Chinese LLMs: Navigating the Shift from API Call Volume to True Value Share

1. The Structural Disconnect Behind Two Sets of Data

Two recently released industry reports outline a compelling contrast in the global large language model (LLM) landscape.

The first dataset comes from an independent evaluation agency’s updated LLM capability index. Incorporating more private benchmark datasets and real-world task scenarios, the leading closed-source model scored 58 points, followed by a Chinese open-source model at 46. Two other Chinese models with open weights scored 32 points. Almost simultaneously, an overseas top-tier AI lab updated its prospectus: its revenue for the first three quarters of 2026 reached $33.42 billion, a 13-fold year-over-year increase, with computing power and infrastructure spending reaching $18.8 billion. Technology, capital, compute, and enterprise customer revenue are forming a mutually reinforcing positive feedback loop.

The second dataset comes from a model aggregation and API platform. In a recent one-month window, Chinese models accounted for 65% of total Token usage but only 29% of the corresponding monetary spend. Overseas models accounted for 25% of Token usage but captured 61% of the spend. Looking at specific providers, one Chinese model supplied about 32% of the Tokens corresponding to only 6% of the spend, while an overseas model accounted for 7% of Token usage but 40% of the spend.

The massive gap between API call share and value share represents the core crisis and opportunity currently facing Chinese LLMs.

2. Benchmark Scores vs. Commercial Value: Different Metrics

Benchmark indices provide a horizontally comparable scale of capabilities, but there is a conversion funnel between these scores and actual commercial value. Understanding this funnel requires distinguishing three layers:

1. Basic Capabilities: Performance in general tasks like reasoning, coding, and multilingual understanding.
2. Task Reliability: Stability in real-world production scenarios involving long reasoning chains, multi-turn interactions, and tool calling.
3. Enterprise Trust: Data compliance, Service Level Agreements (SLAs), auditing capabilities, and ecosystem integration costs.

Leading closed-source models have established high switching costs at the second and third layers. Once enterprise customers integrate their core business workflows into a specific model, switching costs include not only API refactoring but also implicit investments like prompt engineering accumulation, benchmark rebuilding, and compliance review cycles. This allows pioneers to maintain higher pricing power, making it difficult for latecomers to achieve the same premium in the short term, even if they approach parity on certain benchmark metrics.

3. The Ruthless Arithmetic of Token Economics

The conversion rate from Token usage to spend share can be broken down into a straightforward formula:

spend_share = token_share * average_price_per_token_ratio

Take the overseas model mentioned earlier: contributing 40% of the spend with 7% Token usage implies its average price per Token is about 5.7 times the platform average. The Chinese model, with 32% Token usage and 6% spend, has an average price per Token of roughly 0.19 times the platform average. The price multiplier difference between the two is nearly 30x.

This gap is not solely explained by “brand premium.” Behind the price difference lies:

• Inference Cost Structure: Larger parameter models consume more compute power per inference.
• Service Level Differences: The cost of isolation, availability commitments, and technical support for enterprise-grade deployments.
• Capability Scarcity: Limited alternative solutions for complex reasoning, long context, and high-reliability code generation tasks.

The strategy of Chinese models trading low prices for API call volume is effective during the cold start phase of a developer ecosystem. However, lingering in a “low price - low margin - low R&D investment” cycle could lead to a passive market position.

4. Open Weights: The Double-Edged Sword

Chinese models with open weights scoring 32 points shows a clear gap compared to the 58 points of leading closed-source models, but the open-source strategy brings another layer of dynamics.

Advantages:

• Rapid penetration into the developer community, establishing a de facto standard.
• Accelerated iteration through community feedback.
• In data-sensitive scenarios, private deployment is a rigid requirement.

Risks:

• The commercialization path of open-source models relies on cloud provider hosting services, and the homogeneous competition among these services further drives down prices.
• If the capability gap between the open-source and commercial versions is too large, it may damage community trust; if too small, the commercial version struggles to command a premium.

The balance lies in the open-source version covering general capabilities while the commercial version builds differentiation in reliability, compliance, and large-scale service capabilities. This strategy has mature precedents in infrastructure domains like databases and container orchestration.

5. From “Usable” to “Essential”: Model Selection Logic for Developers

For developers and enterprise technical decision-makers, model selection is not simply about comparing benchmark scores; it is a multi-objective optimization problem. Below is a simplified decision framework:

def model_selection(candidates, constraints):
    """
    candidates: List of candidate models, each containing capability score, unit price, latency, compliance level
    constraints: Business constraints, such as max latency, min compliance level, budget limit
    """
    feasible = [m for m in candidates if meets_constraints(m, constraints)]
    # Sort by cost per task, rather than simply by Token unit price
    ranked = sorted(feasible, key=lambda m: cost_per_task(m))
    return ranked
def cost_per_task(model):
    # Cost per task = cost per call * average retries / success rate
    return (model.price_per_call * model.avg_retries) / model.success_rate

The key insight is: Low unit price does not equal low total cost. A model with a low unit price that requires multiple retries and manual fallbacks may have a higher overall cost than a model with a higher unit price but a high first-pass success rate. If Chinese models want to increase their value share, they must establish an advantage in the real-world metric of “cost per task,” rather than relying on the marketing rhetoric of “price per million Tokens.”

6. Breaking the Deadlock: From Price Wars to Value Wars

Based on the above analysis, shifting from an advantage in API call share to an advantage in value share requires sustained investment in the following areas by Chinese LLM providers:

1. Deep Optimization for Vertical Scenarios
Catching up on general capabilities takes time, but in specific vertical domains—such as industrial inspection report generation or public service Q&A—local advantages can be established through domain-specific data fine-tuning and toolchain integration. Customers in vertical scenarios are less sensitive to price and highly sensitive to accuracy and compliance.

2. Structural Reduction in Inference Costs
Through model architecture optimization, quantized inference, and heterogeneous compute scheduling, the compute consumption per Token can be reduced while maintaining capabilities. Cost advantages should be translated into gross margin space, rather than being entirely surrendered as ammunition for price wars.

3. Bolstering Enterprise-Grade Service Capabilities
This includes but is not limited to: private deployment solutions, fine-grained permission management, API audit logs, SLA guarantees, and integration with mainstream identity authentication systems. These capabilities do not directly reflect in benchmark scores, but they are critical weights in enterprise procurement decisions.

4. Continuously Polishing the Developer Experience
Documentation quality, SDK coverage, clarity of error messages, and community response speed—these “soft skills” directly affect developers’ willingness to migrate and their retention rates.

7. Conclusion

The disconnect between 65% Token share and 29% spend share is both a warning and an opportunity. The warning is that reliance on a low-price, high-volume path may solidify a low-value positioning. The opportunity is that Chinese models have already secured scaled, real-world API usage scenarios, which are valuable assets for iterating capabilities, verifying reliability, and accumulating trust.

Moving from “being called” to “being depended on,” and from “cheap” to “worth it,” has no shortcuts. It requires a continuous climb in technical capabilities, restraint and patience in business strategy, and a profound understanding of the real pain points of developers and enterprise users. Benchmark scores will fluctuate, and capital market trends will change, but the continuous optimization of cost per task and the steady improvement of task success rates will always be the ultimate anchor of a model’s commercial value.