--- title: "Interview with KINGSOFT CLOUD's Liu Tao: Capturing the Thickest Slice of the Intelligent Computing Second Half" type: "News" locale: "zh-HK" url: "https://longbridge.com/zh-HK/news/280613652.md" description: "Engineering breakthrough" datetime: "2026-03-26T10:47:12.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/280613652.md) - [en](https://longbridge.com/en/news/280613652.md) - [zh-HK](https://longbridge.com/zh-HK/news/280613652.md) --- > 支持的語言: [简体中文](https://longbridge.com/zh-CN/news/280613652.md) | [English](https://longbridge.com/en/news/280613652.md) # Interview with KINGSOFT CLOUD's Liu Tao: Capturing the Thickest Slice of the Intelligent Computing Second Half Author | Zhou Zhiyu As the global market, reeling from the aftershocks of Amazon and Google's financial reports, repeatedly questions whether AI investment has "peaked" and when free cash flow will return, this arms race is entering a grim elimination phase. Few are destined to win this race; only those who establish a commercial closed loop on the critical line of output-to-input ratio can gain investor favor. KINGSOFT CLOUD, having experienced a "V-shaped reversal," understands this deeply. From a low of $2 per share in September 2024 to over $22 per share at its peak, KINGSOFT CLOUD's value is being re-evaluated by the market. In the fourth quarter of 2025, KINGSOFT CLOUD delivered a crucial performance report: revenue reached a new historical high of 2.76 billion yuan, with intelligent computing cloud revenue climbing again to 930 million yuan, accounting for 49% of public cloud revenue. In the fourth quarter, KINGSOFT CLOUD once again achieved positive adjusted operating profit. Compared to previous losses that often ran into hundreds of millions, this marks the company's finally knocking on the door of profitability in the "intelligent computing era" after a painful three-year transformation. Liu Tao, Senior Vice President of KINGSOFT CLOUD, a veteran who has been in the cloud industry for over a decade, recently sat down with Wallstreet News for an exclusive interview to review this perilous transformation journey. "We experienced the extreme pain of falling from the highest stock price," Liu Tao said frankly. From a peak of nearly $75 per share at the end of 2020 to a low of $1.77 per share in October 2022, the cash value held by KINGSOFT CLOUD was once higher than its market capitalization. The capital market's conclusion was cold and decisive: the market predicted they would quickly run out of funds. This was not just KINGSOFT CLOUD's predicament, but a "scale trap" that the entire Chinese cloud industry collectively fell into over the past decade. In that era, the drivers of valuation came from traffic and revenue scale, with manufacturers fiercely competing in the red ocean of infrastructure like CDN (Content Delivery Network). KINGSOFT CLOUD also once fell into a vicious cycle of "the more we invest, the more we lose." "This logic was a lingering shadow for the industry," Liu Tao stated, but the essence of a company is to make money for its shareholders, and this old development model came to an end in 2022. Liu Tao revealed that his mission when he took over the public cloud team in early 2022 was extremely pure: do not engage in money-losing business. In his first year in charge, he cleared out all customers with negative gross profit, which led to significant fluctuations in both revenue and gross profit for the public cloud business. However, by 2025, KINGSOFT CLOUD's public cloud business had seen steady growth. The most critical chapter in KINGSOFT CLOUD's transformation story is a gamble on "faith." Facing the computing power tsunami triggered by ChatGPT, in April 2023, Liu Tao and the management team faced a decision that could determine the company's fate: whether to invest hundreds of millions to build a large-scale intelligent computing cluster with extremely tight cash flow? "This was a decision that could have dragged the company into a desperate situation," Liu Tao recalled, as the huge depreciation would have been unbearable for the company at the time if the delivery failed. But CEO Zou Tao set the tone with a single sentence: "We must have faith in AI; it's like the advent of the steam engine era." Ultimately, KINGSOFT CLOUD, in a field without any precedent, managed to conquer the engineering challenge of large-scale cluster networking. Looking back, KINGSOFT CLOUD's management made the right decision. Currently, US and Chinese tech giants continue to pour huge sums into computing power, but demanding investors are also starting to measure the effectiveness