The AI Era: From Data Backup to Data Resilience Platform

Having worked in the data protection field for many years, I have witnessed several cycles of industry evolution. I know that each wave of technology inevitably brings with it new demands for data protection. However, today's AI-driven transformation is different. As large AI models and agents are rapidly integrated into enterprises' mission-critical services, the underlying logic of the data protection industry is being rewritten.

The evolution from traditional data centers (DCs) to AI data centers (AI DCs) has now become irreversible. This positions the data protection industry at a crossroads—not simply another performance upgrade, but a paradigm shift from traditional data backup to data resilience platforms.

40 Years: Three Evolutions, Three Leaps in Value

Over the past 40 years, the data protection industry has been shaped by three major milestones:

The first came in the 1990s and early 2000s, when backup software was combined with tape libraries. At that time, physical servers and centralized storage were the mainstream. The core objective of data protection was to create reliable data copies so that data could be restored in the event of a failure. This primarily addressed the challenge of reliable backup.

The second began in 2000 and continued into the 2010s, when backup software and dedicated backup storage became widely adopted. As virtualization and cloud computing gained popularity, data volumes grew exponentially. Technologies like deduplication, compression, forever incremental backup, and fast recovery matured during this period. The focus then shifted from backup to fast recovery.

The third milestone, from 2010 to the 2020s, was the era of secondary storage and all-flash backup storage. Ransomware attacks grew increasingly severe, with backup systems becoming a primary target. Capabilities such as immutable storage, physical isolation (Air Gap), ransomware detection, secure isolation and recovery (Clean Room), and automatic recovery drills became industry standards. This phase focused on guaranteeing secure and reliable recovery.

We are currently undergoing the fourth transformation. Unlike the previous three, this transformation is not driven by storage media or backup technologies themselves, but by profound changes in enterprise IT infrastructure and business models, driven by the rise of AI workloads. This is clearly reflected in three fundamental shifts:

First, protected objects have transitioned from databases and virtualization to AI operating environments and agents. In the past, data was protected at the level of databases, virtual machines, file systems, and containerized applications. Today, AI training platforms, inference platforms, agents, and multi-agent collaboration systems are rapidly entering the production environment. The scope of data protection has expanded from traditional workloads to the full lifecycle of AI operating environments and agents.

Second, the form of data assets has moved from structured and unstructured data to vectors, knowledge bases, and memory banks. The AI era has given rise to new forms of data assets, including training datasets, model checkpoints, vector databases, RAG knowledge bases, agent memory banks, agent workflows, and inference logs. However, while models can be retrained and computing resources can be continuously expanded, the domain-specific knowledge, vectorized indexes, and memories and experiences accumulated through agent operations are non-renewable strategic resources. Data has evolved from a supporting element for services to critical fuel for AI.

Third, data risks have shifted from physical failures, human errors, and ransomware to AI-assisted attacks and uncontrolled agents. Today, hardware failures, human errors, and ransomware attacks remain significant risks. However, AI brings a brand-new attack surface that includes data poisoning, knowledge base pollution, prompt injection, model tampering, and agentic failures. These new risks do not necessarily cause immediate system failures, but they continuously erode the trustworthiness of AI-driven decision-making. In the future, data protection must not only ensure service continuity, but address the evolving security battle between AI systems.

These three transformations—new ecosystems expanding the protection boundary, new data enriching the scope of protected objects, and new risks reshaping threat models—collectively form the underlying logic driving the evolution from data backup to data resilience platforms.

Global regulators are also raising the bar for related requirements. The EU AI Act requires high-risk AI systems to automatically log input data, inference processes, output results, associated data sources, and human oversight records, all while retaining relevant logs for at least six months. This means that, in the future, data protection must extend beyond restoring data itself to restoring model versions, knowledge versions, inference context, and complete audit trails.

The Data Resilience Platform Must Cover All Scenarios

The future-proof data resilience platform must address three core scenarios:

1. Traditional services: superior performance for deeper industry development

Databases, VMs, and file systems remain the cornerstones of enterprise core services. The data protection requirements of these three traditional workloads are growing increasingly stringent, including needs for faster recovery, higher security, and lower total costs. Deep software-hardware collaboration is essential for delivering superior data protection and enabling secure, reliable recovery.

