Physical AI is leaving the demo lab. Three separate announcements last week make that clear. ASUS launched its unified AI factory platform. SFA got selected for Korea's M.AX autonomous manufacturing program. And in Serbia, the first humanoid robot rolled off a production line at Minth Group's facility in Šabac.
Different companies, different continents, different layers of the stack. But they all face the same problem: how do you operate AI when it is physically embedded in a factory, with real equipment, real safety margins, and real production schedules?
ASUS builds the whole floor
ASUS announced its position at AI Tech 2026. The company is no longer selling racks of GPUs. It now sells what it calls an AI factory, an integrated platform that combines compute, power, cooling, and automation under one governance umbrella.
The interesting part is the digital twin workflow. ASUS says customers can design and validate an AI factory in simulation before installing any physical hardware. It works with NVIDIA's DSX Sim Blueprint, plus Schneider Electric, AVEVA, and IBM, to model compute, storage, networking, and facility infrastructure in a single environment.
That matters for a practical reason. Power delivery is the constraint that kills most AI deployments. If a digital twin can show you a cooling bottleneck before concrete is poured, you avoid an expensive retrofit later. I buy the workflow in principle. I am skeptical about how accurately these models represent physical AI workloads, which have latency and reliability profiles very different from batch training.
The governance angle is the more substantial piece. Agentic systems run with increasing autonomy. ASUS positions governance not as a pre-deployment checklist but as an ongoing operational function. Its AI Software Stack includes a Quota & Billing tool and an MLOps Portal. Both are unglamorous pieces of operational software, and they are exactly what an enterprise needs when AI agents start consuming tokens and calling production APIs.
The hardware catalog is dense but specific. The XA P2N-E2 is a 2U NVIDIA MGX platform with two Vera CPUs and room for two dual-slot GPUs. The UF920-E3-RS24 server is built on NVIDIA's STX modular foundation and pairs with OJ340A-RS60 object storage and VS320D-RS26N storage arrays. The message is that reasoning workloads at the edge and data pipelines in the core need different hardware shapes, and ASUS wants to supply both.
The revenue claim is the part I would stress-test. ASUS says its Kubernetes-based integration shortens time to revenue for customers. That only holds if the customer already has data pipelines and defined workflows. Otherwise the platform shortens time to first inference, not time to revenue.
SFA wants the factory to run itself
SFA is a Korean supplier of robotic logistics systems and factory automation. Korea's Ministry of Trade, Industry and Energy and KEIT selected it for the M.AX (Manufacturing AI Transformation) program in the AI solution company category. The selection gives SFA access to government-led manufacturing AI projects, financial benefits, and consulting support through August 2028.
SFA's target is what it calls Level 5 fully autonomous manufacturing. No human intervention in the control loop. AI analyzes production process data, commands robots and equipment, and adjusts for faults, drift, and changing conditions on its own.
The company already deploys robotic logistics in semiconductor, battery, display, and general manufacturing plants. Its M.AX selection rests on field results, not lab demonstrations. SFA says its AI-based control has been validated across multiple industries, which is a meaningful credential because every factory integration exposes different data schemas, safety constraints, and legacy equipment.
The concrete date to watch is September 30. SFA opens its ART (AI Robot Technology) Innovation Center that day. The facility is designed to replicate real client manufacturing conditions so SFA can collect robot motion data, stress-test control software, and verify that its AI passes production line speed and safety thresholds before it touches a customer floor.
I find the Level 5 label borrowed from automotive autonomy imperfect. But the underlying objective is clear: a factory that detects and reacts to quality drift and equipment faults without a human pressing buttons. The hard part is data integration. Robots, PLCs, and MES systems generate data at different rates and in incompatible schemas. That integration work is not glamorous and it is most of the cost.
Serbia has robots, for now
The most consequential story may be the one in Serbia. Minth Group, a Chinese automotive parts manufacturer, partnered with robotics developer AGIBOT to start serial production of humanoid robots at Minth's existing complex in Šabac. Operations began August 29. The companies call it Europe's first mass-production base for humanoid robots.
The initial investment is around 20 million euros. Minth has announced plans for a wider Robotics Industrial Park in Inđija that would lift total investment to roughly 200 million euros, with an eventual target of 20,000 humanoid robots and robotic dogs per year for European and international markets.
Here is the number that matters most: current plans point to 200 to 250 jobs at the Šabac operation. That is a modest employment footprint. The economic value depends entirely on what those people do. If they assemble imported motors, sensors, compute boards, and battery packs, Serbia captures very little value. If Serbian engineers write control software, handle mechanical integration, run testing and certification, and participate in product development, the project changes Serbia's position in the robotics supply chain.
The robots themselves use multimodal perception, tactile sensing, and AI-based control for material handling, industrial inspection, and repetitive or hazardous tasks. That is a factory-first commercial focus, not a consumer one. AGIBOT reports delivering 5,168 humanoid robots in 2025 and about 8,400 units in the first half of 2026. Those are company figures, not independently audited market data. Treat them as directional indicators.
There is also a geopolitical dimension. The US has tightened restrictions on some Chinese robotics technology on national security grounds. A Serbian factory producing Chinese-designed robots for the European market could face procurement barriers at sensitive industrial customers. Cybersecurity, data handling, and software update policies will matter as much as unit cost. Whether the factory can localize enough software and engineering to read as European manufacturing rather than Chinese assembly is an open question.
The bottleneck is not hardware
Across all three stories, the hardware is the easy part. ASUS ships dense MGX platforms with two Vera CPUs. SFA integrates robot controllers into battery production lines. Minth assembles humanoid frames in Šabac. None of that is the limiting factor.
The limiting factor is integration between the AI and the physical process. Data pipelines that normalize factory telemetry. Control loops that respect safety margins. Governance systems that track which model version authorized which action. Digital twins accurate enough to catch a constraint before installation.
The useful metric is not GPU count or robot units shipped. It is the share of engineering captured locally and the proportion of value sitting in software and testing rather than in assembled metal. ASUS's time to revenue claim holds only if the customer already has data maturity. SFA's Level 5 ambition holds only if its data integration keeps pace with its robotics hardware. Serbia's project becomes a genuine industrial cluster only if the initial 200 jobs grow into thousands of engineers, suppliers, and testing facilities.
Each of these projects will look successful in its first year, because the hardware will be visible and the software bugs will not be. The real test comes at the two or three year mark, when deployment complexity outpaces the demo metrics and the teams have to decide whether integration work was built into the system or bolted on after the fact.