From R&D Spend to LLM Competitiveness: Revenue Mix, Gross Margins, and Delivery Efficiency

    This article explains how high R&D spending at large-model companies can translate into competitive strength through model capability, engineering delivery, commercial closure, and ecosystem reuse. Using structural data disclosed in public materials from a large-model company, it analyzes revenue mix, gross-margin layers, customer repurchase, and cost dynamics. It also offers practical evaluation criteria for developers and enterprises, including total cost of ownership, observability, and replaceability.

    Public operating data from large-model companies often produces two sharply different interpretations. One view focuses on loss rates and low-cost API calls and concludes that even strong model capability is merely a loss-making marketing exercise. The other focuses on public benchmark rankings and developer reputation and concludes that the technology has already achieved global competitiveness. When revenue structure, gross margin, customer repurchase, and R&D investment are placed on the same analytical chart, many seemingly contradictory phenomena are not contradictory at all. High R&D investment does not directly equal short-term profit. Instead, it is translated into competitiveness layer by layer through “model capability—engineering delivery—commercial closure—ecosystem reuse.” This article uses structural data from public materials of a large-model company as an example to discuss this translation path and to propose practical evaluation methods for developers and enterprises.

    1. First, Remove the Misreading That “Losses Mean No Business”

    In the early stage, large-model companies typically concentrate high R&D investment on training compute, inference optimization, data governance, algorithm teams, and engineering platforms. Financially, these investments appear as expenses. Operationally, they accumulate as model capability, toolchains, and delivery experience. Therefore, looking only at the operating loss rate can easily misread an “investment period” as “no demand.” Public materials show that revenue at this type of company rose rapidly from tens of millions of yuan to hundreds of millions of yuan, and half-year revenue reached a substantial proportion of the previous full year’s total. This indicates that the question is not whether there is business, but in what form the business appears, at what pace it grows, and in which links economies of scale have not yet formed.

    The revenue structure is even more important. Public materials show that in the first half of the year, about 84.8% of revenue at a large-model company came from on-premises deployment, while about 15.2% came from cloud deployment. This structure breaks a common assumption: at a time when cloud APIs and subscription models are widely discussed, the main revenue of some leading large-model companies still comes from delivering “landed” model systems to enterprise customers, rather than purely from cloud-based calls. It is closer to a “high-end software delivery provider”: it packages model capability, inference services, data security, industry adaptation, and on-site support into an acceptance-based project. This revenue form is not asset-light, but it can support early cash flow.

    2. Revenue Structure: On-Premises Deployment Is Cash Flow; Cloud Services Are Incremental Growth

    The core value of on-premises deployment lies in meeting requirements such as data security, compliance isolation, deep customization, and low latency. For internet and high-tech companies, telecom operators, government customers, and large industry customers, general cloud APIs may not cover their intranet data, proprietary knowledge bases, permission systems, and audit requirements. Therefore, they are willing to pay higher fees for “dedicated deployment + dedicated models + on-site support.” Public sharing mentions that internet and high-tech companies account for about half of on-premises customers, while government customers account for less than 20%. The developer base on the model service platform has reached millions, paid conversion exceeds 10%, and annual repurchase by paying customers is about 70%. These signals show that the on-premises business does not rely only on a single industry; it is driven by high-requirement, high-value scenarios.

    Cloud services follow a different logic. They have low marginal cost, strong replicability, and fast developer reach, but during a period of intense competition they often require low prices or even negative gross margins to gain call volume and ecosystem stickiness. Public materials show that the gross margin of on-premises deployment is about 59.1%, the gross margin of cloud deployment is about -0.4%, and the overall gross margin is about 50%. One positive and one negative does not mean “the model is not good.” Rather, the two types of business are at different stages. The former is project-based delivery, earning a premium through engineering capability and customer trust. The latter is a platform business that gradually dilutes costs through scale, inference efficiency, and ecosystem reuse. In one sentence: large projects are “supplying blood,” while the cloud is still “growing its body.”

    3. Gross Margin Layers: The Scale Logic Behind Positive and Negative Margins

    Gross margin layers are key to understanding the translation of R&D investment. The relatively high gross margin of on-premises deployment shows that “model capability + engineering delivery” is not inherently a loss on every order. However, project-based delivery has weak replicability, and delivery cycles, customization ratios, and on-site support affect the speed of scaling. The negative gross margin of cloud deployment shows that inference costs, price competition, and free quotas are still eroding short-term profit. Yet the marginal cost of cloud services will decline as inference optimization, caching, batch processing, model tiering, and scheduling efficiency improve. If cloud call volume expands, unit token cost falls, and paid conversion rises, negative gross margin can gradually narrow.

