The Five-Layer Architecture of AI DC Data Infrastructure

For decades, the relationship between humans and technology has followed a one-way interaction model—we input instructions, the machines return results. From search engines to e-commerce, this input-output paradigm has fueled trillion-dollar industries. However, AI is now reshaping the paradigm. Instead of simply answering questions, it is capable of taking actions on our behalf. And, as the solid foundation of AI, data infrastructure must similarly evolve. Its positioning, hardware system, and software capabilities all need to be systematically reconstructed.

From Serving Humans to Serving Humans and Agents

In the age of databases and cloud computing, the primary role of traditional IT infrastructure was to support device operations. So its core functions centered on providing static information access, network transmission, and stable server operations, while reactively handling user-initiated database queries, file access requests, and virtualization workloads.

But AI has changed this paradigm. AI agents are now being widely adopted in real-world business scenarios across various industries. They can identify user intents, break down tasks, dynamically adjust strategies, and autonomously coordinate multiple service systems without the need for frequent manual intervention. As a result, organizations can achieve business objectives far faster and at much greater scale. According to IDC, more than 30 million active AI agents are already integrated into our work and daily lives, and this number is expected to reach 2.2 billion by 2030. This rapid growth will fundamentally transform how enterprises productize, operate, and manage their businesses.

Against this backdrop, data is no longer just a static asset to be stored. Instead, it must flow continuously from the moment it enters the data infrastructure, supporting the commercial implementation of AI agents and continuously creating business value. As a result, data is becoming the most powerful driver of productivity, making the design of data infrastructure a key challenge for the industry.

The Five-Layer Architecture of AI DC Data Infrastructure

To support this transformation, data infrastructure—the most critical component of AI data centers (DCs)—must be systematically reimagined and planned around the entire AI lifecycle. We break it down into five layers.

• AI data lake: As the foundation, the AI data lake addresses the issue of corpus production and supply for AI systems. The quality, diversity, and scale of industry-specific corpus data will directly determine the potential of upper-layer operations. Enterprises need to integrate data assets scattered across different systems and regions into a data lake. Then, they need to develop high-quality, easy-to-search, and circulatory corpus libraries that enable global data visibility, manageability, and availability. Although this is the first step in building AI DC data infrastructure, it is also the most challenging—and time-consuming.

• Knowledge and memory platform: Knowledge and memory capabilities are the key to distinguishing AI from previous technologies. By integrating knowledge across fields such as biology, economics, and chemistry, AI understands the logic of the physical and virtual worlds, unleashing its transformative capabilities. To support the evolution of data into knowledge and memory, enterprises need to build dedicated knowledge bases that are highly accurate, traceable, and resilient; continuously evolving memory banks that accumulate knowledge; and intelligent, hierarchical, and massive KV cache stores. Together, these capabilities improve the efficiency and accuracy of upper-layer model inference.

• Compute: The compute layer directly determines model iteration speed, training and inference efficiency, and service scale. The industry's focus has shifted from competing on raw floating-point operations per second (FLOPS) to maximizing token efficiency. To meet these demands, enterprises need to build a diverse computing resource pool, provide customized planning for valuable computing resources, and maximize the utilization of these resources. This enables high-concurrency processing and ultra-low latency interconnection across large-scale clusters, improving the utilization of computing power.

• Model: Sitting atop compute is the model layer. For enterprises today, the priority is no longer building a powerful model from scratch, but efficiently deploying, adapting, continually pre-training, and fine-tuning models. Only truly out-of-the-box model engineering capabilities and end-to-end AI toolchains can accelerate AI deployment and enable AI to precisely solve real-world business pain points.

• Agent: Positioned at the top of this five-layer architecture, agents are the tangible vehicle through which data delivers actual value. Backed by data, knowledge, compute, and models, agents can be deployed intelligently and evolve autonomously, drastically reshaping complex business workflows to support enterprises' real use cases.

The five layers of AI DC data infrastructure support and reinforce one another. None can reach its full potential in isolation. The more graphics processing units (GPUs) an organization deploys, the greater the demand for high-quality data supply, high-speed transmission, high-precision knowledge scheduling, and mature, smart models.

Currently, foundational model capabilities are iterating at a monthly pace, while industry-specific models and tailored use cases are expanding exponentially. Concurrently, top-tier open-source models like DeepSeek are being widely adopted by enterprises and developers globally, making new technology benefit more people than ever before. In addition, domain-specific agents spanning financial risk control, power grid inspection, medical diagnostics, and more are emerging at scale. Together, these trends create a compounding effect that continuously amplifies, driving data infrastructure to become both smarter and far more cost-effective.

What We're Doing and the Gaps We See

At Huawei, we are unswervingly customer-centric. Our mission is to help customers think systematically about how to design data infrastructure truly aligned with AI—the most revolutionary workload of our time. Over the past few years, Huawei has achieved a series of breakthroughs in foundational technologies and rapidly translated them into product capabilities.

For instance, we have taken the high-density capacity and intelligent tiering feature of OceanStor Pacific scale-out storage and combined it with the DME Omni-Dataverse unified data space to build an AI data lake solution that helps enterprises aggregate and supply high-quality data at scale.

Huawei has also pioneered the industry's first Context Memory Storage (CMS) that supports heterogeneous compute, alongside our 3 + 1 AI data platform solution, directly improving model inference accuracy.

Our high-performance AI storage—the OceanStor A series—continues to break boundaries, too. Powered by cutting-edge architecture, massive scalability, and intelligent algorithms, the series has secured the top spot on both the globally recognized MLPerf and IO500 storage performance benchmarks.

And the capabilities of the Huawei ModelEngine AI toolchain continue to enrich, offering features like model gateways, resource scheduling, and streamlined agent deployment, making AI deployment significantly easier for enterprises.

At the same time, we recognize that AI is revealing a set of new frontier challenges across fields. Many critical technologies still require big breakthroughs, including the application of new storage media, the optimization of new networks, upgrading to new storage architectures, the adaptation to new data paradigms, the construction of new resilient runtime environments, and the sustainable development of new energy. Equally important, the storage industry faces a significant shortage of multidisciplinary talent and vast market opportunities remain untapped.

Today, AI has shifted from a technological concept to concrete actions and a core driver of productivity inside enterprises. Looking ahead, we hope that, through the joint efforts of Huawei, industry partners, and academia, data infrastructure will continue to break new ground, enabling AI to evolve from assisting human work to autonomously completing it, from single-point automation to full-spectrum business intelligence, ultimately delivering lasting business value.

Disclaimer: The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy, position, products, and technologies of Huawei Technologies Co., Ltd. If you need to learn more about the products and technologies of Huawei Technologies Co., Ltd., please visit our website at e.huawei.com or contact us.