How Heavy R&D Spending Can Close the Commercial Loop for LLM Companies

    LLM company results often show fast revenue growth alongside large losses and rising R&D and sales costs. This article explains how to judge whether that growth comes from durable demand, whether losses fund lasting capability or inefficient competition, and whether model performance, inference cost, and pricing can form a positive loop—using a typical set of disclosed figures without judging any single firm’s investment value.

    In a fast-expanding AI industry, LLM company financials often combine signals that are easy to misread: revenue grows quickly, but losses are also large; R&D keeps rising, and sales expense rises with it; model usage expands fast, while the price per call does not necessarily rise in step.

    Analyzing an LLM company therefore cannot stop at revenue growth, nor can it rely only on traditional internet metrics such as price-to-sales or net profit. Three questions matter more:

    1. Does revenue growth come from durable, real demand?
    2. Are losses mainly building long-term capability, or do they reflect inefficient competition?
    3. Can model capability, compute cost, and commercial pricing eventually form a positive loop?

    The discussion below uses a typical set of publicly disclosed figures to examine the financial structure and operating logic of an LLM business. It does not conclude on the investment value of any single company. The goal is a more general method of analysis.

    1. Fast revenue growth does not mean the business model is mature

    On the disclosed figures, the company’s 2025 revenue was about RMB 724 million, up 131.9% year over year. The growth rate alone is strong. It suggests that paid demand for model API calls, enterprise solutions, intelligent applications, and related technical services is reaching a meaningful scale.

    Revenue growth still needs to be broken down. At least the following dimensions matter.

    Where the revenue comes from

    LLM company revenue usually includes several sources:

    • Model inference call fees
    • Fees for private deployments or technical services
    • Developer platform services
    • Intelligent applications and industry solutions
    • Related services tied to compute, data, and operations

    Profitability differs sharply across these lines. Model API revenue can scale well, but only if inference cost stays below the call price. Private-deployment projects can bring large one-off contract value, yet they often come with long delivery cycles and customization cost. Industry solutions may look high in average deal size, while sales, delivery, and after-sales effort are also heavier.

    So the question is not only “how much did revenue grow?” It is also:

    • How high is the share of recurring revenue?
    • Do API calls turn into stable repeat usage?
    • Is revenue overly concentrated in a few large customers?
    • Is there a large gap between contracted revenue and actual cash collection?
    • Does growth depend on one-time projects?

    If revenue is mostly one-time projects, the next year can swing sharply. If it is mostly ongoing usage, it better reflects the long-term commercial value of the model service.

    More tokens do not automatically mean more profit

    The industry widely expects total token consumption to keep growing quickly through 2026. As coding assistants, enterprise knowledge bases, office automation, intelligent customer service, and multimodal applications expand, both call volume and context length per call can keep rising.

    More tokens do not necessarily mean profit rising in step. Token revenue and token cost are not in a fixed relationship. Call cost depends on several factors:

    • The mix of input and output tokens
    • Context length
    • Model parameter scale
    • Inference concurrency
    • Peak demand versus off-peak utilization
    • KV-cache hit rate
    • Hardware utilization
    • Service level and latency requirements
    • Data security and isolation requirements

    If new usage comes mainly from low-price competitive scenarios, or if enterprise customers are highly price-sensitive, higher volume may only add more revenue and more cost without meaningfully improving profit.

    2. The structure of the loss matters more than the loss figure

    The company reported an annual loss of about RMB 4.718 billion, an adjusted net loss of about RMB 3.182 billion, and R&D spending of about RMB 3.180 billion. R&D expense is almost equal to the adjusted net loss. That is central to reading the operating position.

    It suggests the main tension is not conventional sales inefficiency. The company is in a period of intense technology investment. Capital is flowing into foundation-model training, compute procurement, inference infrastructure, algorithm engineering, and hiring.

    For this kind of loss, it helps to split costs into three parts.

    Capability-building costs

    These include foundation-model R&D, training-data processing, algorithm staff compensation, experimental clusters, evaluation systems, and model safety systems. Such spending usually does not turn into revenue in the same period, but it can shape future model capability and competitive position.

    The core question is whether the spend creates reusable assets.

    A mature training pipeline, a data-governance system, a model evaluation platform, and an inference scheduling system can serve multiple model versions and retain long-term value. If a large share of cost only repeats trial and error, or if model iteration does not improve retention or pricing, the capital value of that R&D falls.

    Scaled operating costs

    These include inference compute, storage, networking, monitoring, disaster recovery, and customer service. Compared with training cost, inference cost is closer to a direct cost incurred as revenue is earned.

    As call volume grows, inference can become one of the most important cost lines. Useful measures include cost per million tokens, average hardware utilization, peak-capacity redundancy, and gross-margin differences across models.

    Go-to-market and sales costs

    The company’s 2025 selling costs were about RMB 428 million. As competition intensifies, the cost of winning enterprise customers often rises further. In large-account markets, sales cycles, commercial effort, project adaptation, and collection cycles can all lengthen.

    Enterprise buyers care about stability, security, delivery capability, and service assurance, not only model quality. An LLM company therefore needs pre-sales consulting, solution design, deployment, and ongoing operations. Sales expense does not fall linearly just because call volume rises.

