Kingsoft Cloud Holdings - Earnings Call - Q3 2025

Transcript

Operator (participant)

Good day, and thank you for standing by. Welcome to Kingsoft Cloud's third quarter 2025 earnings conference call. At this time, all participants are in a listen-only mode. After the speaker's presentation, there will be a question-and-answer session. To ask a question during the session, you will need to press star one and one on your telephone. You will then hear an automated message advising your hand is raised. To withdraw your question, please press star one and one again. Please be advised that today's conference is being recorded. I would now like to hand the conference over to your speaker today, Nicole Shan, IR Director of Kingsoft Cloud. Please go ahead.

Nicole Shan (Director of Investor Relations)

Thank you, Operator. Hello everyone, and thank you for joining us today. Kingsoft Cloud's third quarter 2025 earnings release was distributed earlier today and is available on our IR website at ir.ksyun.com, as well as on the TI NewsWire services. On the call today from Kingsoft Cloud, we have our Vice Chairman and the CEO, Mr. Zou, and the CFO, Ms. Li. Mr. Zou will review our business strategies, operations, and other company highlights, followed by Ms. Li, who will discuss the financial performance. They will be available to answer your questions during the Q&A session that follows. There will be consecutive interpretation. Our interpretations are for your convenience and reference purposes only. In case of any discrepancy, management statements in the original language will prevail.

Before we begin, I'd like to remind you that this conference call contains forward-looking statements within the meaning of Section 21E of the Securities Exchange Act of 1934, as amended and as defined in the U.S. Private Securities Litigation Reform Act of 1995. These forward-looking statements are based upon management's current expectations and current market and operating conditions, and are related to events that involve known or unknown risks, uncertainties, and other factors, which are difficult to predict, and many of which are beyond the company's control, which may cause the company's actual results, performance, or achievements to differ majorly from those in the forward-looking statements. Further information regarding these and other risks, uncertainties, or factors are included in the company's filings with the U.S. SEC.

The company does not undertake any obligation to update any forward-looking statement as a result of new information, future events, or otherwise, except as required under applicable law. Finally, please note that unless otherwise stated, all financial figures mentioned during this conference call are denominated in RMB. It's now my pleasure to introduce our Vice Chairman and CEO, Mr. Zou. Please go ahead, Zou.

Zou Tao (Vice Chairman and CEO)

大家好，欢迎参加金山云2025年第三季度业绩电话会。我是金山云CEO周涛。在人工智能进入千方百业、重塑技术格局的时代变革中，金山云坚定战略定位，明确发展方向。在稳健满足自算训练需求的基础上，做好推理需求爆发的技术和资源储备。面对模型快速迭代与自算规模化应用的双重趋势，我们为客户提供了稳定高效的训推一体自算服务，并布局模型API业务，将推理场景打造为新的增长引擎。持续的收入高增长和稳健的利润率水平，印证了我们战略举措的稳步落实，实现高质量可持续的良性发展态势。首先，我们三季度收入达到24.8亿元人民币，同比增长率由上季度的24%再次提速到31%。公有云和行业云均实现了同环比的增长，其中公有云同比大幅增长49%，收入实现17.5亿元人民币。其次，自算云业务保持快车道发展。本季度，自算云上单收入达7.8亿元人民币，同比增长近120%。占公有云收入比例达到了45%，相较去年同期的31%，大幅提升。自算和云服务协同共生，从技术、产品到客户销售等方面都深度融合。自算需求不仅仅带动了自算云的快速发展，同时带动了基础公有云的需求增长与技术的创新，驱动了云计算技术的加速迭代。从自算训练集群到自算原生解决方案，我们的算力服务、模型API服务、存储服务、数据服务都进行了升级。第三，小米金山生态根基牢固。本季度，来自小米金山生态的收入达6.7亿元人民币，同比增长84%。占总收入比例进一步提升至28%。2025年1至9月，来自小米金山生态的收入合计达18.2亿元。我们预计今年较为充分地完成关联交易额度下的业务合作，并对明年额度的进一步提升充满信心。最后，本季度公司经调整毛利润率达到3.9亿元人民币，同比增长28%。经调整经营利润率扭亏为盈，达到1,536万元人民币。经调整经营利润率0.6%。经调整净利润，历史上首次实现盈利达2,873万元。公司兼顾收入增长和盈利能力提升，规模效应日渐显著。在加速自算基础设施和技术能力建设的同时，强化成本与费用的管控。

