How manufacturers can evaluate AI servers and computing platforms as Physical AI moves factories from repetitive automation toward genuine autonomy.

For decades, manufacturers have relied on robotic arms, automated production lines, and autonomous mobile robots (AMRs) to drive efficiency. However, as product lifecycles shorten and low-volume, high-mix production becomes the norm, a fundamental limitation of traditional automation has become unmistakably clear: it can only repeat pre-programmed actions under known, fixed conditions. The next stage of smart manufacturing is the profound shift from automation to autonomy — empowering machines with the ability to perceive their environment, learn new tasks, and make decisions and take action based on real-world, real-time conditions. This is the transformative power that Physical AI brings to the factory floor.

For manufacturers, the initial step in adopting Physical AI isn't merely acquiring more robots; it's strategically choosing the right AI computing platform and server infrastructure based on critical factors such as AI workload type, model scale, deployment environment, future scalability requirements, and seamless hardware-software integration. And that’s exactly what Foxlink is doing, partnering with technology leaders including NVIDIA and ASUS to convert decades of frontline manufacturing expertise into AI capabilities that can be accumulated, trained, and deployed at scale. From desktop AI supercomputers to enterprise-grade servers such as the ASUS ESC8000-E13 series, equipped with NVIDIA RTX PRO™ Blackwell Server Edition GPUs, manufacturers like Foxlink can select the optimal compute platform for model training, AI inference, physical simulation, and deployment needs. Selecting the right AI server can effectively transform manufacturing data into tangible AI capabilities, initiating the transition from automation to a new era of Physical AI-driven autonomy.

Why the Factory Floor Is the Ideal Training Ground for Physical AI

Compared to open, unstructured environments like homes or city streets, the factory floor presents a unique advantage: clearly defined workflows, controllable workspaces, and a continuous stream of repetitive, high-value tasks. These include assembly, testing, connector insertion, material handling, and equipment inspection, each with explicit goals and measurable success criteria. This inherent structure makes manufacturing environments not just well-suited for robot deployment, but a critical training ground for Physical AI. Moving forward, the value manufacturers create will extend beyond producing physical goods; it will also stem from generating high-quality data and cultivating machine intelligence as strategic assets in their own right.

From Single-Task Robots to Multi-Skill Workers

Before Physical AI can truly take hold on the factory floor, a fundamental question must be addressed: how does a machine learn to perform work that humans have refined over decades of experience? In electronics manufacturing, for instance, operators routinely perform equipment operation, pressing, and tooling tasks across processes like electrical testing, product assembly, and cable/wire harness assembly. Through systematic data capture, manufacturers can now record the intricate details of these tasks — including hand posture, finger joint angles, movement trajectories, tool usage patterns, and task completion times. This process converts tacit operational knowledge, once residing solely in a skilled worker's hands, into structured data that AI models can learn from, thereby digitizing frontline manufacturing expertise.

Historically, training a skilled technician required years of hands-on experience. In the future, that same technical knowledge can be captured through data collection, systematically structured, and translated into robotic operational capability. This represents a pivotal shift for smart manufacturing, moving it beyond pure production toward knowledge transfer and generalization.

Robot Skills, Unlike Human Skills, Can Be Copied

Traditional industrial robots are typically engineered for a specific workstation and task, which is why automation investment has historically been tied to a single production line. Physical AI, however, introduces a fundamentally different model. Tasks such as picking and placing materials, connector insertion, or component assembly can be isolated for data collection and model training, validated, and then expanded incrementally with new capabilities. As a robot accumulates more validated skills, it evolves from a machine performing one fixed action into something closer to a multi-skill worker: a resource that can be dynamically reassigned across workstations based on production demand.

Here's the critical difference: human skill must be cultivated one person at a time, but a trained model's capability can be copied. Once a robotic skill is trained and validated, that same capability can potentially be deployed to other production lines rapidly, without retraining every unit from scratch. This is profoundly reshaping the economics of manufacturing automation, where companies are no longer investing in one-off automation projects, but building a continuously expanding AI asset: a library of skills that can be reused and redeployed across multiple production environments.

This shift also means that future competitiveness in smart manufacturing won't stem from any single piece of robotic hardware. Instead, it will come from the ability to build a compounding library of models and data. This creates a foundation model capable of understanding the environment, human instructions, and task requirements, with capabilities that extend across diverse on-premises devices, including dual-arm robots, mobile robots, and even humanoid robots, paired with task-specific end effectors such as grippers, custom End-of-Arm Tooling (EOAT), or dexterous hands.

Data collection, in this light, is not a one-time project cost. Every round of real-world data collection and model training becomes a reusable enterprise asset, capable of supporting an expanding set of robotic platforms and manufacturing applications over time. As data and models continue to accumulate, the way smart manufacturing scales changes fundamentally — the goal is no longer just scaling the number of machines, but scaling institutional expertise itself.

