The Declining Monetizable Spread: A Critical Metric for AI Valuations

The ongoing debate over the capabilities of closed-source versus open-source artificial intelligence (AI) models has taken a new turn. While the performance gap between these models is narrowing, a more critical metric-the monetizable spread-has been largely overlooked. This metric represents the subset of the capability delta that enterprises are willing to pay a premium for. As this spread declines faster than the raw performance gap, it poses significant implications for AI valuations and investment strategies.

Why It Matters

The capability spread, or the performance difference between the best closed-source and open-source models, has been compressing over the past two years. According to Epoch AI, open-source models now trail state-of-the-art closed-source models by an average of three months, down from a year in late 2024. However, this compression is only part of the story.

The monetizable spread, which captures the performance aspects that enterprises are willing to pay for, is declining even more rapidly. This divergence suggests that the premium paid for closed-source models may be overvalued if it does not account for the diminishing returns in practical applications. For investors and analysts, understanding this metric is crucial for accurate valuations and strategic decision-making.

Key Risks

1. Overvaluation of Closed-Source AI Companies: If the market continues to price closed-source AI companies based on their raw performance metrics, these valuations may be unsustainable. The declining monetizable spread indicates that the premium paid for proprietary models may not justify their current equity prices.

2. Regulatory and Safety Concerns: While open-source models are catching up in terms of performance, they often lag in regulatory compliance and safety certifications. Enterprises require robust guarantees on data security and ethical use, which closed-source providers can more easily offer. This remains a significant barrier for widespread adoption of open-source AI in regulated industries.

3. Research Talent and Distribution: Closed-source companies like OpenAI and Anthropic have a stronger talent pool and distribution network, which are critical for maintaining their competitive edge. However, as the monetizable spread narrows, these advantages may not be enough to sustain high valuations.

The Opportunity

1. Investment in Open-Source Ecosystems: As the performance gap closes, there is an increasing opportunity for investors to support open-source AI initiatives. These projects can benefit from a broader community of contributors and a more diverse range of applications, potentially leading to innovative breakthroughs.

2. Enterprise Partnerships: Closed-source companies can leverage their existing strengths in enterprise agreements, regulatory compliance, and safety certifications to form strategic partnerships. By focusing on these areas, they can maintain their market position even as the performance gap narrows.

3. Hybrid Models: A potential middle ground is the development of hybrid models that combine the transparency and flexibility of open-source with the security and reliability of closed-source solutions. This approach could address the concerns of both developers and enterprises, creating a more balanced and sustainable AI ecosystem.

A Brief History of Compression

The capability gap between open and closed models has followed a trajectory that should concern investors in frontier labs. At the end of 2023, the best closed model scored roughly 88% on MMLU-the standard knowledge benchmark-while the best open model managed about 70.5%. By early 2026, this gap is effectively zero on knowledge benchmarks and single digits on most reasoning tasks.

The time dimension is even more telling. Epoch AI found that open weight models now trail the state-of-the-art by roughly three months on average. In late 2024, the Epoch team measured this lag at closer to a year. DeepSeek demonstrated the mechanism behind this rapid improvement; its V3 base model used 2.6 million GPU hours versus Llama 3 405B’s 30.8 million, an order of magnitude improvement in training efficiency.

Conclusion

The declining monetizable spread is a critical metric that investors and analysts must consider when evaluating AI valuations. While the performance gap between open-source and closed-source models is narrowing, it is the practical value that enterprises are willing to pay for that truly matters. As this trend continues, it will reshape the landscape of AI investments and business strategies.