For example, in the government sector, government customers face complex O&M challenges due to heterogeneous environments spanning multiple vendors, platforms, and versions. Huawei provides a full-stack solution covering backup software, backup servers, and backup storage, simplifying delivery and enabling one-stop O&M. This significantly lowers technical barriers and ensures secure and compliant government data throughout its lifecycle. In the financial sector, transactions run 24/7, and any data loss may cause significant financial and reputational damage. The all-flash backup architecture supports recovery bandwidth of up to 100 TB/hour, enabling critical data to be recovered in minutes. This helps guarantee zero data loss for financial institutions and allows them to maintain service continuity even in extreme scenarios.

2. AI training and inference: protecting new assets and addressing new risks

Training datasets, model checkpoints, vector databases, prompt templates, and model repositories have become core components of the AI production pipeline. However, new forms of attacks like data poisoning, knowledge base pollution, and model tampering may cause AI to continuously generate incorrect results. To address these challenges, three key capabilities of data protection in AI environments must be strengthened:

First, backup performance. Backup storage performance improvement and active-standby collaboration provide efficient direct backup for storage. When combined with data reduction before transmission, this helps address the performance challenges of backing up data in the AI era. In terms of massive training datasets, PB-scale cluster backup with forever-incremental backup and second-level synthesis, alongside multilayer data reduction technologies—including source deduplication, target inline deduplication, and global deduplication—can significantly reduce storage costs for massive AI datasets.

Second, security detection. A built-in model library powered by AI self-learning, training, and inference should be provided to achieve up to 99.99% ransomware detection accuracy, effectively protecting AI backup copies. Multilayer, built-in ransomware protection capabilities establish end-to-end security for critical AI data, including training corpora and vector databases.

Third, ecosystem collaboration. Protection capabilities for AI data ecosystems, including vector databases, should be enhanced to provide unified backup and protection for AI corpora and vector databases. This includes support for both distributed and vector databases and enables the continuous development of a dedicated protection system for AI workloads.

3. Agents: risk escalation, from assistance to execution

If large models have changed the way we access information, agents are changing the way we execute services. According to IDC, there are already over 30 million active agents worldwide, with this number expected to reach 2.2 billion by 2030. Today, agents can directly access databases, execute O&M scripts, modify service configurations, and even independently execute complex workflows.

The risk model is changing dramatically. In the past, a single misoperation usually affected only a single system. In the future, multi-agent collaboration may result in hundreds or even thousands of operations being executed within minutes, and local errors may quickly escalate into systemic risks. In recent years, there have been frequent incidents involving agents autonomously acquiring permissions, deleting critical data, and causing service disruptions. In the future, data protection must not only prevent malicious attacks, but also address new risks that arise when agents use legitimate permissions to perform incorrect or unintended operations.

Data Resilience Platform in the AI Era: From Reactive Backup to Proactive Protection

In response to the fourth transformation, the data resilience platform is shifting from reactive backup to proactive protection. Huawei believes that a complete resilience loop should cover three capabilities: awareness, protection, and recovery.

Awareness layer: real-time risk visualization and prediction. Based on agent behavior graphs and lightweight intent-recognition models, the system can analyze agent behaviors in near real time and accurately detect potential risks like batch deletion or modification, unauthorized access, and prompt injection, enabling early risk detection and alerts.

Protection layer: real-time protection of fine-grained assets. The production and backup systems are linked based on risk levels. With the help of the AI data platform, assets such as knowledge bases and memory banks are backed up in near real time, preventing misoperations or malicious tampering from overwriting critical assets.

Recovery layer: accurate recovery with contextual consistency. The platform traces the complete causal chain of agent activities, accurately identifies the scope of any risky operations, and matches them with trusted, isolated copies. This ensures 100% consistency during the recovery of service data, knowledge bases, and inference context, thus addressing the challenges of incomplete and inconsistent recovery in AI scenarios.

A Paradigm Shift Has Just Begun

The shift from data backup to data resilience platforms is essentially a paradigm shift from preserving data copies to ensuring the trustworthiness of AI-driven decision-making. It is also a crucial step toward ensuring the trustworthy operation of data infrastructure in the AI era.

As the capabilities and autonomy of agents continue to improve, the scope of data resilience will continue to expand. The future data resilience platform must be fully compatible with traditional IT workloads and full-stack AI assets, capable of defending against risks across the entire spectrum, from traditional infrastructure failures to AI-driven attacks, and meet the requirements for business continuity, data trustworthiness, and regulatory compliance.

This is not a task that can be completed by a single company, but a challenge the entire industry must address together. Huawei looks forward to working with industry partners to advance standards and technologies in this emerging field and build a solid foundation for data resilience in order to support every new innovation in the intelligent world.

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.