My Bets on Where Open LLMs Go Next
An analysis of the current unstable equilibrium between open and closed AI models, arguing that closed models will likely pull ahead due to economic and data feedback advantages. The long-term, stable future for open models lies in a specialized ecosystem of cheaper, faster models, potentially funded by new structures like consortiums.
The current state of AI models presents a surprising and unstable equilibrium: the best open models remain remarkably close in capability to their closed-source counterparts. However, this parity is unlikely to last. The long-term advantage will likely shift to closed model labs due to fundamental economic and strategic factors that favor their approach.
The Widening Gap: Why Closed Models Will Pull Ahead
The primary thesis is that while we are in a period of close competition, the underlying dynamics favor closed models. This is rooted in several key advantages:
- Superior Benchmarks: Closed model labs in the U.S. possess a wide array of private benchmarks that have a higher signal for real-world usage and knowledge work. Public benchmarks, which show open models in a very favorable light, are not diverse or representative enough to capture the full range of capabilities where closed models are more robust.
- The Distribution and Data Flywheel: The most significant long-term advantage for closed models is their vast user base. This distribution network allows for a powerful reinforcement learning (RL) feedback loop, where models are trained and improved in real-time based on user data. This creates a compounding advantage that open models, without equivalent distribution, cannot replicate.
- Economic Moat: To justify investing orders of magnitude more in training compute, labs need to demonstrate vast, real-world usage and a clear path to monetization. The direct user access and API-based business models of closed systems provide a much stronger economic foundation for sustained, large-scale investment.
Deconstructing the Rise of Open Models
The impressive performance of open models, particularly from Chinese labs, is often attributed to distillation—training on the outputs of more powerful, closed models. While distillation is a useful training tool, it is not a panacea. These labs would still be highly competitive even if distillation were somehow regulated away. The narrative that China will inevitably catch up, often fueled by simple benchmark plots, misses this nuance and the deeper strategic moats being built by closed labs.
The True, Sustainable Future for Open Models
The path forward for the open model ecosystem is not to compete head-to-head with frontier closed models on the same terms. Instead, open models must find a different, more sustainable equilibrium. This involves moving away from a monolithic goal of matching closed models and toward building a diverse ecosystem.
The future of open models is to be specific, cheap, fast, and ubiquitous.
- Specialization: Rather than general-purpose “power tools” for knowledge work (the domain of closed models), open models will excel at specialized, automated tasks. Examples include small, private models for navigating a local file system or highly efficient models for command-line interface (CLI) operations. These models can perform repetitive tasks at a fraction of the cost (e.g., 100x cheaper).
- Intelligence per Watt: The focus will shift to maximizing efficiency and minimizing the cost of intelligence for background tasks that permeate our daily lives.
Ecosystem Dynamics and Governance
Measuring the ecosystem reveals that adoption is a “second derivative” of capabilities—it lags behind performance gains. While Chinese models have seen an inflection point in adoption, new, high-quality American models like Google’s Gemma 4 and Nvidia’s Nemotron series are just beginning to change user habits, a process that can take over six months.
The success of these models depends on factors beyond raw performance, including robust tooling, ease of fine-tuning, and permissive licensing. Nvidia’s motivation is clear: building open models helps sell more chips. Meanwhile, startups like Arcee AI are building businesses around helping enterprises deploy and manage open models on-premises.
Looking ahead, as AI becomes more powerful and integral to society, the concentration of power within a few closed labs will become strategically untenable for governments and other sovereign entities. This will create a powerful incentive to support open models, which offer a more decentralized and inspectable paradigm. This leads to the inevitable need for new funding structures, such as an open model consortium. Such a coalition would allow multiple companies to pool the billions of dollars required to create frontier open models, giving them security and influence over the technology’s development without any single entity bearing the full capital burden.