Open-Source AI Models: Analyzing the Widening Gap Between Usage and Revenue

    Open-weight models are capturing an increasing share of enterprise AI token usage, while their share of spending often declines. This article explains why that gap is emerging, how enterprise workloads should be segmented, and why cloud providers and AI platforms must move beyond token-based API pricing toward task completion, workflow orchestration, observability, and governance. It also outlines practical strategies for routing workloads, controlling costs, and turning large-scale open-source model usage into sustainable business value.

    Enterprise large language model applications have moved from pilot projects into large-scale production use. One increasingly clear pattern is that open-weight models account for a growing share of token consumption, while their share of total spending often declines. Usage rises as unit prices fall, creating a widening “scissors gap.”

    Understanding this structure is more important than simply tracking model usage rankings. It determines whether providers can turn market share into profit, and it also shapes how cloud-based inference, orchestration, and governance capabilities should be designed.

    Why Usage and Spending Are Diverging

    After open-weight models enter enterprise workflows, the typical path is for engineering teams to begin with open models for evaluation and fine-tuning. High-concurrency and highly parallel workloads are then moved to cost-sensitive production paths.

    Web scraping, batch extraction, log aggregation, and tool calls or intermediate reasoning steps in multi-step AI Agents are extremely sensitive to latency and unit price. They do not always require state-of-the-art reasoning. As a result, invocation volume can grow exponentially within weeks, while the price customers are willing to pay per request is pushed close to the bare cost of inference.

    At the same time, open weights mean that the same parameters can be hosted by multiple inference providers. Basic completion APIs are becoming increasingly interchangeable. Competition is shifting from “who has the model” to “how many tokens can be generated per unit of compute, how stable is the service, and how predictable is queueing.”

    Prices naturally converge toward the cost of electricity, hardware depreciation, and scheduling efficiency. The growing share of open models in token consumption, combined with their declining share of billing, is a direct result of this workload segmentation and increasing commoditization.

    High-value workloads—such as complex reasoning, long-chain planning, and tightly constrained code generation—remain concentrated in a limited number of closed or strongly aligned services. These workloads have higher prices but lower volumes.

    Lower-value, high-throughput tasks are increasingly handled by open models and self-hosted inference. Total enterprise AI spending continues to expand globally, but the “volume” and “money” associated with open models have already become decoupled. For providers, selling only token-priced APIs will increasingly resemble selling standardized compute, with gross margins constrained by hardware and electricity costs.

    Workload Segmentation Determines Pricing Power

    From an engineering perspective, enterprise workloads can be broadly divided into three layers.

    1. Throughput-Oriented Workloads

    These workloads typically have relatively short contexts, support batching and speculative decoding, and are not highly sensitive to sampling temperature. They are well suited to continuous batching, prefix caching, quantization, and speculative sampling.

    For this category, the maximum acceptable price is often close to the marginal inference cost, plus only a thin premium.

    2. Orchestration-Oriented Workloads

    These include multi-step Agents, tool calls, retrieval-augmented generation, and workflow state machines. They may consume a large number of tokens, but the value users actually perceive lies in task completion rates, observability, failure recovery, and permission boundaries.

    If these workloads are still priced as raw token consumption, the value of the overall workflow is diluted by its intermediate steps.

    3. Decision-Oriented Workloads

    These include compliance reviews, critical code merges, and financial or risk-control decisions. Their usage is limited, but requirements for correctness, auditability, and data residency are extremely high.

    This is where providers have genuine pricing power. These services are often packaged by seat, project, or SLA rather than priced purely by token count.

    Open-weight models from China frequently rank highly in aggregated usage among global developers. Much of that usage comes from the intermediate steps of throughput-oriented and orchestration-oriented workloads. If a product remains limited to inexpensive inference compatible with a standard completion API, the scissors gap will continue to widen.

    To turn usage share into profit, providers must move measurement upward—from tokens to completed task units and governed workflows.

    Infrastructure After Open Weights Become Widely Available

    Open weights reduce switching costs. Enterprises can build their own inference clusters or deploy the same models across multiple clouds.