From Seeing a Problem to Understanding It and Acting

Another critical shift in smart manufacturing is the move from single-point recognition to closed-loop decision-making. Traditional vision AI could answer "what just happened?" It can detect a fire, an equipment fault, or a blocked pathway. But on the factory floor, detection is only the first step. The real question is: once an anomaly occurs, what should the system do next?  In Foxlink's AI agent architecture, visual perception, reasoning models, path optimization, and the LocalMind orchestration platform work together to form a complete perceive-to-act pipeline. For example: when the system receives a material transport task, the AI agent identifies available equipment and plans an optimal route. If a camera detects a fire or a blocked path mid-execution, that environmental information is fed back to the system in real time. The agent then re-reasons and adjusts the task plan, and LocalMind translates the new decision into physical actions the equipment actually executes.

This creates a complete decision loop: perceive → reason → decide → act → feedback. As a result, the smart factory is no longer merely a collection of automated equipment; it becomes an AI system capable of understanding its environment, coordinating equipment, and continuously learning. Physical AI only achieves a true closed loop once data collection, model training, simulation validation, and real-world deployment are all connected end-to-end.

From Automation to Physical AI: How Manufacturers Should Approach AI Compute Platform Selection

As manufacturing shifts from automation to Physical AI, the first step isn't deploying more robots or automated equipment; it's determining which AI computing platform best fits the organization's specific application needs. Compute requirements vary significantly across manufacturing environments. Organizations should evaluate their needs based on AI application lifecycle stage, workload type, deployment environment, and future scalability requirements. Teams still in the model development, validation, and experimentation phase can start with a desktop AI supercomputer. Organizations that require large-scale model training, AI inference, physical simulation, or enterprise-grade on-premises deployment will necessitate higher-performance, more scalable enterprise-class AI servers and compute platforms.

Four Key Dimensions for Evaluating an AI Compute Platform

Manufacturing environments differ significantly in their compute requirements. Here are four dimensions every organization should evaluate:

1. AI Application Lifecycle Stage — Is the organization still validating a proof of concept (POC), or already moving toward full production deployment?

2. Workload Type — Does the use case require model training, AI inference, physical simulation, or graphics computing?

3. Deployment Environment — Is the plan cloud-based, on-premises, or hybrid, and what data sovereignty or latency constraints apply?

4. Future Scalability — Will GPU count, memory, and compute nodes need to scale in step with growing application demand?

From Desktop to Rack-Scale: Matching Compute Platforms to Each Stage

Organizations can select a platform sized to their current development stage. Teams still in model development, validation, and experimentation can start with a desktop AI supercomputer like the ASUS Ascent GX10, built on the NVIDIA GB10 Grace Blackwell Superchip. This allows engineering teams to prototype and validate models at a small scale in a standard office environment, without an immediate investment in data-center-grade infrastructure.

Organizations that need large-scale model training, AI inference, physical simulation, or enterprise-grade on-premises deployment require more powerful, more scalable enterprise AI servers. The ASUS ESC8000A-E13 series, for example, can be configured with NVIDIA RTX PRO™ Blackwell Server Edition GPUs — including the NVIDIA RTX PRO 4500 Blackwell Server Edition and NVIDIA RTX PRO 6000 Blackwell Server Edition — supporting a broad range of workloads including AI model training, inference, graphics computing, physical simulation, and on-premises deployment. This makes it exceptionally well-suited as the core infrastructure for factory-level AI deployments.

Beyond delivering high-performance hardware, ASUS brings a strong, dedicated engineering team capable of providing full-spectrum professional services. These range from initial consultation and in-depth assessment of a customer's requirements and budget, to tailored compute planning and deployment recommendations. ASUS offers a complete, one-stop hardware-and-software AI infrastructure solution, spanning AI/GPU servers, storage servers, full-rack systems, large-scale clusters, and data center build-out. Regardless of where an organization sits in its AI development lifecycle, this comprehensive approach ensures compute infrastructure can scale seamlessly alongside business growth and innovation needs.

Foxlink's Approach: Built on the NVIDIA Physical AI Ecosystem

Foxlink is progressively deploying Physical AI into real manufacturing environments, working with technology partners including NVIDIA and ASUS within the NVIDIA Physical AI ecosystem. Based on the specific demands of different AI workloads, Foxlink selects the appropriate compute platform — from desktop platforms for model development and validation, to enterprise-grade servers capable of supporting factory-level AI training and inference — progressively building out its own AI factory.

The Factory Isn't Just Where AI Is Deployed — It's Where AI Keeps Learning

As data generated in real manufacturing environments continuously feeds back into model training, and is validated through simulation and real-world deployment, the factory ceases to be merely a deployment site for AI. It transforms into a dynamic source of ongoing learning and evolution. The manufacturing expertise an organization accumulates over years can, in turn, be converted into intelligent assets that scale across equipment, production lines, and even entire facilities.

This is the arc of transformation now underway: from single-station automation to robots that can learn multiple skills; from dependence on senior engineers to autonomous maintenance systems that replicate specialized expertise; from isolated point solutions in AI to smart factories capable of perceiving, reasoning, deciding, and acting. Ultimately, competition in smart manufacturing will no longer be defined by who owns the most robots; it will be defined by who can turn the data, models, and experience generated on the factory floor into a continuously compounding intelligence asset. That is the true inflection point in manufacturing's journey from automation to Physical AI.