Hello everyone, thank you and welcome to Kingsoft Cloud's third quarter 2025 earnings call. I am Zou Tao, CEO of Kingsoft Cloud. In the era that artificial intelligence is implemented across various industry verticals and reshaping the technological landscape, Kingsoft Cloud has firmly established its strategic positioning and defined its development orientation. On the premise of steadily meeting the demands of model training, we have made adequate technical and resource reserves for the explosive growth of inference. In the face of the dual trends of rapid model iteration and increasing adoption of artificial intelligence, we have provided our clients with stable and efficient integrated training and inference intelligent cloud computing services and have laid out model API business to turn inference scenarios into new growth engines.

The substantial high growth in revenue and the stable profit margin level validates the steady execution of our strategic measures, achieving high quality and sustainable development. First, our revenue in the third quarter reached RMB 2.48 billion, with year-over-year growth rate accelerating from 24% in the previous quarter to 31% this quarter. Both public cloud and enterprise cloud achieved year-over-year and sequential growth, among which public cloud revenue increased significantly by 49% year-over-year, reaching RMB 1.75 billion. Second, intelligent computing cloud business remains on a fast development track. This quarter, gross billings of intelligent computing reached RMB 782 million, with a year-over-year growth around 122%. It accounted for 45% of the public cloud revenue, realizing a significant increase from 31% in the same period last year. Generative artificial intelligence and cloud are symbiotically integrated in many aspects, including technology, products, and customer cross-sales.

The demand for artificial intelligence not only drives the rapid development of intelligent computing cloud, but also leads to the growth and technological innovation of basic public cloud and accelerates the iterative process of cloud computing technologies. From training clusters to native technologies, our computing power services, model API services, storage services, and data services have all been upgraded. Third, the Xiaomi and Kingsoft ecosystem continued to offer a solid foundation. This quarter, revenue from the Xiaomi and Kingsoft ecosystem reached RMB 691 million, increasing by 84% year-over-year, and its proportion in the total revenue further rose to 28%. From January to September 2025, the total revenue from the Xiaomi and Kingsoft ecosystem reached RMB 1.82 billion. We anticipate adequately fulfilling the business cooperation under the continuing connected transactions annual quota this year and are optimistic in the further increase of the quota next year.

Finally, our adjusted gross profit for this quarter reached RMB 393 million, representing a year-over-year increase of 28%. The adjusted operating profit turned from loss to profit, reaching RMB 15.36 million, and the adjusted operating profit margin was 0.6%. The adjusted net profit recorded a historical positive profit of RMB 28.73 million for the first time. The company is aiming at both revenue growth and profitability improvements as the economies of scale are becoming increasingly prominent. While accelerating the construction of intelligent computing infrastructure and technological capabilities, we are also strengthening the control of costs and expenses.

下面我向大家具体介绍2025年第三季度的业绩情况。公有云方面，本季度实现收入17.5亿元人民币，同比增长49%。自算云业务保持强劲的增长，我们成功地支持了多家互联网行业头部客户的大规模训练和推理需求，为客户提供高质量、高性能、高稳定、高效率的云计算服务。特别是诸多生成式人工智能和互联网企业，在模型训练和推理业务兼具需求的情况下，我们为客户提供了稳定的训推一体自算服务。同时，我们在客户拓新以及自算云基础公有云协同售卖方面积极拓展。生态客户方面，我们持续为小米和金山提供优质服务，并继续为生态客户新增底层资源，夯实自算需求的快速响应能力。行业云方面，本季度实现收入7.3亿元人民币。我们坚信，在人工智能产业快速迭代的今天，智能化必将从模型能力向行业解决方案演化，赋能和重塑千方百业。而云作为智能化不可或缺的载体，数字化赋能千方百业的天地广阔，大有可为。在这样万亿级持续扩张的市场，金山云深度挖掘ToB企业服务的基因能力传承，精选优势垂直行业和地域区域，打造面向未来的核心竞争力，赢得了客户与市场的广泛好评。举例来说，公共服务领域，我们旨在成为政府行业自算推理云的首选云伙伴。以甘肃沁阳为例，作为国家东述西算八大节点之一，自算业务的聚集地，我们将负责建设甘肃沁阳政务云平台，全面赋能当地政务的智能化、数字化。数字健康领域，我们实现了里程碑式的人工智能+中医临床场景项目突破，不仅实现了中医理论与人工智能的深度融合，抢占慢病管理技术的制高点，更在临床层面验证了智能化在提升患者生活质量和疾病控制率的实际价值。企业服务领域，在银行授信报告智能生成的标杆性项目落地后，我们继续推行从单一的授信报告发起到授信全流程的智能化转型，打造从客户准入、授信报告生成到贷款发放、监控预警与贷后报告的智能体系。我们坚信，这些已沉淀的成功经验、市场口碑及可复用的核心解决方案，将使我们在这一产业浪潮中占得先机，构筑坚实的核心竞争力，实现高质量、可持续的沟通回报。

