A fascinating paradox is unfolding in the artificial intelligence landscape, as revealed by Decagon CEO Jesse Zhang’s recent theory: “Everyone is wrong about open source AI in the enterprise.” He observes that while more mature AI deployments are indeed shifting towards lighter, more efficient models, the overall spending on expensive, state-of-the-art models remains stubbornly high.
Zhang proposes a new way to understand the relationship between cutting-edge “frontier” models and their open-source counterparts. He argues they aren’t competitors locked in a zero-sum battle; instead, they represent two distinct phases within the same AI lifecycle. Frontier models serve to prove out complex use cases, which then transition to more cost-effective open-source alternatives as they mature.
This dynamic creates a curious equilibrium. As existing, mature applications migrate to lighter, cheaper models, a constant stream of new, innovative use cases emerges. These fresh applications initially rely on the most advanced, often proprietary, frontier models, ensuring that overall investment in them remains robust.
The Data Doesn’t Lie: Usage vs. Spend
While Zhang’s theory is compelling, concrete data supports his observations. The Vercel AI gateway dashboard, for instance, shows significant shifts in token volumes, with open-source models rapidly gaining traction. In a single week, DeepSeek surged into the lead, now processing over a third of the tokens flowing through Vercel’s infrastructure, while Z.ai’s popular GLM-5.2 model also secured a respectable fourth place.
However, when you shift your focus to overall AI spend on the Vercel platform, the picture changes dramatically. Anthropic still accounts for more than half of the total expenditure. Even with slight dips over the past month due to Anthropic’s own rising prices, its financial dominance remains largely undiminished.
OpenRouter, a platform catering to a broader, slightly less enterprise-focused market, tells a similar story. Here, Deepseek V4Flash is the undeniable leader in raw usage, processing an astonishing 5.3 trillion tokens weekly. In contrast, Opus 4.8, a prominent frontier model, handles just over 2 trillion tokens during the same period.
Despite lower usage, Opus 4.8’s financial impact is substantial. OpenRouter data reveals its average token cost is roughly 23 times higher than V4Flash ($1.37 per million tokens compared to just 6 cents). This dramatic price difference strongly suggests that frontier models like Opus 4.8 continue to capture the lion’s share of overall spending, even with lower token volumes.
It’s also worth noting the recent arrival of Nvidia’s Nemotron, poised to become another significant player. Given Nvidia’s deep industry connections and the model’s exceptional adaptability, Nemotron could quickly leap to the forefront, further influencing this evolving market dynamic.
Understanding the AI Lifecycle: Discovery to Production
These figures, while not providing a complete proof of Zhang’s AI lifecycle theory, certainly indicate that frontier labs like Anthropic are not yet suffering from the ascent of open-source solutions. One key explanation is the sheer, rapid expansion of the AI-addressable task market. Top models maintain their position by dominating the crucial early-stage deployments where new use cases are forged.
As Zhang aptly puts it, “The frontier labs will keep owning discovery. Open source will increasingly own production.” This suggests a natural division of labor, where innovation and cutting-edge performance command a premium in the initial phases, while efficiency and cost-effectiveness take over as solutions mature.
Another contributing factor is the inherent difficulty of many AI use cases. Even as clients explore open-source alternatives, numerous complex challenges simply cannot be fully addressed by cheaper models, necessitating the continued use of high-performance frontier solutions. This maintains a steady demand for their superior capabilities.
Consequently, this “two-tiered economy of models,” with frontier models handling complex discovery and open-source models excelling in efficient production, appears to be settling in as a relatively stable feature of the broader AI economy. It’s a symbiotic relationship, not a purely competitive one, fostering growth across the entire ecosystem.
The Enduring Value of Frontier AI
Not long ago, there was speculation that foundational AI labs might become mere commodity suppliers, like coffee bean growers selling to a brand like Starbucks, with the application layer reaping most of the benefits. While some aspects of this prediction materialized—for instance, vertical AI applications did indeed switch to lighter models, and the economics of “GPT wrapper” startups remained stable—the full picture is more nuanced.
What we’re witnessing instead is the remarkable ability of frontier providers to hold onto the most desirable segment of the marketplace. They consistently command the premium token price for their advanced capabilities. This signifies the inherent value placed on groundbreaking performance and reliable, sophisticated outputs.
This trend of frontier models retaining their premium positioning doesn’t show any signs of abating soon. As the demand for innovative AI solutions continues to accelerate, the unique strengths of these advanced models ensure they will continue to be a significant driver of spending in the AI sector.
Source: TechCrunch – AI
