Crisis and Opportunity for Chinese LLMs: Rebalancing Volume and Value

1. The Industry Enters a Reassessment Phase

Over the past year, the LLM industry has exhibited three parallel trends:

• Expanding parameter scales: The parameter counts and training budgets for flagship models continue to rise. Reasoning models have introduced “test-time compute” as a new variable in addition to training compute.
• Rapid popularization of open-source models: Open-weight models have seen a massive increase in deployment within the global developer community, with private deployment becoming a key implementation pattern for enterprises.
• Plummeting token prices: Hardware iteration, inference optimization, and intense competition have driven unit inference costs down continuously. In some scenarios, API calling prices have dropped by nearly an order of magnitude within a year.

However, beyond the boom in scale and low prices, the market has begun to re-examine this industry with stricter metrics: How much do model capabilities actually lead? Can commercial revenue cover R&D investment? Who is truly paying for “intelligence”? When the capabilities, revenue, and investments of Chinese and overseas LLMs are placed within the same coordinate system for comparison, both gaps and opportunities become clearly visible.

2. Evaluation System Upgrades: Re-quantifying the Capability Gap

2.1 From Static Benchmarks to Real-World Tasks

Traditional public benchmarks (multiple-choice questions, fixed datasets) face two issues: First, scores are reaching saturation, with top models approaching perfect scores, reducing differentiation. Second, there is a risk of data contamination, where evaluation questions leak into training data, artificially inflating scores. Therefore, the next generation of evaluation systems exhibits two characteristics:

• Private question banks: Evaluation questions are not made public to prevent targeted training.
• Real-world tasks: Models are measured by real workloads such as long-context coding, multi-step tool calling, and agent task execution.

A recent update to a comprehensive capability index by a third-party evaluation agency, which incorporated more private evaluations and real-world tasks, revealed a clearer landscape: Top overseas closed-source models lead with a significant advantage; the first tier of Chinese models sits in the middle with a considerable score gap from the top; open-weight models lag further behind. It should be noted that the choice of evaluation weights and task types influences specific rankings, and a higher index does not exactly equate to “better usability.” However, the trend is stable: On the highest difficulty tasks, top laboratories still maintain a generational advantage.

2.2 Where Does the Gap Come From?

The capability gap is not caused by a single factor but is an overlay of three dimensions:

1. Pre-training investment: The gap in compute scale and data engineering directly determines the upper limit of a model’s foundational capabilities.
2. Post-training methods: Technologies such as Reinforcement Learning from Human Feedback (RLHF), verifiable rewards, and process supervision determine the model’s performance in complex reasoning and agent tasks.
3. Iteration speed: Top laboratories release improved versions on a weekly basis. Engineering organizational efficiency is a competitive advantage in itself.

3. The Scissors Gap Between Usage and Revenue

3.1 A Set of Contrasting Data

Statistics from a public model API aggregation platform over the past month (a single-platform sample, not representative of the entire market, but directionally insightful) show:

• Chinese models contributed about two-thirds of the platform’s token usage, yet the corresponding billed spend share was less than 30%.
• Overseas models captured over 60% of the spend with only about a quarter of the usage.

Zooming into a single vendor dimension makes this more typical: A Chinese open-source model supplied about 30% of the platform’s tokens, but its spend share was only in the single digits; meanwhile, an overseas top laboratory captured about 40% of the spend with less than 10% of the usage.

def value_capture_ratio(token_share: float, spend_share: float) -> float:
    """Value capture ratio = spend share / token share
    <1 indicates "trading volume for price"; >1 indicates high revenue from low-volume, high-price usage"""
    return spend_share / token_share
print(round(value_capture_ratio(0.65, 0.29), 2))  # Approx. 0.45: Chinese models overall
print(round(value_capture_ratio(0.25, 0.61), 2))  # Approx. 2.44: Overseas models overall

This metric quantifies the structural issue: For every unit of usage supplied by Chinese models, the value captured is less than one-fifth of that captured by overseas models during the same period.

3.2 Three Causes of the Scissors Gap

First, pricing strategy. Chinese models generally trade low prices for penetration rates, with unit token prices significantly lower than overseas flagship models. For the same one million tokens, prices can differ by an order of magnitude, directly compressing the conversion ratio from “usage share” to “spend share.”

Second, scenario structure. Low-priced models are mostly routed to high-throughput, low-value-added tasks—batch summarization, translation, content rewriting, and data cleaning. For complex coding, agent planning, and deep reasoning tasks, users still prefer the most capable closed-source flagships. Usage volume can be inflated by being cheap, but the average revenue per user can only be supported by capability.

Third, open-source diversion. Opening weights means a massive amount of inference demand flows to third-party hosting services or enterprise self-built clusters, leaving the official API to capture only a fraction of the value. Open source brings ecosystem influence and developer mindshare, but objectively dilutes direct monetization capabilities.

