The Shift from Closed to Open Weights AI Models
Matthew Bermango watch the original →
the gist
While frontier closed-source models like those from OpenAI and Anthropic capture the majority of revenue, open-weights models are rapidly gaining token volume share, offering enterprises better data control, lower costs, and reduced platform risk.
The Divergence of Token Volume and Revenue
Data from Vercel indicates a significant trend: while closed-weights models (OpenAI, Anthropic) maintain a dominant hold on revenue, open-weights models are capturing an increasing percentage of total token volume. This market split is defined by two distinct tiers of usage. The absolute frontier models, such as those from OpenAI and Anthropic, command high premiums because they provide the incremental intelligence necessary for high-stakes tasks like high-frequency trading, where a slight performance advantage justifies the cost. Conversely, open-weights models are becoming the commodity standard for general-purpose tasks, offering performance that is often sufficient for most enterprise needs at a fraction of the cost.
The Strategic Value of Open Weights
For enterprise companies, the primary advantage of open-weights models is the ability to build long-term equity rather than renting intelligence from third-party labs. By hosting and fine-tuning models internally, companies gain three critical benefits: ownership of the data, reduced platform risk, and bargaining power. When businesses rely on closed-source APIs, they risk exposing sensitive operational data to the model provider, which could eventually be used to train competing products. Furthermore, the ability to switch between dozens of inference providers prevents vendor lock-in, forcing frontier labs to compete on price and performance.
Geopolitical Risks and Infrastructure
Despite the benefits of open-weights models, a significant portion of the current high-performing open-source ecosystem originates from China, including models like Qwen and DeepSeek. This creates a potential long-term geopolitical risk for the United States. As these models become increasingly co-designed with domestic Chinese hardware, US enterprises building their infrastructure on top of these models may inadvertently become dependent on Chinese chip ecosystems. While open-source adoption currently drives demand for Nvidia hardware, the lack of a robust US-led open-source strategy leaves a gap in the domestic AI supply chain.