of every expenditure with ROI. This means that while the computing power arms race continues, intelligent computing cloud is gradually shifting towards the second half, with "operational efficiency" at its core. From KINGSOFT CLOUD's perspective, intelligent computing cloud is no longer a simple trade of resources but an extreme engineering challenge. Large-scale intelligent computing clusters are extremely fragile, with optical module failure rates as high as 2‰ and machine failure rates at 1%-2%. For large-scale clusters, any single point of fluctuation can trigger a chain reaction, leading to the interruption of overall training tasks. To cope with such losses, KINGSOFT CLOUD has reshaped stability through a series of hardcore engineering measures. KINGSOFT CLOUD has built a fault-aware self-healing system that, relying on containerized scheduling and hot-swapping mechanisms, compresses training task recovery time to within 20 minutes, ensuring an actual cluster operational SLA of over 99.9%; through PD separation (pre-filling and decoding separation) and KV Cache reuse technology, it achieves response latency stability in office scenario services that is superior to the original manufacturer's solution, effectively increasing profit margins on the inference side; the resource management model has been upgraded to be "task delivery" oriented, solving the pain point of training task interruptions in the intelligent computing era and significantly improving the effective utilization of underlying hardware. Looking ahead to 2026, Liu Tao believes the true watershed will be "applications." Compared to the uncertain free-for-all of general large models, robotics and autonomous driving offer clearer scenario-based ROI. The practical application of video generation and the generalization of VLA models in robotics will lead to exponential growth in inference demand. Liu Tao has keenly observed that most robotics companies are proficient in physical commands but are almost bare-handed when it comes to cloud engineering, security, and data loops. KINGSOFT CLOUD's strategy is to become the underlying "circulatory system" for this trillion-dollar market: providing leading companies with a full lifecycle closed loop from data production and storage to model training and simulation; building end-to-end security defenses against extreme risks like robots being hacked, preventing malicious hijacking of AI "brains"; and achieving rapid and accurate responses in robotics, akin to biological intuition, through extreme inference optimization. "Securing the high ground in the robotics scenario means obtaining the ultimate ticket to reshaping the intelligent rules of the physical world," Liu Tao predicts that in the next three to five years, robots will evolve from simple actions like picking up socks and towels to gradually penetrating every household, and KINGSOFT CLOUD aims to be the cloud foundation they natively rely on from the moment these enterprises are born. "Not developing large models" is KINGSOFT CLOUD's strategic bottom line for maintaining calmness amid the intelligent computing tide. Precisely because it does not compete in model development, KINGSOFT CLOUD is able to maintain extreme openness and purity in its ecosystem. Whether it's "new model contenders" or large models within the ecosystem, they are growing steadily on KINGSOFT CLOUD's intelligent computing foundation. Liu Tao believes that with the exponential explosion of inference-side demand, Tokens will become a universal resource similar to water and electricity, and the winner in the cloud vendor race will be whoever can provide the most stable service with the lowest engineering cost. The golden age of intelligent computing cloud will not last forever; only those vendors who can solve extreme engineering challenges and establish industry data loops will remain at the forefront after the tide recedes. From a stock price low of $1.7, to intelligent computing business now accounting for over one-third and driving the entire company towards break-even, KINGSOFT CLOUD's "V-shaped reversal" is not accidental. It is essentially a mature enterprise's resolute return to operational efficiency and business fundamentals after a deep understanding of industrial depth. This race for productivity dominance has just entered its most brutal, and most exciting, deep water zone. Liu Tao and KINGSOFT CLOUD are attempting to prove that a cloud service provider, by sufficiently respecting financial principles, being willing to delve into engineering details, and maintaining foresight regarding technological trends, has the opportunity