LLM MaaS Financials: Growing ARR and Improving Gross Margins
This article examines the shift from project-based AI revenue to recurring Model-as-a-Service (MaaS) usage, using disclosed financial figures as a reference point. It explains why ARR and recognized revenue differ, why Token-linked inference costs can constrain margins, and how inference optimization, pricing design, and customer mix can improve MaaS unit economics.
LLM MaaS Financials: Growing ARR and Improving Gross Margins
1. From Project Revenue to Recurring Usage: The Core Shift in MaaS
Public financial disclosures show that a large language model company generated more than RMB 724 million in revenue for the year ended December 31, 2025, up 131.9% year over year. Its revenue was RMB 57 million, RMB 125 million, and RMB 312 million from 2022 to 2024, respectively, indicating accelerating growth. Based on disclosed figures, its revenue scale is among the leading companies in the LLM market.
More important than a single revenue number, however, is the business-model transition reflected in its revenue structure. The company has gradually moved from project-based delivery toward a Model-as-a-Service (MaaS) model, providing enterprise customers and developers with AI capabilities through API usage.
Project-based revenue is typically recognized once or according to delivery milestones. Growth depends on winning new contracts and investing implementation resources. In contrast, MaaS revenue is tied to ongoing model usage. Customers must continue making API calls and consuming Tokens for revenue to recur. The growth model therefore shifts from winning projects to retaining usage.
In this example, the MaaS API platform reported annual recurring revenue (ARR) of approximately RMB 1.7 billion, representing a 60-fold year-over-year increase. ARR is commonly used to assess the durability of recurring revenue. In MaaS, it is particularly useful for showing whether customers continue using models and consuming Tokens, rather than relying on one-time projects.
ARR should not be treated as the same as recognized accounting revenue. It is an annualized indicator of recurring revenue capacity, while actual revenue recognition remains subject to contract terms, delivery progress, and usage settlement cycles. ARR should therefore be evaluated alongside usage retention, customer concentration, consumption volatility, and gross-margin quality. Annualized revenue should not be interpreted as realized profit.
2. Why MaaS Can Grow Revenue Without Growing Profit
A common question surrounding MaaS is whether it can fall into a “revenue growth without profit growth” trap. The reason is straightforward: MaaS revenue is directly linked to Token consumption, while every Token also carries ongoing compute costs. When revenue and cost are tied to the same usage chain, scaling volume can put pressure on profit margins.
Without cost engineering and disciplined customer management, a MaaS provider may end up in a situation where more usage produces thinner margins.
A simplified MaaS gross-margin formula is:
Gross Margin = (Token Usage Revenue - Inference Compute Cost - Platform Operations Cost) / Token Usage Revenue
Inference compute costs are affected by model size, context length, concurrency utilization, cache hit rates, and scheduling efficiency. Platform operations costs include gateways, metering, monitoring, security, and customer support. If revenue grows through low-price volume while inference costs do not decline at the same pace, gross margins will erode.
The referenced company provides a positive signal: gross margin for its MaaS API platform increased from 3.3% in 2024 to 18.9% in 2025, showing a meaningful improvement in profitability. Based on public disclosures, its overall gross margin remained relatively stable from 2022 to 2024, staying above 50% over the longer term.
This suggests that MaaS is not inherently a low-margin business. The key is whether a provider can convert higher usage volume into higher gross profit through engineering efficiency and a stronger customer mix.
3. Three Paths to Higher Gross Margins
3.1 Improve Inference Efficiency and Reduce Token Costs
Inference is the central cost driver in MaaS. Common engineering approaches include:
- Continuous batching to improve GPU utilization
- KV cache usage to reduce repeated context computation
- Quantization and distillation to reduce per-inference overhead
- Model routing to send requests to models of an appropriate size
- Caching to reduce the cost of repeated queries
- Elastic scheduling and hot/cold workload tiering to improve resource utilization
The shared goal is not simply to reduce model capability. It is to lower the cost per Token while maintaining acceptable output quality.
For example, request-level cost accounting can be represented with the following pseudocode:
def request_margin(request, model_tier, cache_hit, unit_cost):input_tokens = request.input_tokensoutput_tokens = request.output_tokens# Different model tiers have different unit costscost = (input_tokens + output_tokens) * unit_cost[model_tier]if cache_hit:cost *= 0.3 # Cache hits reduce repeated computation costsrevenue = request.billed_amountreturn revenue - cost
In production, Token metering, cache-hit data, model routing, and cost attribution should be connected to create a gross-margin view by tenant, model, API endpoint, and time period. Total API call volume alone is not sufficient.
3.2 Token Economics and Tiered Pricing: Raising Price Without Losing Usage
“Raising price without reducing usage” does not mean applying a simple price increase. It means aligning price with delivered value.
For use cases requiring high availability, low latency, strong compliance, or long-context processing, customers may be willing to pay for service-level agreements, security governance, and dependable access. MaaS platforms can use common pricing approaches such as:
- Different pricing for input and output Tokens
- Tiered pricing based on context length
- Discounts for cache hits
- Discounts for committed usage volumes
These approaches can make customer spending more predictable while preventing low-quality, low-price traffic from eroding margins. Pricing adjustments must be accompanied by improvements in model quality, service reliability, and engineering efficiency. Otherwise, customers may reduce their usage.
3.3 Increase the Share of High-Value Enterprise Customers
A larger share of high-value customers can also improve gross margins. Large customers often have deeper usage, stronger use-case stickiness, and higher requirements for reliability and service assurance. Their willingness to pay is not determined solely by unit pricing; they also value the overall service proposition.
Moving from broad but shallow low-price trials to deep, stable, production-grade usage can improve both average revenue per user and margin quality. At the same time, customer concentration risk must be monitored. Changes in usage by a small number of major customers can affect ARR stability.
4. Engineering Metrics and the Operating Loop
To grow ARR while improving gross margins, MaaS providers need an operating loop that connects metering, cost, pricing, and retention:
- Track Token consumption and gross margin by tenant and model.
- Set budgets, quotas, and alerts.
- Optimize routing for high-cost requests.
- Provide SLAs and dedicated support for high-value customers.
- Regularly assess cache-hit rates, cost per Token, customer retention, and ARR quality.
Usage volume, cost, revenue, and retention must appear on the same operating dashboard. This helps avoid an undisciplined growth model that focuses only on revenue while ignoring margin quality.
5. Risks and Limitations
Several factors should be considered when interpreting MaaS financial results:
- ARR and recognized revenue are different metrics.
- Prepayments and committed usage may inflate short-term annualized figures.
- Customer concentration can affect the stability of recurring revenue.
- Changes in compute pricing and model iteration costs can affect margins.
- Pricing competition may compress profitability.
Healthy MaaS growth should not focus only on call volume. It should prioritize sustainable usage retention and gross-margin quality.
6. Conclusion
This example shows that the LLM MaaS model is moving from project-driven growth toward recurring usage-driven growth. ARR of approximately RMB 1.7 billion, with a 60-fold year-over-year increase, demonstrates the scaling potential of recurring API usage. The improvement in MaaS API gross margin from 3.3% to 18.9% shows that inference optimization, Token cost reduction, tiered pricing, and a higher-value customer mix can help MaaS businesses address concerns about revenue growth without profit growth.
For technical teams, the central tasks are clear: understand Token economics, reduce inference costs, and deepen customer value. Each API call should become a measurable, manageable, and sustainable unit of growth.