    Once basic capabilities become commoditized, differentiation can come from only three areas:

    • Unit economics
    • End-to-end latency distribution
    • Depth of integration with enterprise systems

    For unit economics, continuous batching, KV-cache hit rates, and the trade-off between quantization precision and output quality should be exposed as configurable strategies rather than reduced to the lowest possible bit width by default.

    For latency, providers need to distinguish between time to first token and time to the final token. Interactive Agents require reliable tail latency, while batch workloads prioritize throughput.

    For integration, identity, key management, audit logs, content safety, and enterprise knowledge bases should be connected through a unified control plane. Replacing a model weight file should not require replacing the entire production system.

    On the cloud side, object storage can hold model weights and datasets, while containers and elastic compute provide inference. Message queues and workflow engines manage Agent steps, and logging and monitoring systems maintain both token-level and task-level accounting.

    The value of a general-purpose cloud platform such as Baidu AI Cloud lies in turning this control plane into a reusable platform, rather than selling another interchangeable completion API.

    From Selling APIs to Selling Completion

    The next stage of product development should focus on three directions: moving measurement upward, moving risk management downward, and combining capabilities into an integrated platform.

    Move Measurement Upward

    Token-based billing can remain the foundation for throughput workloads. Orchestration workloads can be priced by successful sessions, successful tool calls, or successfully processed documents. Decision-oriented workloads can be priced by seat and compliance package.

    A single workflow should support mixed billing models. This prevents intermediate reasoning tokens from inflating customer bills, while also protecting provider revenue from token price competition.

    Move Risk Management Downward

    The security and copyright responsibilities associated with open-weight models are more distributed. Enterprises need model cards, data-composition disclosures, red-team findings, and runtime guardrails.

    Making guardrails, auditing, and key management default platform capabilities is more likely to earn a place on enterprise procurement shortlists than competing solely on price.

    Combine Capabilities Across the Lifecycle

    Retrieval, memory, tool sandboxes, and evaluation regression should be released and managed alongside inference as part of the same lifecycle.

    Customers need repeatable business metrics—such as extraction F1 scores and ticket resolution rates—not a weekly ranking of token volume. Evaluation datasets, gradual rollouts, and rollback procedures should be integrated into a deployment pipeline so that open weights can move from experimentation to production defaults.

    Several engineering principles can be applied immediately:

    • Design stable system prompts and tool descriptions for prefix caching to reduce unnecessary KV-cache invalidation.
    • Use smaller, faster specialized models for intermediate Agent steps, and upgrade to stronger models only at critical decision points.
    • Use asynchronous queues to absorb traffic spikes and warm pools to reduce time-to-first-token variance.
    • Report both token-level and task-level billing so finance teams can understand which workloads are creating the scissors gap.

    How Enterprises Should Use the Scissors Gap

    For procurement and architecture teams, the scissors gap does not simply mean that open models are cheaper. It is a reminder to separate workloads.

    Batchable tasks should not use expensive interactive APIs. Cacheable computations should not be repeated. Tasks that can be completed by a smaller model should not default to the largest available model.

    Organizations should establish internal routing policies that dynamically select model weights and quantization levels based on task type, latency SLOs, and data sensitivity.

    At the same time, enterprises must guard against “usage growth without business growth.” Each Agent workflow should have defined business KPIs linked to token costs. If costs continue to grow faster than business metrics, the workload may still be stuck in the throughput layer. The organization may need to productize orchestration and decision capabilities or tighten unrestricted generation.

    From a data and compliance perspective, open weights make private deployment easier, but they do not automatically guarantee that data remains within the required jurisdiction or security boundary. An inference gateway, auditing, and data redaction should be part of the default architecture rather than patches added after launch.

    Conclusion

    Open models are now deeply embedded in enterprise processes, and their scale of adoption demonstrates their engineering viability. The more important question is how to turn increasingly commoditized tokens into measurable task completion and auditable decision quality while inference prices continue to approach marginal cost.

    Providers that build reliable workload routing, mixed billing, and governance control planes will be better positioned to protect margins as usage grows. Otherwise, they will simply sell APIs at increasingly lower prices.

    Cloud platforms such as Baidu AI Cloud can better address this stage of the market by offering elastic inference, storage, queues, and observability through a unified control plane. The need is no longer just another model name. It is a production system that turns large-scale model usage into an operable and economically sustainable business.