Now I would like to walk you through the key business highlights for the third quarter of 2025. In terms of public cloud services, revenue reached RMB 1.75 billion in this quarter, making a year-over-year increase of 49%. The intelligent computing cloud business has maintained strong growth. We have successfully supported the large-scale training and inference demands of various top internet customers, providing high-quality, high-performance, high-stable, and highly efficient cloud computing services. Especially for many artificial intelligence and internet enterprises, facing the simultaneous demands for model training and inference, we have provided customers with stable and integrated intelligent computing services for different scenarios. Meanwhile, we actively expanded customer coverage and the cross-selling of intelligent computing cloud and basic cloud.

In terms of ecosystem customers, we continued to provide high-quality services to Xiaomi and Kingsoft and continue to prepare underlying resources for ecosystem customers to enhance the rapid expansion capability for intelligent computing demands. In terms of enterprise cloud services, revenue in the quarter was RMB 730 million. We firmly believe that in today's rapidly evolving generative artificial intelligence landscape, intelligence will evolve from model capabilities to industry solutions, empowering and reshaping diverse sectors of the economy. As the indispensable carrier for intelligent computing, cloud services enjoy tremendous potential for such digitalization and intelligentization. In this trillion-dollar sustainably expanding market, we have deeply explored our inherent DNA of ToB enterprise services, targeted advantageous selected verticals and geographical regions, and built core competitiveness for the future. As a result, it has received widespread recognition from our customers and the broader market.

For example, in the public services sector, we aim to become the preferred cloud partner for intelligent computing in the public services agencies and enterprises for their inference demands. Taking Qingyang City in Gansu Province as an example, as one of the eight major nodes of the national project East Data Web Computing and a central area for intelligent computing business, we will be responsible for building the public services cloud platform in Qingyang to fully empower the local public services affairs with intelligence and digitalization.

In the field of healthcare, we have achieved a milestone breakthrough in a project integrating artificial intelligence with traditional Chinese medicine clinical scenarios, whereby not only have we achieved a deep integration of traditional Chinese medicine theory and artificial intelligence, seizing the commanding position in chronic disease management technology, but we have also verified the practical value of artificial intelligence in improving patients' quality of life and disease control rate at the clinical level. In the enterprise services sector, following the successful implementation of a landmark project for intelligent generation of bank credit reports, we continue to advance the intelligentization transformation across the entire credit approval process. This evolution extends from the single function of credit report initiation to a comprehensive intelligence system, including customer onboarding, credit report generation, loan disbursement, monitoring and early warning, and post-loan reporting.

We firmly believe that these proven accumulated successful experiences, market reputation, and replicable core solutions will enable us to seize a pioneering position in the emerging industry wave, build a solid core competitiveness, and achieve high-quality and sustainable shareholder returns.

产品技术方面，公有云领域，本季度自算云持续强化新流平台能力，在以下三个方面取得了重要进展。首先，我们发布了模型API服务，提供高可用一级层的模型调用与管理能力，为后续提供多样化的模型服务模式打下了扎实的基础。其次，模型在线服务升级，整合多款开源模型，具备自动扩缩容能力，为推理服务提供高可用平台底座。第三，我们上线的数据标注及数据集广场，旨在为客户提供数据流转全流程的支撑，助力客户高效推进模型训练进程。行业云领域，我们基于私有化部署场景，建设了算力调度平台、轻量化MaaS平台和生成式人工智能知识库等，并紧密与WPS AI政务办公应用相结合，打造行业云客户可信赖的智能产品架构。同时，我们将通过北京、武汉双研发中心的组织建设，吸引各地优秀人才，进行人才梯队建设，持续保持在自算领域的投入强度。截至三季度末，武汉员工数量已达22年武汉战略发布之初的2.8倍。总体而言，我们将坚定以小米金山生态带来的历史性机遇为抓手，继续投资基础设施，专注于打磨核心产品及解决方案能力，为我们的客户、股东、员工和其他利益相关方持续创造价值。接下来有请我们的CFO李毅女士为大家介绍三季度财务业绩，谢谢。

