[YOU] on AI asks who controls the infrastructure of human intelligence. Commodity complementation is the mechanism by which that control is contested. When a powerful model layer is freely available, the value of intelligence migrates to the data, the distribution, and the user-facing applications built on top—assets that are not equally distributed. Granovetter’s framework would observe that open weights democratize the bridging capital of AI capability while leaving the social graph and the behavioral data that sit above it as concentrated as before.
The geopolitical dimension is real and underappreciated. Open models have become a resource for AI development programs in countries that cannot afford to license proprietary frontier models. The developer in Lagos, the research institute in Bangalore, the national AI program without Google’s or OpenAI’s resources—all gain meaningful capability from Llama’s open release. Whether this constitutes genuine democratization or whether it reproduces the Free Basics dynamic—making a curated version of the infrastructure freely available while the company retains control of the layers that matter most—is the open question the cycle cannot yet answer.
The concept is most systematically articulated in the work of Joel Spolsky, whose 2002 essay “Strategy Letter V” formalized the logic: smart companies make the products that complement their core product into commodities. The historical cases he analyzed—Netscape open-sourcing the browser, Apple championing open standards against Microsoft, IBM supporting Linux—each illustrate the same underlying dynamic: commoditizing the complement strengthens the proprietary layer.
Ben Thompson’s Stratechery analysis of Meta’s Llama strategy extended this framework to the AI era: Meta’s commodity is the model, Meta’s proprietary moat is three billion users generating continuous behavioral data. The open-source model benefits Meta not by generating direct revenue but by ensuring that no competitor can charge rent for the model layer that Meta’s products run on. Zuckerberg made the logic explicit in his 2024 essay, framing the commodity-complementation strategy in the language of democratization without using the economic terminology.
The complement must be identified correctly. For Meta, the AI model layer is a complement to the distribution layer (three billion users) and the hardware layer (custom silicon, data centers, wearable devices). Making the model free strengthens both. For a company whose core business is the model itself—OpenAI, Anthropic—the same strategy would be self-defeating. The logic is structurally sound only for the entity whose business sits outside the layer being commoditized.
Democratization and competition produce the same policy until they diverge. Zuckerberg’s open-source argument is both genuinely democratizing—it has lowered barriers for researchers, startups, and national AI programs—and structurally advantageous for Meta. Both things are simultaneously true. The interesting question, which Meta’s 2025 pivot toward charging for its most capable Llama variants began to answer, is what happens when the two motivations produce different policies. At that moment, the pure democratizing argument becomes a residual, and the competitive logic becomes determinative.
Open weights versus open infrastructure. Releasing model weights is not the same as creating open infrastructure. Open weights give developers a copy of the model; they do not give developers access to the training data, the compute cluster, or the RLHF pipeline that produced it. The distinction matters because the most significant competitive advantage in AI is increasingly upstream of the weights: in the data flywheel that the social graph provides and in the compute infrastructure that Meta is building at scale. Commodity complementation at the weight level does not affect Meta’s advantage at the data and compute levels.