祝融说。

We stand at the threshold of a magnificent era.

Generative AI (AIGC) technology, led by large language models (LLMs), is reshaping every corner of software development, content creation, and scientific research with unprecedented force. This is not merely a technological iteration, but a profound paradigm revolution. Countless enterprises and developers, filled with aspirations for an intelligent future, have plunged into this exhilarating "AI gold rush."

Yet beneath the bustle and noise, a deep and formidable challenge confronts every dreamer: compute power.

The beautiful vision of large models rests upon an immense, intricate, and fragile physical reality -- tens of thousands of expensive GPUs and NPUs, labyrinthine high-speed networks, and parallel storage systems with breathtaking throughput. When algorithm engineers conceive clever model architectures in Jupyter Notebooks, they may not realize that what sustains it all is an intelligent computing center costing hundreds of millions of dollars, consuming energy like a small town, and demanding upkeep comparable to running a large industrial facility.

When a training task worth tens of millions fails due to an "NCCL Timeout," when an inference service serving millions of users crashes from a "GPU OOM," when a CEO asks, "What is the actual utilization rate of the AI servers we invested so heavily in?" -- the AI halo vanishes in an instant, revealing cold, hard engineering problems. The quantitative examples in this book are computational examples or synthetic cases used to demonstrate methods; they do not constitute commitments about the costs or timelines of any specific project.

It is at this moment that a new, pivotal role steps onto center stage: the AI Infrastructure Engineer.

They are the "arms dealers," the "energy ministers," the "infrastructure maniacs" of this era. They do not produce models directly, yet every model born depends on the "factories" they build. They do not write business logic directly, yet every AI application runs on the "highways" they pave. They are the bridge connecting AI dreams with engineering reality, the ultimate stewards of the most expensive and scarcest resource of our time: compute power.

This is the very reason we wrote this book.

It should be noted that this book is not an operations manual detached from a line of thought. The planning, scheduling, and billing of intelligent computing centers are, in essence, the organization of compute power -- a scarce resource -- into order across multiple scales: chips and nodes, clusters and queues, headquarters and edge. This is precisely the extension of "finite order" and "stability continuation" from the RC system in Process Realism: Observational Convergence and the Generation of Certainty into the infrastructure domain: HFU, MFU, and linear scaling are yardsticks of convergence efficiency at different observation scales; Gang Scheduling, Binpack, and fault-tolerant redundancy are concrete implementations of the four engineering conditions -- survival safeguards, minimal investment, path diversity, and continuous correction -- on compute platforms; and the two-tier intelligent computing architecture with internal billing models is the organizational design of evaluation rights and resource quotas. Readers interested in the philosophical foundations may trace back through this entry point.

We found that while tutorials on "how to use" AI abound, systematic, hands-on guides on "how to support" AI -- on how to build and operate an enterprise-grade intelligent computing center from scratch -- are exceedingly rare. Many excellent SRE and DevOps engineers and cloud-native architects, when confronted with this "new species" called AI, feel their existing knowledge systems overturned, and urgently need a cognitive map to guide their transition from "traditional IT infrastructure" to "AI infrastructure."

This book is written for you. It will take you on a deep ascent from "cloud-native" to "AI-native." We will not stop at the simple use of any single tool; rather, we are committed to:

• Building a complete knowledge system: From the lowest-level GPU/NPU hardware selection, to the mid-level Kubernetes platform construction and scheduling, to the upper-level model training and inference serving, and finally to the top-level operations system design, this book will link together the entire full-stack technical chain of AI Infra.
• Emphasizing first principles: We will not merely tell you "how," but also delve into "why." You will understand why native K8s is unsuited to AI, why Binpack scheduling is needed, why PagedAttention exists, and what core contradictions these technologies are designed to resolve.
• Focusing on practice and quantification: We will abandon empty talk, using executable code, reproducible experiments, and computable formulas to turn vague "experience" into precise "engineering." You will learn how to scientifically calculate memory usage, estimate training time, design stress-test plans, and build monitoring dashboards.

Structure of This Book

The book is divided into five parts, like a complete journey of building an intelligent computing center:

Part I: Introduction. We will establish a macro perspective, understanding the overall architecture and technical challenges of an intelligent computing center.

Part II: Foundation. We will lay the groundwork with our own hands, solving the three core platform problems of compute access, image construction, and resource scheduling.

Part III: Core. We will go deep into the "production workshop," mastering the full-process operations of large model training and inference, and learning how to accurately account for compute costs.

Part IV: Insight. We will install "eyes" and a "brain" for our vast system, building a full-link monitoring framework and learning how to handle all manner of complex ailments.

Part V: System. We will move from technology to business, designing a two-tier operations system and billing model, and looking ahead to the future of LLMOps and AI Infra.

We know all too well that "what you gain from books is shallow; only by putting it into practice do you truly understand." To that end, the appendix of this book provides a detailed "hands-on manual" and an out-of-the-box "efficiency toolbox," hoping to help you put what you have learned into real-world practice.

If you aspire to be a "core builder" of the AI era, if you are not content merely to be a user of AI technology but wish instead to be the "helmsman" steering this wave of technology -- then this book is for you.

Let us together embark on this hardcore journey of exploring the art of compute.