In terms of product and technology, in public cloud space, we continue to enhance the technology of intelligent computing cloud this quarter, strengthening the capabilities of the StarFlow platform and made significant progress in the following three aspects. First, we have launched our model API service, delivering highly available and easily integrable capabilities for model invocation and management, laying a solid foundation for the subsequent provision of diverse model service paradigms. Second, we upgraded our online model services, integrating multiple open-source foundation models and equipped with automatic scaling capabilities, offering a highly available inference level. Third, we launched our data annotation and dataset marketplace, aiming to provide customers with end-to-end support for data flow and help them efficiently advance the model training process.

In enterprise cloud space, in order to meet the demands for private deployment scenarios, we have built a computing power scheduling platform, a lightweight mass platform, and a generative artificial intelligence knowledge base. We have closely collaborated with WPS AI to build a trusted intelligent product architecture for public services use cases. Meanwhile, through the organizational development of the dual R&D centers in Beijing and Wuhan, we attract talents from various regions, build a talent pipeline, and maintain sustained investment intensity in the intelligent computing field. As of the end of Q3, the number of employees in Wuhan is 2.8 times the headcount back in 2022 when we first launched our Wuhan strategy.

Overall, we will firmly seize the historic opportunities presented by the Xiaomi and Kingsoft ecosystem, continue to invest in infrastructure, focus on refining core products and solutions, and to create long-term value for our customers, shareholders, employees, and other stakeholders. I will now pass the call over to Ms. Li Yi, our CFO, to go over our financials for the third quarter of 2025. Thank you.

Li Yi (CFO)

Thank you, Zou Tao, and Jingjing Lang. Good evening and good morning, everyone. Thank you all for joining the call today. Before we walk through the details of financial results for the third quarter, I would like to highlight the following aspects. First, revenue has consecutively achieved year-over-year growth for six quarters, reaching RMB 2,478 million this quarter. This represents an accelerated year-over-year growth rate of 31%, up from 24% in the previous quarter. Revenue from public cloud service sold at RMB 1,752.3 million, a significant increase of 49% from RMB 1,165.5 million in the same quarter last year. Meanwhile, robust demand for our intelligent cloud, which also called as AI Cloud Business, drove around 120% year-over-year billing growth, which totaled RMB 782.4 million. Second, profitability has seen substantial improvement.

Our adjusted gross margin rose to 16%, up from 50% in the previous quarter, and adjusted EBITDA margin improved to 33%, compared with 17% last quarter. Notably, we turned quarterly adjusted operating loss and adjusted net loss into profit simultaneously for the first time. This gain validates our strong execution in pursuing high-quality, sustainable development, as well as our ability to monetize opportunities in the intelligent cloud space. Third, we would like to express our gratitude to shareholders for their support during our risk equity financing in September. We successfully raised HKD 2.8 billion, and 8% of the fund will be allocated to further investment in AI infrastructure and 20% to general operational needs. This funding will fully underpin the growth of our intelligent cloud business and enable us to create long-term value for all stakeholders.

Now, I will walk you through our financial results for the third quarter of 2025 and use the RMB as currency. Total revenues were RMB 2,478 million. Of these, revenues from public cloud services were RMB 1,752.3 million, up 49% from RMB 1,175.5 million in the same quarter last year. Revenues from enterprise cloud services reached RMB 725.7 million, compared with RMB 710 million in the same quarter last year. Total cost of revenues were RMB 2,097.1 million, up 33% year-over-year, which was mainly due to our investment into infrastructure to support intelligent cloud business growth. IBC cost increased by 15% year-over-year, from RMB 673.8 million to RMB 775.7 million this quarter. The increase was mainly due to the purchase of racks, which serve the expanding intelligent cloud business, as well as the basic computing and storage cloud demand brought by AI development.

