Engineers are increasingly scrutinizing their monthly API bills, prompting a migration from high-performance frontier models to lightweight, cost-effective alternatives. Yet, a look at the data from AI gateway dashboards, such as those provided by Vercel, reveals a striking disconnect between raw usage and actual financial expenditure. While developers are routing massive amounts of traffic toward budget-friendly models, the revenue landscape remains firmly anchored to the industry's most expensive offerings.

The Token Volume Paradox

Recent data from Vercel’s AI gateway dashboard illustrates this shift in infrastructure load. Over the past week, DeepSeek has processed more than one-third of all tokens passing through the platform, securing the top spot in total volume. However, this surge in throughput does not translate into a proportional share of the platform's total AI spending. Despite the high volume of DeepSeek tokens, Anthropic continues to capture more than half of the total AI expenditure on the platform.

This trend is further corroborated by data from OpenRouter, which tracks usage across various models. DeepSeek V4 Flash currently leads the market with 5.3 trillion tokens processed per week. In contrast, a leading frontier model like Opus 4.8 processes just over 2 trillion tokens in the same timeframe. The financial disparity is stark: the average cost per million tokens for Opus 4.8 is $1.37, while V4 Flash costs a mere 6 cents. Consequently, the per-token price for Opus 4.8 is approximately 23 times higher than that of the V4 Flash model.

The Two-Stage AI Lifecycle

This divergence suggests that frontier models and open-source alternatives are not locked in a zero-sum game, but rather function as distinct stages in a service lifecycle. Enterprises are increasingly bifurcating their AI budgets into two specific phases: discovery and production. In the discovery phase, teams utilize high-cost, high-performance frontier models to validate business value and prove complex use cases. Once a use case is mature and the technical requirements are well-defined, companies migrate to lower-cost, open-source, or lightweight alternatives for the production phase.

This structure allows frontier labs to maintain their market dominance. By providing the superior reasoning capabilities required to solve the most difficult technical challenges, these labs capture the most valuable segment of the market. Even as vertical AI services transition to lighter models, the premium pricing of frontier models remains resilient because the value provided by their advanced reasoning capabilities effectively offsets price resistance. The frontier models open the path, while the low-cost models handle the high-volume execution.

The Future of AI Infrastructure

As the market for AI-solvable tasks continues to expand, frontier labs are positioned to retain their revenue share by dominating the initial deployment phase of every new, complex task. While Nvidia’s Nemotron is poised to enter the top tier of this market, leveraging its robust network connectivity and model adaptability, the fundamental economic structure remains unchanged. High-cost models will continue to serve as the engine for innovation, while low-cost models provide the scale for operational efficiency.

Companies must now treat AI spending as a two-tiered budget, separating the costs of discovery from the costs of production. This strategic division is becoming the new standard for achieving real-world cost optimization in AI operations.