In a provocative manifesto published this Monday, Jesse Zhang, CEO of the AI infrastructure firm Decagon, challenged the prevailing narrative of the artificial intelligence sector. Titled "Everyone is wrong about open source AI in the enterprise," Zhang’s post touches on the central paradox currently defining the AI economy: while corporate AI deployments are increasingly migrating toward lightweight, cost-effective models, the overall expenditure on elite, "frontier" AI systems remains stubbornly high.
For months, industry analysts have debated whether the rise of powerful, accessible open-source models would cannibalize the market share of the "Big Three" labs—OpenAI, Anthropic, and Google. Yet, as Zhang points out, the reality is far more nuanced. We are not witnessing a zero-sum game of displacement; instead, we are observing the emergence of a two-tiered, symbiotic AI ecosystem.
The Life Cycle Theory: Discovery vs. Production
At the heart of Zhang’s theory is the concept of the "AI Life Cycle." He argues that frontier models and open-source models occupy distinct, non-competitive niches. In his view, frontier models—the high-cost, high-capability systems—act as the R&D engines of the AI world. They are the tools used to "prove out" complex, high-stakes use cases.
Once a specific task is sufficiently understood and the necessary workflows are optimized, companies are increasingly shifting those processes to leaner, more cost-efficient open-source alternatives. This process allows enterprises to maintain high performance without the prohibitive overhead of running massive models for routine, repetitive tasks.
"The frontier labs will keep owning discovery," Zhang writes. "Open source will increasingly own production."
This paradigm shift explains why the "death of the frontier" has been greatly exaggerated. While individual use cases may migrate to cheaper models, the rate at which new, complex use cases are being invented is keeping the demand for frontier-grade intelligence at an all-time high.
Chronology: The Evolution of the Model Market
The current landscape represents a massive departure from the state of the industry just eighteen months ago.
- Late 2024: The market was characterized by a "frontier-first" mentality. Almost all enterprise-grade applications relied on proprietary, closed-source models, as the performance gap between them and open alternatives was considered too wide for professional use.
- Early 2025: The "commoditization" narrative took hold. Observers began to wonder if foundation labs would suffer the same fate as infrastructure providers in other tech booms, essentially becoming low-margin utility providers—"selling coffee beans to Starbucks" while application developers captured all the value.
- Mid-2026 (Present): We are seeing the crystallization of the two-tiered market. Vertical AI startups have successfully optimized their tech stacks to favor lightweight models, yet the total spend on frontier models remains largely unchanged. The rise of efficient, high-performance models like DeepSeek and the anticipated impact of Nvidia’s Nemotron has cemented this split.
Supporting Data: The Disconnect Between Volume and Value
The evidence supporting Zhang’s theory is visible in the telemetry provided by AI gateways. According to Vercel’s AI gateway dashboard, the shift toward volume-based, cheaper models is undeniable. In just the past week, DeepSeek surged to command over one-third of total token volume processed through Vercel’s infrastructure. Similarly, Z.ai, the lab behind the GLM-5.2 model, has climbed into the top tier of volume usage.
However, a scroll down the same dashboard to the "overall token spend" metric reveals a starkly different story. Anthropic continues to account for more than half of the total AI expenditure on the platform. Despite the massive migration of tokens to cheaper providers, the "premium" dollars remain firmly entrenched with the frontier labs.
The data from OpenRouter further illustrates this disparity. DeepSeek V4 Flash currently processes 5.3 trillion tokens weekly, compared to just over 2 trillion for the leading frontier model, Opus 4.8. Yet, the cost per million tokens for Opus 4.8 is roughly 23 times higher than that of V4 Flash ($1.37 versus $0.06). This implies that while the usage has moved to the bottom of the pyramid, the revenue remains anchored at the top.
Implications for the Enterprise
For the Chief Information Officer (CIO) or the CTO, this bifurcation represents a new mandate: architectural agility. The ability to route tasks dynamically—sending complex, ambiguous requests to frontier models while delegating routine, high-volume tasks to specialized open-source models—is becoming the hallmark of a mature AI stack.
The "Difficulty Ceiling"
One reason frontier models have maintained their dominance is the sheer complexity of current enterprise demands. There is a "difficulty ceiling" for open-source models; while they have become incredibly adept at summarization, basic coding, and pattern recognition, the most nuanced enterprise tasks—those requiring high levels of reasoning, multi-step planning, or deep context integration—still require the "frontier" touch.
The Rise of Specialized Hardware
The introduction of Nvidia’s Nemotron model adds a new variable to this equation. By combining deep hardware integration with extreme adaptability, Nvidia is positioning itself to capture the middle ground. If a model can be both efficient enough for production and powerful enough for discovery, the current bifurcation may face further disruption.
Official Responses and Industry Outlook
While many industry leaders have remained guarded about the specific economics of their model usage, the trend toward "model-agnostic" application development is clear. Companies are no longer signing long-term, exclusive contracts with a single provider. Instead, they are building modular systems capable of switching model backends based on cost-per-token and performance benchmarks.
Critics of Zhang’s theory argue that it underestimates the pace of open-source improvement. If open-source models continue to improve at their current velocity, the "discovery" phase currently owned by frontier labs may shrink. If an open-source model can solve in one day what previously took a frontier model a week to refine, the necessity of paying for those premium tokens may evaporate.
However, for now, the economic data suggests that frontier labs have successfully protected their margins. By acting as the "source of truth" and the primary engine for high-end reasoning, they have ensured that even as the AI industry moves toward efficiency, the value of elite intelligence remains high.
Conclusion: A New Equilibrium
The narrative that foundation labs are destined to become mere commodity providers has failed to manifest in the way many predicted last year. Instead, we have arrived at a state of equilibrium.
The enterprise AI economy is currently defined by a persistent demand for both the "expensive, smart" and the "cheap, fast." As long as the frontiers of AI research continue to push the boundaries of what is possible, businesses will pay the premium for that capability. Once that capability is codified and the use case becomes a standard operational requirement, they will pivot to the cost-optimized, open-source alternatives.
This cycle—from frontier innovation to open-source commodity—is not a sign of failure for the major labs; it is the natural engine of the AI economy. We are not looking at a future where one model wins, but a future where the right model for the right task determines the success of the enterprise. As the dust settles on the current market, the only certainty is that the "two-tiered" model is here to stay, creating a stable, if complex, foundation for the next decade of digital transformation.
Ali Ikhwan
rifanmuazin
Neng Nana
