From Project-Based Delivery to MaaS: How LLM Companies Are Restructuring Revenue and Gross Margins

    This article examines the shift in how large language model (LLM) companies monetize their technology, from one-off custom project contracts to Model-as-a-Service (MaaS) API billing. It explains why ARR (Annual Recurring Revenue) is the key metric for measuring the health of a MaaS business, breaks down the structural logic behind improving gross margins—amortizing training costs and optimizing inference efficiency—and outlines supplementary metrics such as NDR, customer concentration, and per-token margin. Developers and enterprise buyers will also find practical guidance on what to consider when adopting MaaS platforms.

    1. The Revenue Model Divide: From “Selling Projects” to “Selling Calls”

    Over the past two years, the commercialization paths of LLM companies have split into two main routes. The first is project-based delivery: vendors build or fine-tune models for individual customers, sign contracts per project, and see revenue arrive in pulses tied to contract signings. The second is MaaS: the vendor exposes a trained foundation model through an API, customers pay based on the tokens they actually consume, and revenue accumulates smoothly as call volume grows.

    A recent earnings report from a leading LLM company offers a useful sample of this transition. Its annual revenue has grown from tens of millions of yuan a few years ago to over RMB 700 million, with year-over-year growth exceeding 130%. The core driver of this growth is the MaaS business, which bills primarily on API usage—its MaaS platform’s Annual Recurring Revenue (ARR) has reached approximately RMB 1.7 billion, up by tens of times year over year. Behind this number, a key industry thesis is being validated: the value of an LLM lies not in one-time delivery, but in being continuously called.

    For developers, the difference between these two routes is anything but abstract. Under the project model, model capability is “sealed” inside a customer’s private environment, with low reuse rates. Under the MaaS model, the same model capability is consumed concurrently by thousands or tens of thousands of businesses, and marginal costs get diluted as scale grows. This is the AI-era realization of the cloud-native “capability as a service” idea. Today, mainstream cloud providers—including Baidu AI Cloud—offer LLM API open platforms as infrastructure-level products for developers.

    2. ARR: The Core Metric for MaaS Business Health

    ARR (Annual Recurring Revenue) comes from the valuation and operating analysis frameworks of the SaaS industry. Its definition: the total recurring revenue you can expect over the next 12 months, assuming the current customer base’s subscription or usage behavior remains stable.

    In traditional software subscriptions, ARR is typically the annualized value of all active subscriptions. In a MaaS context, a customer’s “subscription” is effectively continuous usage of model capability, so ARR must be calculated from usage behavior itself. A simplified estimation approach looks like this:

    1. # Illustrative ARR estimation for a MaaS scenario (not a real API; for demonstration only)
    2. def estimate_arr(monthly_token_revenues, months=12):
    3. """
    4. monthly_token_revenues: monthly API call revenue over recent months
    5. Approach: use the average of the last 3 months as the baseline, then annualize
    6. """
    7. recent = monthly_token_revenues[-3:]
    8. avg_monthly = sum(recent) / len(recent)
    9. return avg_monthly * 12
    10. # For example, if a MaaS platform averages RMB 140 million per month
    11. # in API revenue over the last three months, then ARR ≈ 140M × 12 = RMB 1.68 billion

    Why does ARR reflect business health better than total revenue? The answer lies in sustainability and predictability:

    • Sustainability: If revenue is dominated by a few large custom projects, income dries up once those projects end. ARR, by contrast, is built on high-frequency, continuous, real usage—as long as a customer’s business runs, tokens keep being consumed.
    • Predictability: Project-based revenue is hard to forecast. Usage-based ARR lets vendors project future revenue curves with reasonable accuracy, so they can rationally plan GPU procurement and capacity expansion for inference clusters.
    • Quality of growth: Growth in ARR means existing customers are deepening usage and new customers are sticking around, rather than relying on a stream of newly signed projects to stay alive.

    The reported “60x year-over-year ARR growth” essentially signals that the company’s revenue composition has shifted from project-driven stock replacement to usage-driven, compounding growth.

    3. The Structural Logic Behind Gross Margin Improvement: Amortization and Reuse

    MaaS drives significant gross margin improvement because it re-aligns the cost structure with the revenue structure.

    1. Training cost is a sunk cost; inference cost is the marginal cost. Once a model is trained, that cost has already been incurred. Each additional API call mainly adds inference-side expenses: GPU compute, memory, and power. Under the project model, training cost is amortized against a single customer; under MaaS, the same model is shared across a massive customer base, spreading the training cost across billions of token calls.

    2. High concurrency dilutes unit inference cost. The key operational metric for inference serving is GPU utilization. Through engineering techniques such as batching, continuous batching, KV cache reuse, and model quantization, the effective throughput per GPU card rises sharply, and the inference cost per million tokens falls accordingly. These are general-purpose inference optimization techniques, not proprietary capabilities of any single vendor.

    3. Usage-based pricing on the revenue side, declining unit cost on the cost side—gross margin expands naturally. A simplified model captures this:

    1. Gross margin ≈ 1 − (amortized training cost + marginal inference cost) / API revenue
    2. As call volume Q grows:
    3. Amortized training cost ∝ 1/Q, declining continuously
    4. Marginal inference cost per token falls as utilization improves
    5. ⇒ Gross margin rises with scale

    This also explains why MaaS platforms typically adopt tiered pricing: the higher the usage, the lower the unit price. This incentivizes deeper customer usage while further growing the denominator and diluting costs.

    4. Additional Dimensions for Evaluating MaaS Business Quality

    Beyond ARR, several complementary metrics matter in operating analysis:

    • Net Dollar Retention (NDR): The change in spend from the same cohort of customers versus the prior period. Above 100% means existing customers are deepening their usage.
    • Customer concentration: If ARR depends heavily on a few large customers, volatility risk is significant. A healthy structure should be underpinned by stable consumption from a long tail of developers.
    • Token consumption curve: Distinguish “onboarding spikes” from “steady-state consumption.” The former may come from one-off evaluations; the latter reflects real business workloads.
    • Per-token gross margin: Monitor how inference optimizations continuously improve costs, to avoid the trap of growing revenue without growing profit.

    5. What This Means for Developers and Enterprises

    For developers, the maturing MaaS model dramatically lowers the barrier to adopting LLMs: there is no need to build your own training cluster—you can access foundation model capability through standard APIs and build applications with common patterns such as retrieval-augmented generation (RAG) and tool calling. For enterprise decision-makers procuring intelligence via MaaS, the focus should be on the predictability of usage costs, API stability and concurrency capability, and the feasibility of data security and private deployment—rather than unit price alone.

    From an industry perspective, the signal in this earnings report is clear: LLM commercialization is completing the leap from “capability demonstration” to “economic model validation.” Only when revenue is composed of continuous, real usage and gross margin improves with scale does an LLM company truly possess the classic scale economics of the software industry. For the broader AI application ecosystem, this means the supply of foundation model services will become more stable and costs will keep falling—and the ultimate beneficiaries are the developers calling AI capabilities in the cloud every day.