Cloud vendors compete for the AI foundation
```
Author | Zhou Zhiyu
Editor | Zhang Xiaoling
While the world is still debating whether large models are just an expensive "bubble," Kingsoft Cloud Senior Vice President Liu Tao says this is not a bubble.
What supports his assertion is the ongoing "Vibe Coding (immersive development)" craze—when Claude Code can already skillfully iterate itself with its own code, the "singularity" of robots building robots and code writing code has truly touched the backbone of the industry.
Faced with this change, Kingsoft Cloud has made its move. On January 21, Kingsoft Cloud announced a comprehensive upgrade of its intelligent computing platform "Kingsoft Cloud Xingliu," which now not only includes a training and inference platform covering the entire lifecycle of models, but also officially launched a robotics platform and model API services.
This veteran cloud service provider is attempting to actively participate in the contest over the future dominance of productivity by reinventing itself.
Over the past year, demand for intelligent computing has continued to rise, with the main driver quietly shifting from the training needs of leading enterprises to a boom on the inference side.
The data is very intuitive: Volcano Engine's average daily token calls have reached 50 trillion; spillover demand for models like Doubao, Qianwen, and Yuanbao is expanding at an incredible speed. This explosive token consumption essentially marks the process of AI being implemented into real-world scenarios. For companies, large models are no longer just embellishments in PowerPoints, but are real tools for reducing costs and increasing efficiency.
"We've always been watching for when inference would explode—this growth rate has surpassed all previous understanding of IT infrastructure," marveled Sun Xiao, Assistant President of Kingsoft Cloud.
Against this backdrop, Kingsoft Cloud positions itself as an engineering companion, following a very pure logic: since large models are becoming the "brains" of everything connected, cloud vendors need to provide the matching "circulatory system"—a stable, efficient, and supremely cost-effective token service.
Kingsoft Cloud follows a "task-driven" path of logical evolution. In 2023, the industry theme was "building large-scale intelligent computing network infrastructure," competing on the ability to manage underlying heterogeneous resources. In 2024, the focus has evolved into "platformization and serverless," with the core being the shift from resource delivery to task delivery. By 2026, the upgraded "Kingsoft Cloud Xingliu Platform" will anchor itself to three core themes: efficiency improvement, building industry platforms, and accelerating inference implementation.
This shift is due to the fact that training tasks in the intelligent computing era are extremely fragile—any minor hardware fluctuation in a large scale compute cluster could cause the entire training task to be interrupted. To solve this "nail" issue, Kingsoft Cloud has independently developed a set of self-healing technology based on fault perception.
This system can handle hardware faults and combined hardware-software faults with tiered responses. Some faults can be handled by a simple restart, while others require immediate replacement strategies. Sun Xiao revealed that this mechanism enables "second-level perception" and rapid handling. This means that even if underlying hardware fluctuates, a customer's training task can continue smoothly without interruption.
Embodied intelligence is, in Kingsoft Cloud’s view, the "second half" of the intelligent computing cloud, and a key focus for the future.
Whether in autonomous driving or humanoid robots, the industry is still in a "chaotic scenario"—a flourishing of many approaches, but persistent pain points. Different vendors focus on the "brain," the "cerebellum," or are stuck at data simulation.
Kingsoft Cloud’s newly released "Kingsoft Cloud Xingliu Robotics Platform" attempts to connect the entire closed loop from data collection, storage, labeling, to model training, deployment, and simulation. Sun Xiao believes that the robotics scenario needs to solve the hard problem of "implementing from algorithm development to real-world deployment."
Take autonomous driving as an example: the model is on the vehicle end, but training happens in the cloud. Here, the requirement for computing density may not be high, but the demands for memory and multi-modal point cloud data processing are extremely high. Kingsoft Cloud, by building a closed-loop data platform, allows customers to receive and process massive data more conveniently.
Looking to the future, Liu Tao describes a vision: from 2026 onwards, household robots will gradually roll out. Starting with simple tasks like helping elderly people pick up socks and towels, and ultimately assisting with daily living, this is a trillion-yuan track spanning five to ten years. Kingsoft Cloud's goal is to become the "foundation" and "engine" of this massive industry.
As the traditional public cloud market turns into a stock competition, intelligent computing clouds are bringing brand-new growth opportunities. Kingsoft Cloud's achievement of 120% year-on-year growth in Q3 last year is essentially due to seizing this wave of productivity restructuring.
Sticking to its promise of "not making big models" has made Kingsoft Cloud extremely open in ecosystem building. Sun Xiao admits their job is to, based on open-source models and proprietary technology, provide the most stable and cost-effective token service. Whenever any hot model is released in the industry, Kingsoft Cloud can launch inference service on the same day—this response speed is among the industry's best.
In the intelligent computing era, supporting products are undergoing drastic changes. Previously, the focus was on computing, storage, and networking; now, it's a technology stack centered on inference acceleration (including engines, operator optimization), as well as ecosystems built around Agents. Kingsoft Cloud, through PD separation (pre-fill and decoding separation) and quantization technologies, is pushing inference latency to the minimum and improving throughput performance.
Even behind certain blockbuster games, Kingsoft Cloud is providing full-stack cloud services. In the heavy-load, high-concurrency launch phase, Kingsoft Cloud’s platformization plus engineer protection mechanisms have smoothly supported the influx of massive numbers of players. These extreme stress test experiences, accumulated from game support, are now being transferred to the large model inference battlefield.
The cloud market of the past decade was a contest over resource scale, where cloud vendors played the "public utility" role like water, electricity, or gas. By 2026, technical "internal competition" will continue. From larger parameter sizes to more advanced computing methods (such as MLA or linear Attention), domestic large model vendors are still relentlessly pursuing the limits of efficiency.
But the true dividing line lies in "applications." The practical use of video generation, generalized VLA models for onboard and robotic use, and the penetration of Agents into people's daily lives will bring inference demand to exponential growth.
The dividend period for intelligent computing clouds will not last forever. Only those vendors who can solve extreme engineering problems, connect industry data closed loops, and provide supremely cost-effective tokens will remain standing when the tide recedes.
"We have laid a solid customer foundation—over the next three to five years, Kingsoft Cloud will see significant changes," Liu Tao said confidently.
As the gears of intelligent computing accelerate, this race for productivity dominance is just entering its most brutal—and most exciting—deep-water zone.
Risk Warning and DisclaimerThe market involves risk, investment must be cautious. This article does not constitute personal investment advice, nor does it consider the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their particular situation. Investing accordingly is at your own risk.
```