deploy it.
Carlota Perez's (2002) framework on technological revolutions identifies the structural transition. An installation phase driven by financial capital is followed by a deployment phase driven by production capital. The installation phase favors technologists and investors. The deployment phase favors integrators and operators. DeepSeek, small language models, and open-source release are collectively signaling the turn from installation to deployment. Competitive advantage is shifting decisively toward organizations that know how to weave AI into the fabric of their operations.
Henderson and Clark's (1990) concept of architectural innovation identifies why incumbents miss this shift. Architectural innovations change how components relate to each other without necessarily changing the components themselves. Small language models, open-weight distribution, and model composition represent an architectural innovation: the individual components are familiar, but the way they are combined, deployed, and governed is fundamentally different from the frontier-model paradigm. Organizations built around procuring the biggest model from the most prestigious vendor may find themselves architecturally mismatched to a world where the winning strategy is composing fit-for-purpose models into integrated workflows.
My Two Cents
I wrote about this question before DeepSeek published in Nature. That paper confirmed what the structural analysis suggested: the compute moat was always thinner than the investment narrative implied. But the small language model trend adds a dimension I did not fully anticipate. It is not just that frontier capability is getting cheaper. It is that the definition of frontier is shifting from largest to most fit-for-purpose. An organization that deploys a fine-tuned 7 billion parameter model on its own infrastructure, trained on its own domain data, governed by its own policies, may have a more durable advantage than one paying for API access to the largest model available. The moat is organizational, and it always was.
Stress-test your AI strategy against a scenario where model costs drop by another order of magnitude within eighteen months. If your strategy depends on access to expensive capability, it is fragile. If it depends on your organization's ability to learn, integrate, compose, and govern faster than competitors, it is durable. Build for the second scenario.
Read to Learn More
Academic: Bresnahan, T. F., & Trajtenberg, M. (1995). General purpose technologies: Engines of growth? Journal of Econometrics, 65(1), 83–108.
Industry: Liang, W., et al. (2025). DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. Nature.
References
Bresnahan, T. F., & Trajtenberg, M. (1995). General purpose technologies: Engines of growth? Journal of Econometrics, 65(1), 83–108.
Henderson, R. M., & Clark, K. B. (1990). Architectural innovation: The reconfiguration of existing product technologies and the failure of established firms. Administrative Science Quarterly, 35(1), 9–30.
Liang, W., et al. (2025). DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. Nature.
Moore, G. E. (1965). Cramming more components onto integrated circuits. Electronics, 38(8), 114–117.
Penrose, E. T. (1959). The theory of the growth of the firm. Wiley.
Perez, C. (2002). Technological revolutions and financial capital. Edward Elgar.
Porter, M. E. (1985). Competitive advantage. Free Press.