Here’s a question I’ve asked myself more than once after fifteen years around manufacturing execution systems: if a plant already has PLCs running the machines and SCADA or a DCS supervising them — often decades of it, deeply reliable — why did anyone ever buy an MES on top of that?
The honest answer is that control and execution are two different jobs. A PLC keeps the process running correctly right now. ISA-95 draws the layers: sensors and actuators at the bottom, PLCs and SCADA/DCS doing real-time control above them, MES sitting at Level 3, ERP at the top. MES was never about the millisecond control loop. It existed to answer a slower, harder set of questions — how well are we operating, how do we make it better, and how do we stay agile when the same line has to build many different things? It turned machine reality into operations: scheduling, dispatch, traceability, OEE, quality, genealogy — the eleven functions the MESA model lays out. The essence of MES was visibility, improvement, and flexibility — not execution.
Control runs the process; MES (Level 3) exists to run the operation — and every layer is a different technology, joined by standardized seams.
That last one has only gotten louder. Manufacturing has moved hard from single-product lines toward mixed-model production — many variants intermixed on one line, lot size one — and from build-to-stock toward build-to-order and mass customization, where a unit isn’t made until someone orders that exact configuration. Picture a paint line running a mix of car models and colors in sequence, or an assembly line where consecutive units are each a different build, delivered just-in-sequence. That’s the environment MES actually has to run in now, and it’s a fundamentally different problem from a line that makes one thing all day.
And the people who got the most out of it were the plant managers and shift supervisors running the floor day to day, and the process and continuous-improvement engineers behind them. Their whole job is to make the operation better — more yield, more throughput, less scrap, tighter quality, and the flexibility to absorb a changing product mix without the line falling over. MES gave them the data to see what was happening.
But it never gave them the one thing they actually needed: a safe place to try what might work. To test a change, you changed the physical line — took the downtime, risked the scrap, and hoped. So improvement was gated by the cost of experimenting on a running plant. The MES could show you the problem in high resolution. It couldn’t let you rehearse the fix.
Two assumptions that became the ceiling
Here’s where I land, and why I think Physical AI is a genuine step change rather than a smarter feature bolted onto Level 3. MES was built on two assumptions that were completely reasonable when they were made, and that quietly became its ceiling.
The first assumption: model everything explicitly, up front. Every entity, every route, every exception had to be defined in advance and wired together. That’s why MES projects drowned in integration — the standards even name the handshakes (OPC UA between control and MES, B2MML between MES and ERP) because the layers were built by different people in different technologies and never spoke natively. It’s also why the data models got enormous: I worked with a FactoryTalk ProductionCentre deployment whose schema ran well past a thousand relational tables, because that’s what it takes to model every operation explicitly — and mixed-model, build-to-order production multiplies that, since every variant, option, and route has to be defined up front, per unit, before the line can run at all. And it’s why the reporting sprawled. Every persona — operator, quality, maintenance, supervisor, plant manager, corporate — needed a different view, so through the 2010s we built dashboards until there were more than a hundred of them, then stood up SAP BusinessObjects Universes to keep them fed (a Universe is a hand-built semantic layer: you define every dimension and measure before a single report can render). At one customer it got complex enough that we added dedicated BI engineers to the MES team just to keep the reporting alive.
The industry’s own answer had exactly the same shape. Rockwell acquired an analytics company called Incuity in 2008 and evolved it into FactoryTalk VantagePoint; Siemens had XHQ Operations Intelligence, GE had Proficy, and Wonderware — later folded into AVEVA — shipped its own Intelligence layer. Across roughly 2010 to 2020, nearly every major vendor sold an “enterprise manufacturing intelligence” product whose entire job was to normalize MES, ERP, historian, and LIMS data into one “Unified Production Model” and unify the sprawl the other systems had created. The tooling got better every year. The underlying problem — a model you had to build and maintain by hand — never went away.
The second assumption: show the data, and a human decides. Improvement was report-and-decide by design. The MES surfaced the number; a person read it and judged what to do. Which means the human was the throughput limit on improvement — you could only get better as fast as people could read reports, meet about them, and act. The BI engineers weren’t a bug; they were what it took to keep feeding a fundamentally human decision loop.
MES sat where two kinds of fragmentation met — and the industry answered each by adding more layers.
