How the MaaS Model Solves the "Rising Revenue, Stagnant Profit" Dilemma in LLM Services

    This article examines the Model as a Service (MaaS) delivery model and the widely discussed risk that LLM platforms grow revenue without growing profit. It explains why revenue and cost are tightly coupled in MaaS, breaks down the unit economics with a simple model, and presents two levers for fixing the problem: pushing inference efficiency as low as possible (quantization, KV cache reuse, continuous batching, elastic scheduling) and continuously upgrading toward high-value customers. It also outlines the common capabilities a sustainable MaaS platform needs and explains why ARR, read alongside gross margin and customer mix, is a better signal of business health than headline revenue.

    1. What Is MaaS? The Business and Technical Essence of Model as a Service

    Model as a Service (MaaS) is a delivery paradigm that has matured alongside large language models. It packages a model’s inference capability behind standardized, elastically scalable API endpoints, so enterprises and developers no longer need to build expensive training and inference clusters. Instead, they call the capability on demand through an API and integrate natural language understanding, content generation, and complex reasoning directly into their own systems.

    Unlike traditional software that charges per seat or per license, MaaS uses the Token as its basic unit of measurement and billing. Every time a user inputs or outputs a certain amount of text, a corresponding number of Tokens is consumed, and the platform bills accordingly. This model has two immediate consequences:

    1. Revenue is tied to actual depth of usage. The more frequently a customer calls the API and the more deeply the model is embedded in their workflows, the more recurring revenue it generates.
    2. Cost scales with usage in lockstep. Behind every inference call sits real consumption of compute, GPU memory, and power.

    It is the second point that creates the thorniest and most debated commercial challenge of the MaaS model.

    2. The Core Tension: Where Does “Revenue Without Profit” Come From?

    A long-standing concern about the MaaS model is whether it will fall into the trap of “rising revenue but stagnant profit” — scaling the top line while margins fail to keep up.

    The root cause is the strong coupling between the revenue structure and the cost structure. In MaaS, revenue is proportional to Token consumption, and Token consumption corresponds directly to continuously incurred inference compute cost. When the revenue curve and the cost curve are bound to the same chain, business growth does not automatically improve the profit margin. If the rate at which the cost per Token falls cannot keep pace with price concessions and scale expansion, the business slides into an awkward situation: the more it does, the thinner its margin becomes.

    A simplified unit-economics model makes this mechanism clear:

    1. Gross margin per request = Revenue per call − Compute cost per call
    2. Revenue per call = Token count × Price per token
    3. Compute cost per call = (Token count × Compute cost per token) + Amortized fixed resources
    4. If the compute cost per token falls quickly with scale and optimization,
    5. then per-request margin rises and gross margin improves;
    6. conversely, if the unit price keeps dropping due to market competition
    7. while compute cost stays relatively rigid,
    8. then per-request margin is continuously squeezed, and the larger the scale, the less economical it becomes.

    This reveals the key to breaking the deadlock: what truly determines profitability is not call volume alone, but whether you can systematically drive down the “compute cost per Token” while simultaneously improving the quality of the revenue mix. These two directions correspond to inference-efficiency governance on the technical side and customer-mix upgrades on the business side.

    3. The Fix (Part 1): Pushing Inference Efficiency to the Limit

    To improve unit economics, the most direct technical lever is to drive down the compute consumed by each inference. There is no single silver bullet here; instead, the gains come from stacking several general-purpose engineering techniques:

    • Model compression and quantization: Reducing memory footprint and computation by quantizing weights (for example, converting high-precision floating-point parameters into low-bit representations), trading acceptable accuracy loss for higher throughput and lower cost per Token.
    • KV Cache reuse: Caching historical key-value pairs during autoregressive generation avoids redundant computation and is central to cost reduction in multi-turn conversations and long-context scenarios. For requests that share a highly identical prefix, prefix caching goes further, so that common prefixes such as the system prompt are computed only once.
    • Continuous batching: Compared with static batching, continuous batching allows requests to be dynamically inserted and replaced during generation, significantly improving GPU utilization and reducing idle compute.
    • Scheduling and elastic scaling: Intelligently routing across multiple inference instances based on real-time load, combined with on-demand scale-up and scale-down, keeps resources resilient during peaks and avoids idle capacity during troughs.