Depreciation and amortization costs increased from RMB 297.5 million in the same quarter of 2024 to RMB 649.7 million this quarter. The increase was mainly due to the depreciation of newly acquired and leased servers and late work equipment, which were mainly allocated to intelligent cloud business. Solution development and services cost increased by 90% year-over-year, from RMB 499 million in the same quarter of 2024 to RMB 595.9 million this quarter. The increase was mainly due to the solution personnel expansion. Fulfillment costs and other costs were RMB 5.2 million and RMB 70.6 million this quarter. Our adjusted gross margin for the quarter was RMB 392.6 million, increased by 28% year-over-year and 12% quarter-over-quarter. It was mainly due to the expansion of our revenue scale, the inadequate distribution for intelligent cloud, and the cost control of IBC racks and servers. Adjusted gross margin increased from 50% last quarter to 16% in this quarter.

On the expense side, excluding share price compensation costs, our total adjusted operating expenses were RMB 420.9 million, decreased by 70% overall year and 25% quarter-over-quarter, of which our adjusted R&D expenses were RMB 188.4 million, decreased by 90% from same quarter last year. The decrease was mainly due to the decrease of personnel costs resulting by our strategic adjustment for research team, as well as the expense serving to from Beijing Wuhan dual research same strategy. Adjusted selling and marketing expenses were RMB 127.6 million, increased by 50% year-over-year. Adjusted general and administrative expenses were RMB 104.9 million, decreased by 29% year-over-year due to the reverse of credit loss. The impairment of long-range assets was near this quarter, compared with RMB 190.7 million in the same quarter last year. Our adjusted operating profit was RMB 15.4 million, totaling profit from adjusted operating loss of RMB 140.2 million in the same period last year.

The improvement was mainly due to the expansion of revenue scale and gross profit, the expense control, as well as the reverse of credit loss. Adjusted operating profit margin increased from minus 7% in the same period last year to 0.6% this quarter, representing an increase of 8 percentage points. Our long gap EBITDA profit was RMB 826.6 million, increased by 3.5 times of RMB 185.4 million in the same quarter last year. Our long gap EBITDA margin achieved 33%, compared with 10% in the same quarter last year. It was mainly due to our strong commitment to intelligent cloud development, strategic adjustment of business structure, strict control of costs and expenses, as well as the long recovery impact of subsidy in other income. As of September 30, 2025, our cash and cash equivalent totaled RMB 3,954.5 million, decreased from RMB 5,464.1 million as of June 30, 2025.

The decrease was mainly due to our infrastructure investment for intelligent cloud. This quarter, our capital expenditures, including those financed by third parties and the right-of-use assets obtained in 2024 finance lease liabilities, were RMB 2,787.8 million. Looking forward, AI technology drives the revolution of cloud computing. We do more than just fulfill the computing demands of model training, inference. We also empower enterprises to invoke an API and apply AI capabilities to their business. Stepping into the phase of rapid development in AI applications and explosive growth in demand, we will further invest into infrastructure, strengthen technology, enhance service stability, and provide customers with high-value-added cloud services. That's all for the introduction of our operational and financial results. Thank you all.

Nicole Shan (Director of Investor Relations)

Thank you, Operator Wu. I now want you to start a Q&A session. Please ask your question in both Mandarin, Chinese, and English if possible. Operator, please go ahead.

Operator (participant)

Thank you. Thank you. As a reminder, to ask a question, you will need to press Star 1 and 1 on your telephone and wait for your name to be announced. To withdraw your question, please press Star 1 and 1 again. Please stand by while we compile the Q&A queue. Our first question comes from the line of Xiaodan Zhang from CICC. Please go ahead. Your line is open.

Xiaodan Zhang (Analyst)

Hey, Zou Tao, Li Yi, Zhong, Nicole, Clarke, 晚上好。感谢接受我的提问。我这边有两个问题。那第一个是我们看到三季度这个AI的收入是取得了非常亮眼的增长。那想请管理层拆解一下就是背后主要的一个驱动因素。那我们的这个内外部的客户的需求是否有出现一些结构性的变化。第二个问题是我们看到今年以来就是由于我们这个算力资源获取模式的变化，虽然毛利率有所下降，但是EBITDA的利润率的这个改善的幅度是非常超预期的。那管理层对于后续季度的这个利润率的趋势如何的展望，那预计就是后续不同的这个资源获取模式大概是一个怎么样的结构的比例。那我快速翻译一下我的问题。 Thank you, management, for taking my questions. First of all, what are the key drivers of AI revenue growth in Q3? Has there been any structural change in the demand of your ecosystem and external clients for the past quarter? Secondly, how does management see the margin trend in the coming quarters? What is the expected mix of different computing resources acquisition models? Thank you.

