The next wave of artificial intelligence isn't happening in data centers anymore; it's happening on the edge, inside devices and robots that need to think and act in real time without waiting for cloud responses. But there's a problem that's getting far less attention than the chips powering these systems: storage. As AI workloads move closer to where they're actually used, the demand for fast, high-capacity local storage is exploding in ways that most people haven't noticed yet.

What Exactly Is "Physical AI" and Why Does It Need So Much Storage?

Physical AI refers to intelligent systems that can perceive, understand, reason, and interact with the physical world in real time. Think autonomous vehicles, warehouse robots, drones, and especially the emerging wave of humanoid robots. Unlike chatbots that answer questions, these systems must continuously learn from their environment, store massive amounts of sensor data, and make split-second decisions without relying on a distant server.

Humanoid robots are projected to reach 1.4 million units by the mid-2030s, and investment in the sector is growing faster than anticipated. Each of these robots will be a data-hungry machine. An industrial humanoid robot, for example, needs to store large Vision Language Action datasets (which combine images, text, and movement instructions), data from multiple cameras and LiDAR sensors (light detection and ranging, a technology that creates 3D maps), local AI inference models, high-frequency motion logs, and maintenance telemetry. The result: some industrial robots will require multiple terabytes of NVMe solid state drives (SSDs), which are the fastest type of storage available.

A general-purpose humanoid robot, designed to handle multiple tasks across different environments, will actually need even more storage. These robots run multimodal foundation models and world models locally, covering speech, vision, and control, often with large language models (LLMs, which are AI systems trained on massive amounts of text) and diffusion models deployed at the edge to reduce latency. Field or hazard-zone robots could demand even more storage capacity than general-purpose models.

How Is Edge AI Storage Different From Cloud Storage?

The shift from cloud-based AI to edge-based AI fundamentally changes how data flows through systems. In the old model, robots and devices would send data to the cloud, where it would be processed and analyzed. Now, with edge inference, data stays local for immediate decision-making. The storage on the edge is used for real-time decision-making and caching, while important data is sent back to the cloud for long-term storage and centralized model training.

This creates a new architecture where edge devices need both speed and capacity. Flash storage, particularly NVMe-powered SSDs, has become central to smooth operation because these devices need faster access to data than traditional storage can provide. As on-edge intelligence increases, so does the need for edge or local storage. Today, autonomous mobile robots in warehouses incorporate flash storage measured in gigabytes, but tomorrow's humanoid robots will need storage measured in terabytes.

Steps to Understanding Edge Storage Requirements for AI Systems

• Real-Time Inference Needs: Devices performing local AI inference require fast access to models and data, making high-performance flash storage essential for millisecond-level response times in applications like autonomous vehicles and warehouse robots.
• Continuous Learning Requirements: Physical AI systems learn continuously from their environment without manual coding, which means they accumulate training data, embeddings, and learned patterns locally, requiring substantial storage capacity that grows over time.
• Multi-Sensor Data Collection: Humanoid robots with multiple cameras, LiDAR sensors, and other perception systems generate enormous volumes of real-time recordings for anomaly detection and simulation replay, driving storage needs beyond what earlier generations of robots required.
• Model and Knowledge Base Storage: Local deployment of large language models, vision models, and knowledge bases for retrieval-augmented generation (RAG, a technique that lets AI systems reference external documents) means robots must store gigabytes to terabytes of model weights and reference data on-device.

What Does This Mean for the Data Center?

The explosion of edge AI storage isn't just changing what happens on devices; it's reshaping data center architecture. A large amount of data collected by humanoid robots will be sent to data centers for training and inference purposes. This brings a major shift in how data centers operate. AI workloads require more frequent and faster access to data, driving the need for faster data lakes. Consequently, high-performance and high-capacity SSDs based on UltraQLC technology (a newer flash storage approach that packs more data into the same space) are becoming foundational infrastructure.

Storage vendors are responding to this shift. SanDisk announced that its 128TB and 256TB NVMe SSDs with UltraQLC technology set a new benchmark for hyperscale flash storage, purpose-built for the fast, intelligent data lakes powering AI at scale.

How Are Companies Building Complete Edge AI Systems?

The challenge of supporting both edge inference and persistent AI workloads has led to new system architectures. MINISFORUM, a global edge computing brand, unveiled two new local AI computing solutions powered by the AMD Ryzen AI MAX+ PRO 495 processor: the AI Mini Workstation MS-S1 MAX-P495 and the AI Agent NAS N5 MAX-P495.

The N5 MAX-P495 combines powerful AI computing with up to 200TB of local storage, enabling users to store, manage, and process large volumes of data within a single AI platform. It's engineered for local data management and AI processing, offering massive bandwidth and high-density compute. The system keeps data and AI workloads local, enabling continuous and private AI operations for applications like retrieval-augmented generation knowledge bases, AI models, embeddings, databases, and AI agents.

The MS-S1 MAX-P495, by contrast, delivers powerful local AI computing for demanding model inference, AI development, and computationally intensive workloads. With up to 131 TOPS of AI compute performance (a measure of how many trillion operations the processor can perform per second), 192GB of memory, and up to 160GB of graphics memory, it supports larger models and more complex AI workloads. The system is designed for low-latency AI workloads, enabling responsive local inference for AI agents, computer vision, generative AI, and other real-time applications.

Together, these two systems form a complete edge AI computing architecture: high-performance AI compute at the frontend for real-time inference, combined with compute, massive storage, data, and AI knowledge at the backend for persistent workloads and scalable local AI applications.

Why Should You Care About Edge Storage Right Now?

As physical AI becomes the latest sector to develop at lightning speed, the growth and success of humanoid robots will inevitably depend on a foundation built on flash storage. This isn't just a technical detail for engineers; it's reshaping how companies think about infrastructure, from the devices themselves to the data centers that support them. The storage bottleneck is becoming as important as the compute bottleneck, and solving it will determine which AI systems can actually operate reliably in the real world.