Physical AI and agentic AI, running on one shared model, overturn both assumptions. The model no longer has to be hand-built and rigid — it can be a living, simulated representation of the operation. And decisions no longer have to route entirely through a person. That’s not a better dashboard. That’s a different foundation.
Walk the floor
Rather than argue that abstractly, here’s how it changes in the modules I know best. The thing to notice in each is that Physical AI and agentic AI aren’t two separate initiatives — they run in the same loop, on the same model. You don’t pick one.
The same shift, module by module — physical AI and agentic AI in one loop.
Material movement and the warehouse. Today the MES tracks moves, reconciles locations, and keeps WIP honest in its tables — a system of record for where everything is. Tomorrow the material moves itself: AMRs and mobile robots that perceive their surroundings and navigate autonomously (Physical AI), coordinated by an agent that re-routes the fleet against live demand instead of a fixed route plan (agentic). The robots get trained without a year of real-world footage because NVIDIA Cosmos can generate photorealistic synthetic video of the exact warehouse, and they act in the sub-second zone on edge GPUs like Jetson Thor. The MES stops being the clipboard that records the move and becomes the model the movement runs against.
Scheduling and JIS/JIT sequencing. Today the schedule is largely fixed — MRP and heuristics produce a plan, and when reality diverges, a planner re-sequences by hand. That’s hardest exactly where modern plants live: a mixed-model line building a different configuration one unit after the next, where the sequence decides whether the right parts arrive just-in-sequence and whether paint or tooling changeovers stay efficient. Get the order wrong on a build-to-order line and the cost lands immediately. Tomorrow an agent continuously re-sequences on live orders, machine status, and labor availability — but the part that matters most to me is that it can simulate the re-sequence on the digital twin before committing it. That’s the safe place to try that the CI engineer never had. Build the operation in OpenUSD, run the “what if” in Omniverse, and only push the sequence that survived simulation to the floor.
Recipe and process design. Today recipes are authored and then validated the expensive way — trial and error on real equipment, wrapped in heavy change control. Tomorrow you author and optimize in simulation, iterating a thousand times virtually before touching the line, with an agent proposing and tuning parameters against the twin. Digital twins are far cheaper to run than the real world, which is the whole point: continuous improvement stops being throttled by the cost of trying.
Quality inspection. Today, for a lot of lines, quality is sampling plans, SPC charts read after the fact, and a defect that mostly just gets flagged. Tomorrow it’s in-line vision AI that inspects every unit — NVIDIA Metropolis turning cameras into inspection systems — paired with a closed-loop agent that doesn’t stop at “defect,” but adjusts upstream parameters and traces the root cause. And this is real, not a slide: at Foxconn, a Metropolis-based SOP-verification agent reported a 3% first-pass-yield improvement and 99% task-level accuracy on critical assembly steps; Corning trained a defect model on eight real images plus synthetic data and hit 95% precision, turning a multi-quarter project into days. The synthetic-defect data comes from the same world models — you can generate the rare failure you don’t have enough real examples of.
Maintenance and asset health. Today most plants run time-based schedules — swap the part every N hours whether it needs it or not — punctuated by reactive firefighting when something fails anyway, with the MES logging the downtime after the fact. Tomorrow the asset carries a live model of itself. Multimodal sensors (vibration, temperature, acoustic, current draw) feed condition models that read the specific signatures of trouble — bearing wear, misalignment, and gear defects each show a distinct vibration pattern weeks before anything is audible, a motor’s efficiency drifting down half a percent a week — and estimate remaining useful life, so you know not just that a machine will fail but roughly when. The rare failure modes you’ve never collected enough real data on get generated synthetically from the physics — the same trick that makes the quality models work when you only have a handful of real defect images. And the response runs end to end: an agent diagnoses the likely cause, checks parts in ERP, books the technician, writes the work order, and negotiates the downtime window with the scheduling agent so the fix lands when it costs production least — with the intervention simulated on the twin before anyone touches the asset. Where predictive maintenance is done well the gains are consistent enough to be almost boring: long-standing U.S. Department of Energy figures put it at 25–30% lower maintenance cost and 70–75% fewer breakdowns than the schedule-based approach it replaces, and Deloitte reports 10–20% lower maintenance cost with 20–25% more uptime. The module stops logging failures after the fact and starts heading them off.
