Open Models vs Closed Models: The $25B Question Explained by Frank Nagle, Linux Foundation

### Overview In this interview, Frank Nagle, Chief Economist at the Linux Foundation, discusses a working paper titled *The Latent Role of Open Models in the AI Economy*. The paper examines enterprise spending patterns across proprietary and open artificial intelligence models, revealing that organizations are leaving an estimated $24.8 billion in unrealized economic value on the table by underutilizing open models. Despite open models closing performance gaps quickly and offering significant cost advantages, closed proprietary models continue to capture the vast majority of enterprise usage.

### Key Topics - **Empirical Market Data via OpenRouter:** The study analyzes data from OpenRouter—a gateway for LLM inference representing roughly 1% of the overall LLM API economy—to observe model adoption, prices, and usage patterns at a granular level. - **Pricing and Performance Disparity:** Closed models account for roughly 80% of observed usage despite being on average six times more expensive than open alternatives. While closed models often lead benchmark performance, open models typically close performance gaps within three to four weeks. - **Adoption Drivers vs. Economic Rationality:** Price and raw performance benchmarks alone do not dictate enterprise decisions. Adoption patterns are heavily influenced by switching costs, vendor lock-in, information frictions, liability concerns, and customer support ecosystems. - **Open Weights vs. True Open Source:** Nagle highlights structural differences between traditional open source software (like the Linux kernel) and open-weight models (such as Meta's Llama), where models are developed internally by a single entity and weights are released without necessarily providing training data or community-driven code governance.

### Technical Details & Practical Considerations - **Misconceptions Around Data Privacy:** A common corporate fallacy is the belief that utilizing open models compromises data privacy. Nagle clarifies that self-hosting open models behind corporate firewalls ensures data never leaves internal environments, whereas querying proprietary cloud APIs inherently transmits corporate data off-premises. - **Support and Liability Infrastructure:** Proprietary vendors (such as OpenAI, Anthropic, or Google) provide contractual liability, enterprise support, and risk offloading. In contrast, open models currently offer limited commercial enterprise support when unexpected issues arise. - **Fine-Tuning and Specialization:** Open models enable deep customization and fine-tuning that proprietary APIs cannot match. Examples include fine-tuning models on Hugging Face for internal enterprise workflows or adapting models for underrepresented African languages. - **Evaluation Checklist:** CIOs, CISOs, and CTOs are pointed to Table 2 of the paper, which lists 8–9 distinct dimensions—differentiating valid operational drivers (such as switching costs and technical support) from invalid justifications (such as privacy misconceptions)—to assess whether workloads should transition to open models.

### Takeaways - Enterprises paying premium rates for closed models could achieve equivalent or superior performance on non-cutting-edge tasks by shifting workloads to open models, capturing tens of billions in aggregate savings. - Over the next five years, the AI market is projected to bifurcate similarly to operating systems: open models will commoditize broad server-side enterprise and customized applications, while closed models will specialize in niche verticals requiring turnkey enterprise support and liability guarantees. - Organizations are advised to pursue a three-step path: educate leadership on true open vs. closed distinctions, evaluate use-case requirements, and begin incremental tinkering and experimentation rather than treating single proprietary tools as default solutions.