Operator (participant)

Please stand by while the speakers reconnect. Please stand by while the speakers reconnect. Please stand by. Once again, please stand by while the speakers reconnect. Thank you. Speakers, you are now reconnected. Please go ahead.

Nicole Shan (Director of Investor Relations)

Yes. Sorry, Xiaodan, we didn't get your question. Could you repeat that again? Thank you.

Xiaodan Zhang (Analyst)

Yes, no problem. My first question is regarding the AI revenue. Could management break down the key drivers for AI revenue in Q3? Has there been any structural change in the demand of your ecosystem and external clients for the past quarter? Secondly, how does management see the margin trend in the coming quarters? What's the expected mix of different computing resources acquisition models going onwards? Thank you.

Zou Tao (Vice Chairman and CEO)

我先回答，你再补充一下。我先来回答一下，然后那个到时候就请你的他补充一下。我觉得主要的增长这个因素啊，就是因为我们呃在二三季度逐步啊这个完成了这个之前的这个订单的交付吧，或者集群的交付，然后呢进入到全量计费的这么一个阶段。这是一个那个最核心的。当然还有一部分就是从那个递延过来的部分递延的收入在在这个呃三季度呢也可以理解成几乎变成全量了。可能在二季度是部分交付，部分计费到三季度已经全部交付，全部计费了。那这是最最核心的这个增长趋势。

Basically, the core of the reason behind the AI revenue growth in Q3 is that we had some clusters that, you know, partially delivered in the previous quarters, for example, like the second quarter of 2025. These clusters and these services have only been partially accounted for revenues from a full quarter basis. Now in Q3, they are starting to be recognized as full quarter revenues. Also, there is the factor of partially delayed revenue as well. Some of the revenue which we had in Q2 but was not accounted for, and then this revenue got delayed into the third quarter. Yeah.

肖丹刚才因为我们断线了，我没有听到你另外一个问题。

Xiaodan Zhang (Analyst)

哦，对，那个我第一个关于那个AI的问题，还有一个是想请教一下，就是我们看到三季度我们内外部的这个客户是否有，就是客户的需求是否有出现一些结构性的变化。然后我第二个问题是我们看到今年以来因为这个算力资源获取模式的变化，虽然毛利率是有些下降的，但是我们的这个EBITDA利润率改善的幅度是非常超预期的。就也想请就是周总分享一下，就是对于后续季度这个利润率趋势的一个展望，然后以及后续就是我们不同的这个算力资源获取的模式大概会是一个怎么样的结构和比例。谢谢。

Zou Tao (Vice Chairman and CEO)

嗯，我先回答等下刘涛和那个李毅做些补充。嗯，内外部客户需求的变化是这样啊，就是我我上个季季度也谈了一下，就是嗯从这个大趋势上看啊，嗯确实就是呃从这个从从我我之前用的话术啊叫做呃那个那个大客户训练逐步向这个普客推理呃这个转变。实际上就是现在我们主要还是在训练这个这个呃建设方面啊，现在逐渐主要是围绕几个大客去去去去去展开资源的部署吧。嗯同时呢那个我们也呃通过这一个Q吧，我们也明显确实是感受到了这种这个呃应该说是模型真正应用到无论是公允还那个还是韩语啊，这种这种进入到千行百业的这种趋势是越来越明显的。实际上我们在9月份也发布了新的平台啊，其实也是为了呃去迎接呃下一阶段的这个呃呃这个这个呃AI应用的这个到来吧。呃这个问题同时呢也也能够去呼应一下你刚才讲的这个毛利水平的变化啊。实际上呃在上个Q我们也谈过，呃随着我们的规模变大啊，反正反正集中在几个大客，所以呢相较我们最早做资产业务的这个毛利水平略有下降，所以我们也在积极的部署呃更高毛利的。说的更直白点，我们从目前的这个发展趋势来看，我们认为呃未来整个推理应用啊，这个毛利水平会会显著显著高于目前的这个训练阶段。对，好吧，这是一个大的一个趋势，好吧。然后具体的看看呃刘涛李毅有什么补充吗？

