KNOWLEDGE ATLAS has demonstrated its potential for profitability, yet the company continues to report losses. If a business must invest 4.4 yuan in research and development for every 1 yuan of revenue it earns, can it be considered a good business? KNOWLEDGE ATLAS AI, a leading global large language model company, is attempting to answer this question. On March 31, 2026, KNOWLEDGE ATLAS released its full-year 2025 financial results, marking its first earnings report since going public.
The financial report shows that KNOWLEDGE ATLAS achieved total revenue of 724 million yuan in 2025, a year-on-year increase of 131.9%. The cost of sales surged by 213.3% to 427 million yuan. Despite this revenue and cost performance, the company's gross profit grew by 68.7% to 297 million yuan. The comprehensive gross profit margin for the year reached 41%. Compared to fellow large model listed company MiniMax, KNOWLEDGE ATLAS's gross margin is significantly higher; MiniMax reported a gross margin of 25.4%.
A breakdown of the revenue composition reveals a divergence behind the 41% gross margin. The gross margin for the on-premises deployment business decreased from 66.0% in 2024 to 48.8% in 2025. Conversely, the gross margin for the cloud deployment business rose from 3.3% in 2024 to 18.9% in 2025. From this perspective, KNOWLEDGE ATLAS has sufficiently demonstrated its profitability potential.
However, the company remains in a loss-making position. Impacted by research and development expenses of 3.182 billion yuan during the same period, KNOWLEDGE ATLAS reported an adjusted net loss of 3.182 billion yuan, representing a 29.1% year-on-year increase in the loss amount. This loss figure is equivalent to 4.39 times the company's total revenue and 10.7 times its gross profit. It is noteworthy that the losses primarily stem from R&D expenditures. According to the financial report, KNOWLEDGE ATLAS's R&D costs in 2025 were 3.184 billion yuan, a 44.9% increase year-on-year, while capital expenditure for 2025 was 74.7 million yuan.
The report explains that the growth in R&D costs was mainly due to: (1) Increased employee costs, including the expansion of the R&D team and higher share-based payment expenses. (2) Payments to third-party computing power suppliers for computational services, covering model iteration and investment in more advanced model training infrastructure.
It is worth mentioning that the computing power costs specifically used for large model training are not classified as R&D expenses. Instead, they are separately accounted for as capital expenditure under a computing power leasing model. In KNOWLEDGE ATLAS's context, the former refers to the cost of utilizing GPU resources from suppliers based on actual model training hours, which is considered a flexible expense and included in R&D costs. In contrast, locking in GPU resources through long-term contracts with specific suppliers is categorized as capital expenditure.
Compared to MiniMax, KNOWLEDGE ATLAS operates on a larger overall scale, a difference primarily driven by variations in business composition and organizational structure. For instance, KNOWLEDGE ATLAS's workforce is twice the size of MiniMax's. This larger scale also contributes to KNOWLEDGE ATLAS incurring higher R&D costs and more substantial losses, while MiniMax demonstrates higher personnel efficiency.
A notable point in this earnings report is that, similar to MiniMax, KNOWLEDGE ATLAS has benefited from the "Lobster" trend. Starting from the first quarter of 2026, KNOWLEDGE ATLAS's performance growth has been largely dependent on the launch of AutoClaw in March, which enables one-click Lobster deployment. According to KNOWLEDGE ATLAS CEO Zhang Peng, the company's API call pricing increased by 83% in the first quarter. This price hike coincided with a surge in demand. The Lobster trend had already been gaining momentum for a month. Within half a month after the price increase, KNOWLEDGE ATLAS began deploying Lobster. Consequently, despite the higher prices, the call volume for KNOWLEDGE ATLAS's GLM model still grew by 400%. The report states that the subscription user base surpassed 100,000 just two days after the plan's launch and exceeded 400,000 users within 20 days.
Correspondingly, a key indicator of KNOWLEDGE ATLAS's profitability is its MaaS platform, which is the company's primary strategic focus. It is reported that the MaaS API platform achieved an Annual Recurring Revenue of 1.7 billion yuan, representing a 60-fold increase year-on-year. In essence, this financial report confirms KNOWLEDGE ATLAS's profitability potential on one hand, while also showing that losses have not ceased on the other.