    The following illustrative scheduling and caching pseudocode shows the basic idea of “prefix reuse plus dynamic batching” (for logic illustration only, not a specific product interface):

    1. def handle_request(req):
    2. # 1. Prefix cache hit: reuse already-computed KV for shared prompts and system instructions
    3. prefix_key = hash(req.system_prompt)
    4. if prefix_key in kv_cache:
    5. cached_kv = kv_cache[prefix_key]
    6. else:
    7. cached_kv = prefill(req.system_prompt)
    8. kv_cache[prefix_key] = cached_kv
    9. # 2. Add the new request to the dynamic batch, grouping in real time under memory and latency constraints
    10. batch.add(req, kv=cached_kv)
    11. # 3. Continuous batching: free a slot as soon as a request finishes and slot in new requests dynamically
    12. return scheduler.run(batch)

    It is worth stressing that inference-efficiency optimization is an ongoing, iterative process. As model size, context length, and concurrency patterns change, optimization strategies must be adjusted dynamically rather than set once and forgotten.

    4. The Fix (Part 2): Continuously Upgrading Toward High-Value Customers

    Compressing costs alone is not enough; the quality of revenue matters just as much. Not all Tokens are of equal value — a customer who calls frequently and is deeply integrated into a core business flow has a very different long-term value to the platform than one who tries the service occasionally and may churn at any time.

    Continuously raising the share of high-value head customers is another main line for improving overall gross-margin structure, for several reasons:

    • Scale dilutes fixed costs: The steady, high-frequency calls of large customers more fully amortize fixed investments such as cluster construction, operations, and R&D, so unit cost falls as scale grows.
    • Predictable demand means higher utilization: A stable call curve makes capacity planning and elastic scheduling easier and reduces the redundant resources reserved to handle peaks.
    • Stickiness comes from deep integration: When a model’s capability is embedded in a customer’s core business processes, switching costs rise, renewals and expansions become more sustainable, and revenue becomes significantly more predictable.

    Correspondingly, a platform’s revenue quality can shift gradually from “one-off project driven” to “sustained usage driven.” As revenue relies more on customers’ steady, continuous use of the model rather than sporadic project-based delivery, the predictability and overall health of the business improve accordingly.

    5. An Architectural View: Building a Sustainable MaaS Platform

    From an engineering standpoint, a MaaS platform capable of supporting the goals above usually requires several general-purpose capabilities working together:

    • A unified inference gateway: Handling authentication, rate limiting, routing, and metering — the foundation for accurate billing and multi-tenant isolation.
    • Multi-model routing and tiered service: Routing requests to differently sized service instances based on their complexity and value, serving long-tail demand at lower cost while reserving premium resources for high-value scenarios.
    • An observability stack: Using logging, monitoring, and alerting, together with metrics analysis, to track Token consumption, latency, error rates, and resource utilization in real time, providing the data needed to optimize both cost and quality.
    • An elastic resource foundation: Leveraging container platforms and elastic scheduling to deliver compute on demand and keep costs under control.

    Most of these capabilities are common, mature, and highly composable industry components. When building in-house, enterprises can often integrate these general-purpose capabilities on demand through a cloud platform and focus their energy on model quality and scenario tuning rather than reinventing the wheel.

    6. Key Metrics: Reading Business Health Through ARR

    To assess the sustainability of a MaaS business, Annual Recurring Revenue (ARR) is a better reflection of long-term value than “current total revenue.” It measures the scale of revenue the business can generate stably and repeatably over a year on top of its existing customer base.

    In the MaaS context, ARR carries particular weight: it reflects not how many one-off projects were signed, but whether customers keep calling the model and steadily consuming Tokens. Healthy ARR growth usually means customer usage is shifting from “trial and experimentation” toward “deep reliance,” and that the business can weather short-term volatility. Only when ARR is read together with gross margin and customer mix can you form a fairly complete picture of a model-service platform’s commercial health.

    7. Best Practices: The Bottom Line

    Overall, escaping the MaaS “revenue without profit” dilemma comes down to moving forward on two legs:

    1. Continuously driving down per-unit compute cost on the technical side: Systematizing and normalizing optimization techniques such as quantization, KV Cache, continuous batching, and elastic scheduling, so that the “cost per Token” keeps falling as scale grows.
    2. Steadily raising the share of high-value customers on the business side: Using deep integration and stable, high-frequency usage to upgrade revenue quality, and using scale to amortize fixed costs.

    When the speed at which the cost curve falls consistently outruns the speed at which prices decline, and when revenue becomes increasingly recurring and predictable, the MaaS model returns to its commercial essence — not “the more you do, the thinner you earn,” but “the deeper the usage, the higher the value, the more stable the profit.” For developers and platform operators alike, building a clear unit-economics model and persisting with two-way optimization across technology and business is the key to going far and steady on this path.