    3. R&D may keep rising, but spending efficiency must rise with it

    If the company keeps iterating larger and stronger models in 2026, R&D costs above RMB 5 billion would not be surprising. The increase is not only from training. It also includes the long-run cost of running models after launch.

    Training cost and inference cost are two different ledgers

    Training cost is usually episodic. It clusters around pre-training, post-training, and large-scale evaluation. Inference cost is continuous. As long as customers keep calling the model, it keeps accruing.

    Many companies emphasize training scale when they promote a model and understate the long-run cost after deployment. Commercial viability is not only “can it be trained?” It also includes:

    • Can it run stably on limited compute?
    • Can it support high concurrency?
    • Can it control memory pressure from long context?
    • Can it route tasks of different difficulty to different models?
    • Can it hold acceptable latency at peak?

    Engineering optimization is necessary, not a lasting moat by itself

    The industry commonly uses several engineering approaches: separating resources for input processing and output generation, improving attention computation and caching, using more efficient parallel scheduling, and cutting deployment cost through quantization, sparsity, and distillation.

    These techniques matter for lowering training and inference cost, but most of them are shared infrastructure that the industry evolves together. Falling behind quickly creates a cost disadvantage. Keeping up does not by itself create a durable exclusive advantage.

    Differentiation more often comes from:

    • Deep understanding of data and tasks in a specific industry
    • A higher-quality, more stable data loop
    • Tool-use capability for complex tasks
    • A more mature enterprise deployment and security stack
    • Higher switching costs and more frequent continued use
    • The ability to turn model capability into business outcomes

    In short, engineering optimization decides whether the company can run at a reasonable cost. Productization and scenario fit decide whether it can sell at a reasonable price.

    4. A price increase does not prove pricing power

    In February 2026, the company raised API call prices by about 83% and raised prices for coding-oriented packages by about 30%. A price increase is often read as a sign of stronger demand or better product capability. It is not the same as pricing power.

    Whether the increase works depends on at least three observations.

    Do customers keep using the service after the increase?

    If many customers switch to other model services, substitutability is high. If call volume stays high, the model is already embedded in workflows and switching cost is higher.

    Does the increase come with a capability upgrade?

    Model prices are rarely adjusted in isolation. They are tied to context length, complex reasoning, response speed, stability, and enterprise-grade service. A pure price hike invites pushback. If a new version clearly reduces manual work, improves code-generation quality, or shortens a business process, customers may care more about cost per completed task than about the price of a single token.

    Does the increase improve gross margin?

    After the price rise, added compute, service, and sales costs still have to be deducted. If the model upgrade also raises inference overhead, the gross-margin gain can be smaller than the headline increase.

    In China’s enterprise market, buyers often have strong bargaining power, long procurement processes, strict budget approval, and slower collection. The size of a price increase is therefore constrained by what customers will actually pay. A company cannot simply copy the pricing of premium overseas models. It has to balance performance, service, compliance, and cost.

    5. Five metrics that matter more than revenue and net profit alone

    A combined set of indicators is more useful than revenue and net profit in isolation.

    Revenue quality

    Look at recurring revenue, renewal rates, customer concentration, and cash collection against contracts. A larger revenue base does not mean higher-quality revenue.

    Unit economics

    Calculate revenue and cost per token, per task, or per active customer. For enterprise applications, also ask how much model cost is required to complete one effective business task.

    R&D return

    After R&D spending rises, does model capability, customer growth, pricing, or cost improve? If R&D keeps rising while commercial metrics do not, spending efficiency needs to be reassessed.

    Compute utilization

    Idle compute directly erodes profit. Peak periods need capacity, but holding too much spare capacity for long periods creates fixed-cost pressure. A strong scheduling system balances performance, stability, and utilization.

    Cash-flow margin of safety

    LLM companies can lose money for a long time, so cash reserves and financing capacity matter. Watch operating cash flow, capital expenditure, payables cycles, and funding needs over the next one to two years.

    6. Competition is shifting from model races to operating efficiency

    Foundation-model capability still matters, but competition is moving from “who trains a larger model first” toward “who can deliver stable business outcomes at lower cost.”

    Companies need to solve four problems at once:

    1. Is model capability sufficient for complex tasks?
    2. Can the inference system control unit cost?
    3. Can the product sit inside the customer’s daily workflow?
    4. Can revenue cover ongoing R&D and operating spend?

    Doing only the first may create technical influence without stable profit. Pushing only on cost can sacrifice model quality and customer experience. A sustainable path is a closed loop among model capability, engineering efficiency, customer value, and commercial pricing.

    Overall, this kind of report describes an LLM company still in a high-investment phase. Revenue growth shows demand taking shape. Very large R&D spend shows a contest for foundational capability. Large losses also mean the business model is not yet fully proven. Investors, developers, and enterprise customers should not stop at model leaderboards or revenue growth. They should also examine cost structure, call quality, retention, unit economics, and cash flow.

    For an LLM company, technical leadership is only the starting point. Long-term value depends on turning R&D into reusable engineering capability, turning that capability into a stable product, and turning the product into sustainable cash flow.