to transform from a resource integrator to a value co-creator in the AI wave. ## **How to Emerge from the Valley** **Wallstreet News:** You've been with KINGSOFT CLOUD for ten years, and cloud vendors have evolved from the mobile internet to PC and now to the AI era. What period has been the most memorable or agonizing for you personally? **Liu Tao:** The most unforgettable and painful period was actually the "V-shaped reversal" process. KINGSOFT CLOUD went public in 2020, and its stock price surged to $75, with a valuation once reaching over $17 billion. Lei (Lei Jun) was very happy at the time, and everyone felt the company had a bright future. By 2021 and 2022, the capital market no longer looked at cloud companies based on PS (price-to-sales ratio) but rather on PE (price-to-earnings ratio) and profit/loss. Our stock price fell from $75 to $1.7. At that time, the capital market's evaluations were very harsh, with negative comments like "insufficient team capability," "cash value higher than market cap," and "likely to quickly deplete funds." The CDN business was operating at a negative gross profit margin, meaning the company was essentially subsidizing its business to maintain scale. This extremely unhealthy operating model forced us to reflect deeply and initiate a comprehensive transformation. **Wallstreet News:** In the past two years, was there any moment or meeting that made you realize AI was truly reshaping the company's logic? **Liu Tao:** In 2021, a large model startup client required high-performance computing power even more than some leading clients. This made us realize that large model companies' demand for computing power was on a completely different scale. Later, the client proposed building a large-scale cluster, involving a huge sum of money that posed a severe test to the company's cash flow. As I come from a technical background, I am very rational in my approach. I asked Zou (Zou Tao): "If this doesn't go well, it could drag the company down. Should we do it or not?" Zou Tao's decision at the time was: "Let's do it through the night!" Zou Tao said: "You must have faith in AI; this is the arrival of the steam engine era, and the future space is unimaginable." The facts have proven the foresight of this judgment. **Wallstreet News:** What were the most direct feelings and mindset changes for the frontline teams? What were the most focused discussion points among everyone? **Liu Tao:** The mindset of the sales and technical teams has shifted from "following" to the confidence of "leading." The sales team is becoming more energized with each success, as this kind of customer benchmark effect is cumulative. It's important to remember that in the era of general-purpose computing, we had a clear gap with our competitors and were always in a passive, following state. However, with the advent of the AI era, market demand is focusing on intelligent computing networks and storage, which are precisely the areas we have deep expertise in and possess advantages. ## **It's Time to Abandon Biases Against Cloud** **Wallstreet News:** Investors believe that cloud business is not a good business. Is this view outdated? **Liu Tao:** I believe this view has misconceptions, especially in the Chinese market. Compared to other B2B businesses, cloud is one of the few sectors with high sustainability. While Chinese clients are not accustomed to paying for pure software, they are willing to pay for hardware capabilities and technical services. The core value of cloud is to maximize hardware capabilities through technological services. Currently, clients generally emphasize cost control, and enterprises are no longer willing to bear the high costs of building their own infrastructure teams. For example, when clients host their equipment with us, we only charge a certain percentage of service fees, which is far lower than what it would cost for a company to build and operate its own. Of course, if a cloud business gets bogged down in endless customized projects, it indeed becomes difficult to be a good business. However, by focusing on standardized and scalable operational projects, cloud becomes a highly valuable and premium sector. **Wallstreet News:** Compared to larger competitors, what are KINGSOFT CLOUD's greatest features and competitive strategies? **Liu Tao:** We adopt a "Tian Ji Sai Ma" (an ancient Chinese strategy of using inferior horses against superior ones in races, by pairing them strategically) competitive strategy. As a smaller company, we do not compete on sheer volume or engage in homogeneous competition. Instead, we build differentiated advantages