So in regarding the second part of your first question, which is about the structure of internal and external customers, I think I used to say that from a large trend, general trend perspective, we're currently in the phase of transitioning from large and top customers' training demand to general and wider spread customers' inference demand. Most at the current stage, we still see, you know, the majority of our demand coming from the larger customers in their training demand. However, especially in the latest quarter, we're increasingly seeing the trend of customers adopting artificial intelligence models into their diverse industries. In face of this general trend, we have also, as we mentioned in the prepared remarks, launched our StarFlow platform to meet the demands of such general trend. This also goes back to the margin question that you also asked about.

We generally think that in the future, the inference demand will tend to exhibit higher margin profile than the current stage of training. Therefore, we think that when that wave of demand comes, we expect to have higher margins.

Li Yi (CFO)

Thank you, Xiaodan. At the EBIT level, as the proportion of the AI business continues to rise and its cost structure is mainly dominated by depreciation, we expect this year's EBITDA margin will still remain above 20%. I have to mention that the significant quote unquote improvement this quarter was mainly driven by a one-time other income, which will return to the normal level next quarter. Thank you, Xiaodan.

Nicole Shan (Director of Investor Relations)

Operator, next question, please.

Xiaodan Zhang (Analyst)

Thank you.

Operator (participant)

Now, next question comes from the line of Wenting Yu from CLSA. Please go ahead. Your line is open.

Wenting Yu (Analyst)

Hey, Zou Tao, Li Yi, Zhong, Clarke, Nico, 晚上好。感谢给我提问的机会。我这边有两个问题想请教。那第一个是能否请管理层分享一下明年的这个收入的指引和增速的一个预期。在今年已经铺开的这个互联网厂商模型的后训练以及巨声智能这些应用场景的基础上，那我们还预期未来在哪一些行业跟场景会出现强劲的这个算力的需求，从而去推动我们收入的进一步增长。那第二个问题是当前我们看到国内外多家云厂商在算力资源配置里面都提高了服务器租赁的这个比例。那结合市场上采购跟租赁这两块的市场情况，那我们在性价比跟利润率的角度会怎么去考虑优化这两种方式的这个分配。那我很快翻译一下我的问题。 The first question is, could management share the outlook and guidance on the revenue outlook for next year and beyond the internet complex post-model training and embodied intelligence scenarios that are already underway this year? Which other industry and application scenarios are expected to have strong computing power demand that could drive the revenue growth next year? The second question is, with multiple cloud providers in both China and US increasing the proportion of server leasing in their computing resource mix, how does management view the current market dynamics for procurement versus leasing? From a cost effectiveness and profit margin perspective, how will the company allocate the resources between these two approaches? Thank you.

Li Yi (CFO)

Wenting, thank you for your question. The company's budget process is currently underway and expected to be completed around the beginning of the next year. We will share specific details with you once it is finalized. However, regarding the demand for our AI business, we are fully confident in the subsequent demand growth. For your second question regarding the procurement method, we primarily align our capital channels with actual customer needs, including cluster scale, delivery time, and the supply inventory levels. There is no rigid top-down allocation target. From the cost effectiveness perspective, both approaches have their own pros and cons. The leasing model is fine to our supply chain channels and provides a certain degree of flexibility in resource allocation, with the flexibility also offered through short-term and long-term contracts. Self-procurement, on the other hand, gives us great autonomy in controlling delivery timelines and managing clusters.

It also reduces the profit sharing with suppliers, thereby elevating our appreciation on profit margin. Thank you, Wenting.

Zou Tao (Vice Chairman and CEO)

要不你先回答一下咱们那个未来哪些行业。Yeah, you know, you know, as you mentioned that, you know, the robotic companies in China is a growth in very fast pace. You know, as you this year, we have covered most of the robot companies in China, and we can see the revenue is increasing very rapidly. In the next year, we believe the increase of the robotic companies will also be fast. Meanwhile, you know, as more and more internet companies in China are using token token services, which is the API services, we are seeing the increase in the business is very quickly. We believe in the next year, this will be a very important factor to driving the revenue to increase. Thank you.