    Therefore, evaluating a large-model company should not rely only on overall gross margin or only on loss rate. It should examine three curves: the delivery gross margin and repurchase rate of on-premises projects, the unit call cost and paid conversion of cloud services, and whether model capabilities formed by R&D investment can be reused across customers. If all three curves improve at the same time, high R&D investment is truly translating into competitiveness. If only benchmark scores rise while delivery costs do not fall and repurchase does not increase, technical advantages will struggle to form a commercial closure.

    4. How R&D Investment Becomes Competitiveness: A Four-Layer Funnel

    The first layer is model capability. High R&D investment first accumulates into foundation-model capabilities in reasoning, long context, tool calling, code generation, multi-turn dialogue, and multimodal understanding. Public benchmark rankings and developer reputation are important signals, but they cannot replace real business validation. Competitiveness depends not only on a single leaderboard but also on model stability, error rates, interpretability, and security and compliance performance in complex tasks.

    The second layer is engineering delivery. Enterprise customers often buy not “a model” but a model system that can run, be operated, and be audited. Private deployment requires model compression, inference acceleration, resource scheduling, permission isolation, log auditing, monitoring and alerting, and fault degradation. The following is a generic illustrative configuration and does not refer to any specific product:

    1. route:
    2. - path: /v1/chat
    3. auth: tenant_token
    4. quota: per_tenant
    5. backend: model_pool
    6. fallback: [small_model, large_model]
    7. observability: [latency, error_rate, token_cost]

    Such engineering capabilities determine whether a project can be delivered on schedule, whether total cost of ownership can be controlled, and whether it can run stably on a customer’s intranet. R&D investment has cross-project reuse value only when it is translated into delivery templates, toolchains, and operations standards.

    The third layer is commercial closure. On-premises projects bring cash flow and industry knowledge, while cloud services bring developer scale and call data. The two should form a feedback loop: high-frequency tasks in projects feed back into model optimization, and long-tail demand exposed by cloud calls drives model tiering and toolchain improvement. Paid conversion rate, annual repurchase rate, and customer expansion revenue explain competitiveness better than revenue growth alone.

    The fourth layer is global competition. Global competitiveness is not “number one on a leaderboard.” It is the combined result of unit token cost, service stability, tool-calling ecosystem, private deployment efficiency, compliance capability, and developer migration cost. If high R&D investment remains only on the training side and cannot reduce inference costs, improve delivery efficiency, or generate repurchase, it will be difficult to translate into global competitiveness. Conversely, if model capability, engineering platforms, and commercial closure reinforce one another, investment during the loss period may translate into long-term barriers.

    5. How Developers and Enterprises Should Evaluate: TCO and Replaceability

    For developers, evaluating model services should focus on API stability, context length, tool-calling consistency, rate-limiting policies, error-code standards, SDK compatibility, log and metering transparency, and the ability to fall back between small and large models. For enterprises, the focus should be total cost of ownership, not only unit price: inference resources, private deployment labor, data governance, security and compliance, operations monitoring, and migration costs. During selection, it is advisable to run small-scale benchmark tests covering real business tasks and record accuracy, latency, unit task cost, and failure recovery time.

    At the same time, replaceability matters. A unified gateway, abstract model interfaces, standard logs and metering, and configurable routing can reduce vendor lock-in risk. For data-sensitive scenarios, deployment boundaries, permission models, audit granularity, and data retention policies should be clearly defined. For cloud call scenarios, quotas, circuit breaking, caching, and degradation strategies should be set. Only when model capabilities are incorporated into an observable, operable, and metered engineering system can the capabilities brought by R&D investment be consumed stably by enterprises.

    6. Conclusion

    High R&D investment translates into large-model competitiveness not through a straight line of “R&D—leaderboard—revenue,” but through a funnel of “model capability—engineering delivery—commercial closure—ecosystem reuse.” The revenue structure and gross-margin layers in public materials show that on-premises deployment provides cash flow and high-value scenario validation in the early stage, while cloud services take on ecosystem expansion and cost optimization during the scale-investment period. The two are not opposed; they are a combination at different stages. To judge whether a large-model company has global competitiveness, one should look not only at model capability and developer reputation but also at delivery efficiency, repurchase rate, unit call cost, and compliance capability. For developers and enterprises, the rational approach is not to be led by a single loss rate or a single leaderboard, but to conduct comprehensive evaluation using real tasks, total cost of ownership, observability, and replaceability. Only in this way can high R&D investment be more than a financial expense and instead accumulate into sustainable technical and commercial competitiveness.