through "cost-effectiveness" and "close-knit service," offering more competitive prices and better service. Take the robotics sector, for instance. When a client has a need, I will accompany the SA and after-sales team to the client's site for direct engagement, following up throughout the process. As long as the client raises a reasonable request, we will go all out to solve it. Clients in this sector generally have weaker engineering capabilities, and strong engineering delivery capabilities are precisely our strength, perfectly matching their pain points. This kind of "close combat" deep service has won us the recognition and trust of many leading clients in the robotics sector, something even larger competitors envy. ## **Exploring Niche Opportunities** **Wallstreet News:** What considerations go into selecting clients? **Liu Tao:** Our selection logic is very clear: focus on sectors and clients that can bring incremental growth. Traditional cloud clients' budgets are shrinking year by year; achieving flat growth is considered good. The growth in the intelligent computing sector primarily comes from large models, but relying solely on this one sector is not enough. Therefore, we have focused our efforts on autonomous driving, robotics, and AI for Science. Among these, autonomous driving is a mature sector, and our core approach is to remain patient and wait for opportunities. Robotics is a new sector. Many companies here are founded by scientists who are proficient in algorithms but lack the concept of engineering implementation, which presents an opportunity for us. In the AI for Science field, such as pharmaceutical and biotechnology companies, these companies are themselves very mature. When collaborating with them, the core is to compete on cost-effectiveness, winning recognition with more advantageous solutions. **Wallstreet News:** What specific pain points have you addressed in terms of computing power scheduling and engineering implementation for autonomous driving and robotics clients? **Liu Tao:** The core is to address stability and data loops in computing power scheduling and engineering implementation. For ten-thousand-card clusters, hardware failures are unavoidable. By building a fault perception mechanism and deploying hot standby technology, we achieve rapid self-healing after node failures. Once a node fails, training tasks can be automatically restored within 20 minutes, ensuring the cluster's actual operating SLA remains stable at over 99.9%. For office scenarios, we have jointly researched PD separation (pre-filling and decoding separation) technology with the group, significantly reducing the first-token latency (TTFT). Simultaneously, by optimizing the hit rate and pre-filling capabilities of KV Cache, we can ensure stable and smooth inference even in high-concurrency scenarios. **Wallstreet News:** Autonomous driving is also transitioning from VLA to large world models, placing immense pressure on data throughput. Are there technical challenges in this area? **Liu Tao:** The pressure is indeed very high, with core pain points concentrated in the entire process of data storage, processing, and transfer. The data acquisition threshold in the autonomous driving field is low, and the incremental data is astonishing, easily reaching EB-level. To address this, we have worked closely with clients to develop a full-process data loop platform that covers core functions such as data anonymization, video and point cloud fusion processing, and 3D Gaussian modeling. Furthermore, due to power consumption bottlenecks limiting onboard computing power, there may be a future architecture where "on-board devices are responsible for real-time decision-making, while the cloud runs macro world models." This will bring about a huge demand for cloud-based inference. **Wallstreet News:** The robotics sector has more diverse data dimensions and greater processing pressure. What core services can we provide well in this process? **Liu Tao:** Our goal is to "assist clients in focusing on core algorithms, while all other engineering and data-related tedious work is handed over to us." In data acquisition and processing loops, we provide an integrated platform for the entire process, from local data synchronization, cloud labeling, data cleaning, to final model training. Although we do not directly engage in data labeling services, we act as a bridge, integrating a series of high-quality data customization and labeling suppliers. This "technical foundation + ecosystem integration" model allows us to achieve lightweight operation of upper-layer services like AWS, while firmly