我再补充一下关于你谈到的后面那个问题啊，就是我的理解就是我们未未来怎么选择是自采的服务模式还是通过服务器租赁的模式。我在上个Q基本上讲过一个总原则，我们针对我们那些大客，包括像小米之类的，相对来讲公司现金流充裕，这个整个公司这个这个这个发展态势非常良好的，或者简单讲其实就是可靠的客户吧，我们会会采取一些自己支付Capex这种方式，但对于其他的一些可能有一定风险，或者说在在这个整个发展的这个还是在这个发展过程当中的这一一些，我们基本上逐渐逐渐会倾向于这种租赁的方式来提供服务，这样也是进一步降低我们自身的这个风险，好吧。当然没有一个具体的怎么样的配比，我觉得只能是更多的是从客户角度，对对像小米这类客户我们觉得优质可靠的呢，那我们就会通过这个自己投Capex这种方式来服务，例如一些一些中小的企业呢，我们还是会采取这个资源租赁的这种方式。所以呢这两个呢确实就是从我们上个Q我也谈过，对于我们总体的毛利水平呢会有一定的影响，但是经过这今年应该说这个三个季度跑下来看吧，我觉得基本上尤其包括我们这个Q3的总体的毛利水平比Q2还是有一定的提升，所以我上次也谈到过，我说基本上可能会保持在这么一个水平，等等未来呢主要是看我们的这个推理业务的这个进展情况，总体来讲这个毛利水平有望进一步改善，好吧。

This is the CEO, CEO Zou Tao, and he added that you understand that your question, your second question is really about the choice between the leasing model and the Capex model. We've talked about that before. Generally, there's a general rule of thumb. When we're looking at the larger customers, especially the customers that have solid profile, have solid fundamentals, and are trustworthy, premium customers, for example, like Xiaomi, we would tend to choose the Capex model. While in other growth stage companies, medium and small size, the small and medium size of the companies, we generally tend to adopt the leasing model, which is also a meaningful way to reduce our own risk. As Li Yi rightly mentioned, there's no this kind of like top-down target for the split between these two different methods.

We also talked about in the last quarter as well that the impact of these two different methods have different impacts to gross margins. However, as we have seen the financial results for the past three quarters, which we have adopted various combinations of these two different models, and you know especially when you compare the growth margin for the third quarter versus the second quarter, it actually also improved sequentially. I would say that at the current stage we do not expect material changes to the current status, but generally speaking, in the future we do expect the margin to improve. Thank you.

Nicole Shan (Director of Investor Relations)

Thanks, Wenting. Next question, please, Operator.

Operator (participant)

Thank you. Our next question comes from the line of Timothy Zhao from Goldman Sachs. Please go ahead. Your line is open.

Timothy Zhao (Analyst)

好的，感谢关宇成接受我的提问。那我的问题还是关于这个训练和推理的。那想请教一下我们对于整个AI训练和推理的这个定价方式有什么不同？那过去的几个月或者是今年以来，整个的AI的服务相关的定价的价格有哪些特别的变化？那我们在推理和训练的卡的这个利用率，价格，包括利用率上面，包括利润率上面的这个差别，大概是有多少？这个是我的问题。那我很快翻译一下。 Thank you Madren for taking my question. My question is regarding the differences between AI training versus inferences. Could Madren share what is the pricing methodology between these two kinds of demand and what has been the pricing trend over the past few months or year to date? In terms of the overall utilization rate of the chips of GPUs, pricing and profitability, can you share more color on the gap between training and inferences? Thank you.

Zou Tao (Vice Chairman and CEO)

Okay, let me answer this question. You know, when we're talking about the price strategy for inference and the training, you know there's not too much difference between the two things. The price is based on the qualities, how many sources use, which is the most important factors. Also, comparing the margin rate, you know there are two kinds of inference services. One is, you know, customers buy resource and use our platform to inference. That margin ratio is very similar to the training margin ratios. Another one is, you know, customers do directly buy our API token services. We think that will have a better margin ratio. You know, this business just in the beginning.

We need time to see what is the big difference between the two things. Thank you.

Nicole Shan (Director of Investor Relations)

Thanks, Timothy. Operator, please.

Operator (participant)

Thank you. Due to time constraints, this concludes our question and answer session. I'll hand the call back to Nicole for closing remarks.

Nicole Shan (Director of Investor Relations)

Thank you. Thank you all once again for joining us today. If you have any questions, feel free to contact us. Look forward to speaking with you again next quarter. Have a nice day. Bye-bye.