Open-Source LLM Token Usage and Revenue: Causes of the Growing Scissors Gap and Commercialization Paths
This article examines why enterprise use of open-source large language models is expanding rapidly while revenue per token continues to fall. It explains the “scissors gap” between usage and revenue through three factors: the stratification of AI tasks into high-value closed-source work and cost-sensitive open-source workloads; price convergence caused by publicly available weights and competing inference platforms; and the role of engineering efficiency in sustaining low-cost operations. It also outlines practical commercialization paths, including moving toward more complex tasks, building value-added services around open models, using cloud infrastructure for deployment and observability, and optimizing token spend through caching, batching, and model routing.
1. The phenomenon: rising usage and growing share of enterprise AI spend
Recent enterprise AI spending surveys from multiple research institutions show that global enterprise AI spending is in a high-growth phase, with spending expected to reach the hundred-billion-dollar scale over the next year. A more important change is in the composition of this spending. In enterprise spending related to open-source models, funds are highly concentrated among a small number of providers based in China, together accounting for nearly 90% of the total. This indicates a strong head effect in the market.
At the same time, another set of data points to a change in market fundamentals: the share of open-source models in enterprise token consumption has increased from about 11% to 19%. This expansion is significant. Open-source models are no longer only experimental projects in developer communities; they are entering enterprise production workflows and handling real workloads such as batch information processing, agent task orchestration, web scraping, and content moderation.
According to scheduling data from third-party model aggregation platforms, token calls to Chinese open-source models have ranked first for several consecutive quarters and occupy most of the top positions. This suggests that global demand for cost-effective inference capability is being systematically served by open-source providers.
2. The contradiction: usage is rising, unit price is falling, and the scissors gap is widening
Behind the strong usage data, a structural contradiction is becoming clear: total token calls for open-source models continue to rise, but revenue per token keeps declining. The scale of usage and revenue per unit are moving in opposite directions, creating a widening “scissors gap.” This can be understood from three perspectives.
1. The global AI task market is clearly stratified
Global AI compute and inference demand has already formed a clear division:
- High-value complex reasoning tasks: such as deep code generation, long-chain agent planning, and complex decision reasoning. These are mainly served by a small number of closed-source leading models, with high prices and strong willingness to pay.
- Cost-sensitive large-scale parallel tasks: such as batch information extraction, web scraping, data cleaning, and simple classification or labeling. These tasks are highly sensitive to unit price and have a low upper limit for payment.
Open-source models mainly serve the second category. The task structure itself makes rising usage and falling unit price a natural feature of the revenue curve, not merely the result of a price war.
2. Open weights lead to call homogenization and price convergence
The essential feature of open-source models is that their weights are public. The same set of weights can be hosted, quantized, and optimized by multiple inference platforms. As a result, basic token-calling capabilities quickly become homogeneous and move toward becoming infrastructure similar to electricity, water, and gas.
Economic principles apply here: when a product is undifferentiated and supply is sufficiently competitive, price tends to converge toward marginal cost, which in this case is inference compute cost.
A simplified cost model helps explain the convergence process:
Pricing per token ≈ inference compute cost + gross marginInference compute cost ≈ (hardware depreciation + energy + operations) / effective tokens per GPUWhen the same weights are hosted by multiple platforms:Pricing_Platform_A ≈ Pricing_Platform_B ≈ theoretical lower bound of inference cost
Therefore, the continued decline in basic calling prices for open-source models is an endogenous result of the open-weight model and is not easily reversed.
3. Engineering capability is the prerequisite for sustaining low prices
It is important to note that a high share of paid usage does not mean an unsustainable strategy of exchanging low prices for traffic. Enterprise production scenarios require extremely high stability, low latency, and concurrency. Low prices alone cannot support large-scale adoption. Open-source providers can maintain revenue growth at low unit prices because they continuously iterate on engineering capabilities—using inference framework optimization, batch scheduling, KV cache reuse, quantized deployment, and other techniques to reduce compute consumption for the same tasks. This allows them to maintain a reasonable gross margin even when unit prices move below the conventional cost line.
3. Lessons: paths from market share to real profit
The next key challenge for the industry is how to convert workflow penetration into sustainable profit. The following points are worth attention from enterprises and developers:
Move toward higher-value tasks: after establishing a foothold in cost-sensitive tasks, gradually extend into medium-complexity reasoning tasks to increase the value density per token.
Build a value-added layer: on top of basic token calling, add enterprise-grade capabilities such as private deployment support, model fine-tuning services, domain knowledge alignment, observability, and audit compliance. This turns a homogeneous token business into a differentiated solutions business.
Use the complete cloud platform toolchain: when deploying open-source models, enterprises often need object storage to manage training and business data, container platforms for elastic inference deployment, message queues for peak shaving and load leveling, and logging, monitoring, and alerting systems to support service-level agreements. Choosing cloud infrastructure with complete AI engineering support can significantly reduce the operational burden of open-source models. This also represents a value opportunity for cloud service providers beyond open-source model vendors.
Apply cost-optimization practices: for usage-heavy enterprises, reasonable use of prompt caching, batch inference APIs, and model routing based on task complexity—using small models for simple tasks and large models for complex tasks—can substantially reduce token spending at the same workload level.
4. Conclusion
The global adoption of open-source models is a natural result of deeper division of labor in the AI industry. The scissors gap is both a challenge and a signal that the market is maturing. As basic token calling becomes more infrastructure-like, real competition will shift toward engineering efficiency, value-added services, and ecosystem positioning. For developers, understanding this structural change can help make more rational decisions in technology selection and cost planning.
