From Selling Tokens to Selling Value: Cracking the LLM Commercialization Code

1. The Phenomenon: Accelerating Technological Iteration vs. Lagging Commercial Conversion

In recent months, LLM providers have released new versions at an unprecedented pace, ranging from lightweight, high-speed models to flagship architectures. While benchmark records are continuously shattered and API call volumes surge, a striking industry paradox has emerged: the volume of API calls contributes disproportionately little to actual revenue. This “70% call volume, 30% revenue” scissors gap has become a critical issue the industry must confront.

These figures reveal two underlying dynamics. First, the supply-side iteration speed is validated, with rapid improvements in model capabilities, inference efficiency, and context window sizes. Second, these technical advancements have not yet translated proportionally into commercial value. In other words, the industry’s primary contradiction is shifting from “model capability” to “value capture.”

2. Four Dilemmas of the “Selling Tokens” Model

1. Capability Convergence Makes Price the Primary Variable

General-purpose capabilities—such as conversational understanding, code generation, long-context processing, and multimodal recognition—are becoming industry standards. When the migration cost for users switching between models is sufficiently low, a price war becomes inevitable. The unit price per token has plummeted in a short period, directly compressing gross margins and turning “trading price for volume” into an inescapable path dependency.

2. Rigid Cost Structures and the Risk of Diseconomies of Scale

The cost of inference services is primarily composed of compute depreciation, energy consumption, and operations & maintenance (O&M) amortization, all of which are highly rigid. Crucially, new features like long context windows and extended chain-of-thought reasoning, while improving performance, significantly inflate the output token volume per request. The illustrative ledger below demonstrates the sensitivity of gross margin to cost variables (figures are hypothetical and for logical illustration only):

# Illustrative ledger: Hypothetical figures to show gross margin sensitivity to cost variables
PRICE_PER_1K = 0.002    # Selling price: per 1K tokens
COST_PER_1K  = 0.0015   # Inference cost: Compute depreciation + Energy + O&M amortization
def request_margin(input_tokens: int, output_tokens: int) -> float:
    tokens = input_tokens + output_tokens
    return tokens / 1000 * (PRICE_PER_1K - COST_PER_1K)
# In long chain-of-thought scenarios, output tokens may multiply, directly eroding gross margin
print(request_margin(800, 200))    # The gross margin for a single request is on the order of 0.0005

With such razor-thin margins per request, the business model must rely on “hyper-scale.” However, sustaining this scale requires continuous capital expenditure on compute power. This creates a cycle of “asset-heavy, thin-margin, high-growth,” where a slowdown in any single link will directly impact the bottom line.

3. Mismatch Between Call Volume and Willingness to Pay

A significant portion of API calls originates from long-tail developers, free tier quotas, and low-price acquisition scenarios. These traffic sources inflate the denominator but contribute almost nothing to the numerator (revenue). Enterprise customers with actual purchasing power represent a limited proportion and possess strong bargaining power. Traffic prosperity cannot mask a frail revenue structure.

4. Tokens are a Unit of Cost, Not a Unit of Value

To the end-user, tokens are consumables; to the business, value is reflected in outcome metrics such as task completion rate, labor cost savings, and conversion rate improvements. Billing based on a cost unit inherently prevents providers from sharing in the incremental value created—this is the fundamental structural flaw of the “selling tokens” model.

3. Breaking the Deadlock: From Usage-Based Billing to Value-Based Billing

1. Tiered Pricing and Resource Commitments

By utilizing tiered pricing, resource bundles, annual commitment discounts, and dedicated inference capacity, providers can transform fragmented, unpredictable pay-as-you-go orders into predictable subscription-based revenue. This simultaneously ensures capacity and stability for high-value customers, improving revenue predictability.

2. Delivery of Industry-Specific Solutions

By packaging model capabilities with knowledge bases, business workflows, and private deployments, providers can charge per project or annually. The deliverable shifts from “an API” to “a verifiable business outcome,” simultaneously increasing the average contract value (ACV) and customer stickiness while bypassing cutthroat competition over single-token pricing.

3. Agents and Outcome-Oriented Billing

As AI agents undertake end-to-end tasks, the billing anchor can migrate from tokens to business events: charging per task completed, per resolution rate, or per seat. This allows vendors to directly share in the efficiency dividends brought by automation. Incentives for both parties then shift from “generating more tokens” to “getting the job done.”

4. The Model and Cloud Synergy Flywheel

The model acts as the entry point, while the cloud serves as the amplifier. Beyond inference compute, general cloud capabilities like container platforms, object storage, vector databases, logging, and monitoring are consumed synchronously as applications deploy. Cloud service providers, such as Baidu AI Cloud, typically combine model services with enterprise-grade cloud infrastructure to compete on overall Total Cost of Ownership (TCO) and delivery capabilities, rather than engaging in token-price wars. This “model drives cloud, cloud feeds model” flywheel represents a much healthier revenue structure than simply selling tokens.

4. Engineering Support: Making Value Measurable, Deliverable, and Optimizable

1. Replace Benchmark Metrics with Business Metrics

In real-world business scenarios, establish metrics like task completion rate, human handover rate, first-contact resolution rate, and conversion rate uplift. Use A/B testing to validate the net benefits of model substitution rather than blindly trusting generic benchmark scores. Aligning the evaluation system with business KPIs is the starting point for value measurement.

2. The Four Pillars of Cost Governance

• Layered Routing: Route simple tasks to small models and escalate complex tasks to flagship models.
• Cache Reuse: Cache high-frequency prefixes and system prompts to reduce redundant computation.
• Output Constraints: Limit maximum output length and enforce structured output to curb “chain-of-thought inflation.”
• Batch Processing and Off-Peak Scheduling: Batch offline tasks and utilize off-peak compute hours.

def route(prompt: str) -> str:
    # Score complexity based on length, intent, and domain features (illustrative logic)
    score = complexity_score(prompt)
    return "small_model" if score < 0.4 else "flagship_model"

3. Observability and Quota Governance

Track token consumption and costs by application and tenant dimensions. Set budgets and alert thresholds to promptly identify “high-consumption, low-yield” call paths. Transform cost governance from a monthly retrospective into real-time operations.

4. Application Engineering Encapsulation

Engineering techniques like Retrieval-Augmented Generation (RAG), domain-specific fine-tuning, function calling, and agent orchestration encapsulate general model capabilities into verifiable business capabilities. This layer is critical not only for delivery quality but also for value capture.

5. Four Recommendations for Enterprises and Developers

1. Adopt a TCO Perspective Over Unit Price: A low unit price does not equate to a low total cost. Rework rates, latency, and stability should all be factored into the cost calculation.
2. Start with Scenarios, Not Models: First define measurable business tasks, then reverse-engineer model selection and engineering architectures.
3. Design Lock-in-Free Architectures: Isolate model interfaces through abstraction layers to retain multi-model routing and switching capabilities.
4. Accumulate Data Assets: Knowledge bases, annotated data, and feedback loops are long-term moats. The models themselves will continue to commoditize.

6. Conclusion

Call volume measures a traffic mindset, whereas revenue quality measures a value mindset. As model capabilities gradually converge, the decisive factor in competition will shift from “who has the higher benchmark score” to “who can translate capabilities into business results.” For model vendors, this means restructuring business models—moving from selling tokens to selling solutions and outcomes. For enterprise users, it represents an opportunity to transition from “experimenting and playing” to “refined operations.” The token business may be getting harder, but the value business has just begun.