KNOWLEDGE ATLAS's growth logic has shifted but has not been fundamentally restructured. Looking at the overall revenue structure, the most critical variable in this report is not the total revenue itself, but the source of that revenue. By examining subtle changes in revenue sources, one can discern KNOWLEDGE ATLAS's new growth logic and its sustainability.
A detailed breakdown shows that KNOWLEDGE ATLAS's growth focus is tilting towards the cloud, specifically MaaS. This segment now accounts for 26.3% of revenue, compared to only 15.5% in 2024. Following the earnings release, KNOWLEDGE ATLAS also stated that its strategic focus would continue to be on MaaS. However, although the proportion of cloud deployment business has increased significantly numerically, several variables within this shift are particularly crucial.
First, the core driver is API calls. In other words, this round of growth for KNOWLEDGE ATLAS is essentially driven by an increase in call volume. Within this, Lobster is the most direct variable. As Agents begin to autonomously execute tasks, a single request often corresponds to multiple rounds of calls, amplifying token consumption exponentially and driving up API call volume.
Second is the primary source of MaaS revenue. The report mentions that nine out of ten internet companies have integrated KNOWLEDGE ATLAS's models. A noteworthy change here is that these internet companies generally have their own large models, but they do not rely solely on them. Instead, they call upon different models for different business scenarios. This indicates that, in the short term, even with proprietary models, companies may still choose KNOWLEDGE ATLAS for specific use cases. However, this does not guarantee that these nine major internet firms will maintain this strategy long-term. The call volume from these companies essentially constitutes half of KNOWLEDGE ATLAS's MaaS revenue. Therefore, losing any single major client would significantly impact the current MaaS business.
Third, MaaS growth also comes from token "exports." Over the past year, KNOWLEDGE ATLAS has collaborated with several Middle Eastern and Southeast Asian countries, exporting its model capabilities, which essentially generates revenue through token calls.
Overall, a clear signal from this financial report is that KNOWLEDGE ATLAS is shifting its growth narrative from heavy on-premises deployment to selling model access, i.e., selling tokens. In terms of results, although KNOWLEDGE ATLAS's primary revenue still relies on on-premises deployment, the MaaS model is already showing a trend of sustainable growth.
On this foundation, KNOWLEDGE ATLAS has introduced a new concept: TAC. According to its definition, TAC comprises three components: intelligent call volume, intelligent quality, and economic conversion efficiency. Simply put, it measures how many tokens are called, whether these calls are effective, and if they ultimately convert into revenue. Following the "Lobster" event, the industry has gradually reached a consensus on tokens: when large models possess long-horizon task execution capabilities, a call is no longer a one-time input-output transaction but is organized into a continuously running system. This means a single task often involves multiple rounds of calls, tool usage, and even self-verification. Tokens are not merely consumed but are "orchestrated," reflecting how users organize their interactions with the large model.
The timing of TAC's introduction is understandable. Over the past two years, competition in the large model industry primarily revolved around parameter scale, model capability, and price. However, as price wars subside and model capabilities converge, with Agent applications beginning to proliferate, these metrics are becoming less effective at explaining differences in company growth. Against this backdrop, KNOWLEDGE ATLAS needs a new set of metrics to answer a more practical question: when model capabilities are similar, where does growth come from?
KNOWLEDGE ATLAS faces a "cost trap." Zooming out from KNOWLEDGE ATLAS to the entire industry reveals that the business models for large models are beginning to converge. Apart from one company, the core revenue for the remaining foundational model companies is converging towards API calls. Whether it's KNOWLEDGE ATLAS, MiniMax, or another player, they are all moving towards a path where MaaS drives growth. However, for KNOWLEDGE ATLAS at least, this path did not exist from the outset.
Taking KNOWLEDGE ATLAS as an example, its early business was heavily weighted towards government projects and private deployments, characterized by a project-based model. It was only around half a year before its IPO, in an effort to make its business model more sustainable and scalable, that KNOWLEDGE ATLAS began a noticeable shift towards MaaS, pivoting its growth focus to cloud-based API calls. The results show that this transformation has indeed brought changes: the MaaS share has increased, tokens have become a core metric, and the revenue structure is moving towards a platform model.