locking in the stickiness of underlying computing power services. ## **The Next Few Hard Battles** **Wallstreet News:** Looking ahead to the entire year of 2026, what will be the "core battles"? **Liu Tao:** The core focus will be on three directions: "large clients, deep sectors, and Tokenization." We have three key battles: First, solidify the technological foundation to ensure that in the context of next-generation hardware iteration, the reliability and stability of large clusters remain at the industry's leading level; second, deepen sector penetration, continuously strengthening full-process closed-loop service capabilities in vertical sectors like robotics, not only meeting existing client needs but also proactively identifying and solving potential pain points such as system security vulnerabilities; third, activate new growth momentum by deeply operating the Token business. **Wallstreet News:** Besides the Token business, what other sectors do you see explosive potential in? **Liu Tao:** I have two forward-looking directions that I am quite optimistic about. AI-native games, such as "Commander Mode," where players only give commands, and AI teammates follow them; or AI companions that provide emotional value. Code generation SaaS is the most certain area of incremental growth. Enterprises, due to privacy requirements, will not use public version large models, so we need to provide enterprise-level private code generation services. We are already using it extensively internally. **Wallstreet News:** The proportion of revenue from the Xiaomi-KINGSOFT ecosystem within KINGSOFT CLOUD is continuously increasing. How do you view the "independent third-party" label? **Liu Tao:** The ecosystem is one of our advantages, but it is not our entire competitive strength. Although the intelligent computing investment from the Xiaomi-KINGSOFT ecosystem is continuously increasing, orders from external clients are growing even faster. Currently, the business proportion from clients outside the ecosystem remains stable. From an industry perspective, the Top 3 to Top 5 enterprises in the cloud industry can all develop healthily. Our core advantage lies in being based on the unique intelligent computing sector and having seized the first-mover advantage, allowing us to achieve faster development. **Wallstreet News:** Where will KINGSOFT CLOUD's revenue and profit growth points be in the future? **Liu Tao:** This will be a layered contribution structure. On one hand, by providing large cluster supply to the most high-quality clients, we drive the absolute growth of revenue and total profit. Although this segment has a lower gross profit margin, its business foundation is stable. On the other hand, the robotics sector business, mid-sized clients, and AI Agent application-related businesses are high-profit margin segments. Additionally, the Token business is an explosive point for the future. Although its absolute scale is currently small, its growth rate is rapid and is expected to become a growth engine. In 2026, we will continue to move towards the goal of "positive operating profit." ### 相關股票 - [Guotai CSI Computer Theme ETF (512720.CN)](https://longbridge.com/zh-HK/quote/512720.CN.md) - [ChinaAMC SSE STAR Semiconductor Material & Equipment Thematic ETF (588170.CN)](https://longbridge.com/zh-HK/quote/588170.CN.md) - [GTJA Allianz CSI All Share Semiconductors & Semiconductor Equipment ETF (512480.CN)](https://longbridge.com/zh-HK/quote/512480.CN.md) - [ChinaAMC Guozheng Semiconductor Chip ETF (159995.CN)](https://longbridge.com/zh-HK/quote/159995.CN.md) - [KINGSOFT (03888.HK)](https://longbridge.com/zh-HK/quote/03888.HK.md) - [China Southern CSI Semiconductor Industry Custom ETF (159325.CN)](https://longbridge.com/zh-HK/quote/159325.CN.md) - [Tianhong CSI Computer Theme ETF (159998.CN)](https://longbridge.com/zh-HK/quote/159998.CN.md) - [KINGSOFT CLOUD (03896.HK)](https://longbridge.com/zh-HK/quote/03896.HK.md) ## 相關資訊與研究 - [AsiaInfo leans into ‘AI First’ as profits hold despite telecom slowdown](https://longbridge.com/zh-HK/news/280186084.md) - [15:54 ETOmniTrust Breaks the Certificate Lifecycle Management Model with $99,000 PKI and Unlimited Certificates](https://longbridge.com/zh-HK/news/280520979.md) - [Vibe-coding startup Lovable is on the hunt for acquisitions](https://longbridge.com/zh-HK/news/280202885.md) - [Hitem3D Becomes Callable via OpenClaw Skills, Enabling Structured Image-to-3D Generation](https://longbridge.com/zh-HK/news/280229587.md) - [US move to ban China routers as part of onshoring push could be uphill battle](https://longbridge.com/zh-HK/news/280333136.md)