4. The Flywheel of Top Laboratories: The Positive Cycle of Technology, Capital, and Compute

In contrast to the usage boom, overseas top laboratories have already validated their commercial closed-loop. A recent updated filing from a top laboratory disclosed that its revenue for the first three quarters grew about thirteen times year-over-year, reaching the tens of billions of dollars scale; in the same period, compute and infrastructure spending exceeded half of its revenue.

These numbers illustrate two points:

1. Enterprise customer revenue has become the primary source of R&D investment. The business model of top laboratories is not “burning money for users,” but rather “capability -> enterprise contracts -> cash flow -> reinvestment.” Revenue growth and spending growth interlock to form a positive cycle.
2. The barrier to entry has been raised by capital expenditure. The training cost for next-generation models is measured in billions of dollars. Thin profits from low-priced tokens cannot support investments of this magnitude.

This is the core of the “crisis”: If the value capture ratio does not improve, today’s lead in usage volume may simply be paying for tomorrow’s capability gap.

5. Opportunity in Crisis: Structural Opportunities for Chinese LLMs

The gap is real, but the opportunities are equally structural.

5.1 Early Penetration of the Open-Source Ecosystem

Open-weight models have entered numerous real-world workflows, accumulating three layers of assets: developer mindshare, toolchain adaptation experience, and private deployment use cases. For enterprise customers, private deployment naturally aligns with data security and compliance requirements—this is an exclusive entry ticket for the open-source route. Usage volume itself proves that the product is embedded in real business operations, an asset that is currently undervalued.

5.2 Engineering Efficiency is the Most Realistic Moat

Under the constraint of limited absolute compute, Chinese LLM providers have generally pushed engineering efficiency to the forefront of the industry. The current mainstream inference optimization tech stack includes:

• MoE (Mixture of Experts) sparsification: Achieving larger total parameter capacity with fewer activated parameters.
• Quantized deployment: FP8/INT8 level quantization compresses memory and bandwidth with almost no loss in precision.
• Speculative decoding: A small model drafts while a large model verifies, increasing decoding throughput.
• KV Cache compression and prefill-decode separation: Reducing the memory footprint of long contexts and allowing the two phases to scale independently.
• Continuous batching: Improving GPU utilization and amortizing the cost per token.

The combination of these technologies is the underlying reason for the continuous drop in token prices, and the foundation for Chinese models to maintain competitiveness in cost-sensitive markets.

5.3 From Selling Tokens to Selling Results

The fundamental way out for value capture lies in changing the billing unit: shifting from token-based billing to billing based on task completion, per seat, or business outcomes. This corresponds to three paths:

1. Capability differentiation: Establishing signature capabilities in high-value dimensions such as long context, multimodal, and deep reasoning to support the premium of flagship products.
2. Enterprise-grade delivery: Packaging Service Level Agreements (SLAs), compliance audits, private deployments, and industry knowledge accumulation into solutions, trading service depth for higher average revenue per customer.
3. Toolchain and ecosystem: The more comprehensive the supporting tools for evaluation, fine-tuning, and observability, the higher the developer migration cost, and the greater the retention value.

It is worth mentioning that domestic cloud providers have generally productized these supporting capabilities. Platforms represented by Baidu AI Cloud provide a universal model service pipeline covering data preparation, model fine-tuning, managed inference, evaluation, and monitoring alerting. Enterprises can combine these on-demand without having to build out the entire infrastructure themselves.

5.4 Developer Perspective: Tiered Task Routing

For developers, changes in the industry landscape directly impact model selection strategies. Rather than chasing a single “most powerful model,” it is better to route based on task complexity:

def route_by_complexity(score: float) -> str:
    """Illustrative: Route by task complexity, balancing performance and cost"""
    if score > 0.8:
        return "Flagship tier model"    # Complex reasoning, agent planning, critical business
    elif score > 0.4:
        return "Mid-tier model"         # Routine generation, summarization, code completion
    return "Lightweight tier model"     # High-frequency light tasks like classification, extraction, format rewriting

Coupled with result caching, batch API calls, and asynchronous queues, most business use cases can compress inference costs by more than half without sacrificing user experience. These engineering practices are precisely what the high-volume Chinese model ecosystem excels at.

6. Conclusion: Crisis and Opportunity are Two Sides of the Same Coin

Returning to the initial premise: The “crisis” for Chinese LLMs lies in the generational capability gap and the excessively low value capture rate—leading in usage volume without converting it into equivalent revenue, while the entry ticket for next-generation models is priced in capital expenditure. The “opportunity” for Chinese LLMs lies in the real penetration of the open-source ecosystem, extreme engineering efficiency, and the massive real-world demand for low-cost intelligence across industrial scenarios.

Usage volume itself is not a report card, but it is a necessary gateway to value. The real question is not “whose tokens run the most,” but “who can convert usage volume into capability iteration and a commercial closed-loop.” Evaluation indices will continue to update, and financing and listing documents will continue to be disclosed—the landscape is far from solidified. For Chinese LLMs, crisis and opportunity have always been two sides of the same coin. The only difference lies in whether the lesson of value capture can be learned over the next year.