However, given KNOWLEDGE ATLAS's existing structure, this composition where on-premises deployment outweighs cloud deployment is difficult to change rapidly in the short term. The current growth of MaaS remains highly dependent on a small number of large clients. The financial report indicates that a significant portion of KNOWLEDGE ATLAS's API revenue comes from major internet companies. While these companies have their own models, they choose to utilize external model capabilities for specific business needs. This "multi-model calling" pattern does provide stable demand for MaaS. The problem, however, is that this does not equate to genuine scalable growth. On one hand, top-tier clients contribute the majority of call volume; on the other hand, the long-tail market has not been truly activated. In other words, the platform form of MaaS has emerged, but platform-scale economics have not been established.
This points to another, more core issue: the cost per token and the revenue structure. The report shows that KNOWLEDGE ATLAS's full-year 2025 loss was 4.718 billion yuan, a 59.5% year-on-year increase. Within this, R&D expenditure reached 3.18 billion yuan, up 44.9% year-on-year, while capital expenditure was 74.7 million yuan, down approximately 83.8% year-on-year. The former includes model training costs and employee costs, while the latter stems from costs like computing power leasing. In 2025, KNOWLEDGE ATLAS adjusted its computing power procurement strategy, shifting from relatively fixed leasing arrangements to a combination of leasing and service procurement, leading to a sharp decline in capital expenditure.
Combining the MaaS growth with these cost figures reveals a very direct logical chain: for a company to drive MaaS growth, it must rely on model capability; improving model capability requires continuous increases in R&D investment. The problem is that R&D and computing power costs do not naturally decrease as call volume scales up. In other words, the prerequisite for revenue growth itself pushes costs higher. This traps large model companies in a structural dilemma: to gain more calls, they need to continuously enhance model capability; and to enhance capability, they must keep increasing investment. This has led to the current situation where faster growth brings greater cost pressure. From this perspective, the issue is not unique to KNOWLEDGE ATLAS but a common constraint facing the entire large model industry. Until this problem is solved, MaaS can drive growth but is unlikely to deliver profits.
Why does KNOWLEDGE ATLAS want to benchmark against Anthropic? During the annual results conference call on the evening of March 31, KNOWLEDGE ATLAS CEO Zhang Peng specifically mentioned the US AI unicorn Anthropic before presenting the results, noting that its ARR grew from $1 billion at the end of 2024 to $9 billion by the end of 2025. In fact, nearly all leading large model companies are attempting to follow the US trajectory. One company is targeting OpenAI's path, combining model capability, product, and subscription. KNOWLEDGE ATLAS and MiniMax, however, are trying to align with the Anthropic model, emphasizing foundational model capability, outputting inference computing power via API, and building a developer ecosystem.
Regardless of the path chosen, the essence is to treat the model as infrastructure and achieve scaled revenue through calls. This path has been preliminarily validated in the US. Both OpenAI and Anthropic demonstrate that when model capability is sufficiently strong, a positive feedback loop can form within the developer ecosystem. The problem is that this path is difficult to replicate in China.
First, there is the difference in pricing systems. In the US market, enterprise customers and developers are more willing to pay for capability, allowing model ability to translate into premium pricing. In China, prices were driven down rapidly from the start. After two years of price wars, tokens have gradually evolved into a "basic resource."
Second, there is a difference in demand structure. The large model ecosystem in the US relies more on long-tail demand from developers. In China, calls are more concentrated among top-tier clients, such as major internet companies and government/enterprise customers. In this structure, MaaS resembles "centralized procurement" rather than being driven by a vibrant developer ecosystem.
Third, there are differences in cost and supply. Computing power supply, chip architecture, and the overall cost environment make it harder for domestic model companies' costs to decline with scale.
Viewing KNOWLEDGE ATLAS's predicament through this lens makes it easier to understand. Looking back at the development paths of the internet and cloud computing, profitability at the infrastructure layer often materializes *after* the application layer explodes. Similarly, this implies that at the current stage, whether it's KNOWLEDGE ATLAS or other large model companies, they likely need to wait for application scenarios to be continuously validated before scale effects can emerge.
Comments