Become a TechRadar Insider
We’re entering the golden age of data-driven AI. That was the definitive takeaway during the keynote speeches at Huawei IDI 2026 in Paris. At the event, the message was clear: this is a time of ‘data awakening, infra evolving.’
But this ‘golden age’ comes with its own set of challenges.
As investment in AI technology soars, organizations are investing heavily in models and compute - hoarding GPUs and servers - while neglecting the data foundations required for AI deployment to work reliably at scale.
Effectively, the storage has evolved from being an "external system" to becoming an integral part of the inference system. This shift has become a critical and unavoidable component in the implementation of Agentic AI.
The large models have begun to converge and computing power platforms have taken initial shape at scale. However, very few enterprises have succeeded in converting model capabilities into stable, scalable agent deployments. And all of that severely impacts enterprise AI performance.
Huawei believes it’s engineered the solution: a true AI data platform designed to provide AI agents with data, knowledge, and memory that can be directly invoked, improve inference experience and efficiency, and lower the barrier to large-scale AI adoption. This is a unified system that combines a knowledge base, KV cache store and memory bank alongside file storage - all of which supports the AI to perform in real world scenarios.
Yuan Yuan - Huawei’s Vice President and President of the Huawei Data Storage Product Line - summarized the shift during his keynote speech at IDI 2026. “The first chapter of AI is computing powers. The next chapter is about modelling. The third is agentic.” The fourth, he strongly believes, will be data-driven AI.
Knowledge & Memory
One of the fundamental challenges Huawei’s AI Data Platform seeks to solve is knowledge and memory: the hidden bottlenecks for many enterprises. The transition from simple chatbots to autonomous AI agents absolutely depends on infrastructure systems that can consistently retain, retrieve, and act on context over time.
Yet, many organizations risk overlooking this vital layer, hampering their AI’s execution.
The knowledge base is the necessary layer for AI adoption. Yuan Yuan said: “When we do coding engineering, the first step is to do some planning, give some dependencies, domain knowledge, coding standards and engineering. Those kinds of things will compose the knowledge base. It’s a preliminary to the coding procedure.”
Meanwhile, AI deployments engaging in massive workloads demand hardware scaling, forcing organizations to continually upgrade setups with ever-more expensive GPU chips just to maintain basic workflows.
This issue is made more difficult by an AI agent’s need to repeatedly compute, remembering all data and action points within a live session. Because this short-term data is stashed directly on the GPU chip, running multiple agents simultaneously risks exhausting that memory, leading to a crash.
And that’s the bottleneck: the so-called Memory Wall that occurs when the AI’s compute is processing faster than the ability to move and store the data.
A core component of the AI Data Platform, the Key-Value (KV) cache helps to serve the inference and agentic requirements, giving a better, predictable performance.
Yuan Yuan said, “It’s not necessary to put all the key values in context in every round with the GPU card again and again. It’s not cost effective. And it’s not time saving. The right way is to fill the KV cache. We can get rid of the redundancy of KV with the storage to provide KV access ability to the inference procedure.”
A self-evolving memory system absolutely can help orchestrate the storage meters to balance inference costs and requirements.
AI Data Platform is the key to bringing AI into Production
Huawei’s approach to memory for the AI Data Platform is the Context Memory Storage (CMS) - an industry-first - supporting heterogeneous computing power and KV Cache for ultra-scale inference clusters. This shifts the short-term memory away from a business’s ever-growing stack of ever-full GPU chips and on to an external storage pool.
According to Huawei, the CMS “can expand into a PB-scale shared KV cache pool and reduce the time to first token (TTFT) by 90%,” allowing the AI agent to rapidly execute commands.
Addressing enterprises' AI inference scenarios, Huawei’s 3+1 AI data platform focuses on three core elements: the aforementioned PB-level KV cache acceleration, self-evolving lifelong memory bank, and high-accuracy knowledge base to deliver over 95% retrieval accuracy.
This is coupled with one technology - the company’s Unified Cache Manager (UCM), enabling intelligent three-tier KV caching and scheduling across on-chip memory, DRAM and SSD, improving inference accuracy by 30%.
By solving the problems surrounding speed and accuracy, the platform helps eliminate the core challenge for enterprise AI: enabling systems to retrieve information quickly and reliably retain context over time.
The infrastructure evolution
As enterprises make the transition from exploring AI possibilities to building scalable, autonomous agents, a shift in infrastructure priorities is becoming essential.
Understandably, much of the conversation around AI remains focused on the models and compute. However, the reality for large organizations is that successful AI adoption and deployment are tied, almost entirely, to the underlying infrastructure.
The direction of travel is clear. In this ‘golden age of data-driven AI’, success depends on high-quality, high-speed data operating on a layer that makes it usable, accessible, accurate, and contextually aware.
Yuan Yuan declared, "the next chapter of AI is data. Committed to technological innovation in data storage, Huawei will accumulate the experience of industrial AI adoption, and work closely with the entire industry to help customers accelerate their journey into the intelligent era."
By lowering the barrier for enterprise adoption of agents, Huawei’s AI Data Platform is helping to transform AI into reliable and scalable digital employees.
Summing up the key components that define Huawei’s AI Data Platform, Yuan Yuan explained: “First, you need a high-process knowledge base. Second, you need the KV cache to help you to serve the inference and agentic requirements to give you predictable performance. And third, you need some embedded memory system to help your agent get smarter and smarter.”
Businesses that recognize the importance of this necessary infrastructure are the ones that will successfully accelerate the adoption of AI agents to reap the rewards.