These five are illustrations, not the whole map. MES spans more — the MESA model counts eleven functional areas: detailed scheduling, resource allocation and status, dispatching production units, document control, data collection and acquisition, labor management, quality management, process management, maintenance management, product tracking and genealogy, and performance analysis. I picked the five where the change is easiest to feel, but the same pattern runs down the whole list: whatever the module, the shift is from a hand-modeled, report-and-decide system toward one that perceives, simulates, and acts on a shared model.
What ties all five together
Look at those five and the pattern is the same: none of them works without a single, shared model of the operation that both the physical and the agentic side can read and write. That model is the real story — and it’s a different kind of model than MES ever had. The old Unified Production Model was tabular: rows, dimensions, measures. This one is spatially aware. It holds the environment (the plant, the line, the cell, and their geometry), the assets (machines, robots, fixtures, materials, with their real physical properties), and the data (live state, telemetry, orders, quality) in one place — so the simulation, the robot, and the vision system are all reasoning about the same world instead of three disconnected views of it.
That’s why the right foundation turned out to be OpenUSD, not another database schema. USD — Universal Scene Description — started at Pixar as the way to describe and compose huge 3D scenes, with many artists layering their work into one shared world; Pixar open-sourced it, and it’s now an open standard. That layered, composable, 3D-native design is exactly what a plant needs: one description of the scene and the operation that CAD, simulation, robotics, and execution tools can all read and write, with Omniverse as the place you run simulations against it and Cosmos and Metropolis feeding it synthetic data and perception. It’s the BusinessObjects Universe’s instinct — normalize everything into one model — but spatial, and with the hand labor taken out.
Not another database — a spatially aware model that the machines and the agents both work from.
That’s also why I keep insisting Physical AI and agentic AI belong in the same conversation. They’re two capabilities working on one model: Physical AI perceives, simulates, and acts on it; agentic AI reasons, decides, and coordinates over it. They only combine because they’re reading the same thing.
Where I’d temper this
I don’t think the whole pyramid collapses at once, and it would be dishonest to pretend otherwise.
The safety-critical control edge is the hardest part to converge. Level 1 and 2 control is hard-real-time and certified precisely because it’s deterministic and auditable; you don’t swap a safety PLC for a probabilistic policy casually. What’s actually changed — and this is the part that moved my own thinking — is the hardware. There’s a concrete way to see it. A PLC running a line-level control process typically scans every 50 to 100 milliseconds; but the moment you loaded that same PLC with MES-level functions and heavy data handling, scan times stretched to 500 milliseconds or more — control and data were competing for the same cycle, and you paid for every bit of analysis in latency. GPU compute breaks that trade-off. Edge chips like Jetson Thor now run heavy analysis right next to the control loop and still answer in well under a second — fast enough for a machine to think and act on its own on the floor, instead of shipping the data off to a datacenter and waiting for an answer. That kind of work was too slow to run on the line a couple of years ago; now it fits. The edge is converging faster than a skeptic would have told you in 2024 — but the deepest safety loops will stay specialized for a while, and that’s fine.
The bigger caveat is the one hiding in plain sight: all of this rests on that unified model, and most plants don’t have it yet. The reason we needed a hundred dashboards and a hand-built Universe in the first place is that the data was never coherent — different sources, different meanings. Point an agent at that same fragmented mess and it will hallucinate confidently; run a twin on bad data and it will mislead you precisely. “Just add AI” fails for exactly the reason MES was hard: building that shared model is the work. The cure was never the AI. It’s the shared model — Physical AI supplies it for the machines, agentic AI consumes it for the people, and neither is worth much without it.
So — is it the next level?
Yes, I think it is, but not for the reason it usually gets sold. It isn’t that MES gets an AI feature or a chatbot. It’s that the two assumptions MES was built on — model everything by hand, and route every decision through a person — are the two things a shared model plus autonomy actually remove. MES doesn’t die in this story. It finally gets to be what it was always for: a continuous-improvement engine, where seeing, simulating, deciding, and acting happen in one loop on one model, instead of across a stack of systems held together by integrations and a hundred dashboards.
That’s the version I’d bet on. I’ve been wrong about timing before, and the control edge and the data problem are real gates. But the direction feels less like a trend and more like the shape the whole thing was reaching for all along.
This is my own individual opinion, formed from my personal experience working with this technology. It does not represent an official NVIDIA position. I may have gotten something wrong or missed a detail — always happy to be corrected.