Part 1 of 3: Why the dark factory isn’t coming. It’s already here.

Xiaomi’s Beijing smartphone plant runs without humans on the production floor.

Not “mostly without.” Without. The lights stay off. One factory. Ten million phones a year. Machines that re-orchestrate themselves between shifts. Quality decisions made by models, not supervisors.

This isn’t a press release. It’s already shipping product.

And it isn’t just Xiaomi. Foxconn has dark lines. BMW has them. Siemens has them. The Indian electronics sector — the one I work in — is three to five years behind, but every serious EMS leader I’ve spoken to in the last twelve months has the same question on their desk:

How do we get there from here?

The answer isn’t more robots. It isn’t more cameras. It isn’t another AI vendor pitch.

The answer is the layer of software that sits between your machines and your business — the layer most factories underestimate, undersell, and accidentally pick the wrong one of.

The MES.

What “Dark Factory” Actually Means

The phrase gets thrown around. Most people picture an empty warehouse with robots gliding around. That’s the visual. It’s not the substance.

A dark factory is a factory where decisions are made without waiting for humans.

A board fails AOI inspection at 2:14 AM. The system reroutes it to rework. The rework station is already prepped. The BOM is adjusted. The downstream schedule shifts. The supervisor reads about it in the morning report.

A solder paste printer drifts 18 microns. The system catches the drift, adjusts the next print, and logs the deviation against the upstream lot. No one is paged. Nothing stops.

A customer’s forecast doubles overnight. The MES re-plans the line, signals the SMT machines for a changeover, queues the new BOM, and notifies procurement. The plant manager sees the new plan when she walks in.

That’s a dark factory. Not the absence of light. The absence of friction.

The robots are the easy part. The machines have existed for years. What didn’t exist — until very recently — was the software spine that could orchestrate them in real time, learn from their data, and make decisions faster than humans could intervene.

That spine is the MES. And almost every MES on the market today cannot do it.

Why This Is Suddenly Possible

For twenty years, dark factory was a promise that kept slipping. The reason it’s now real comes down to four shifts that landed at roughly the same time:

Machines got loud. A modern SMT line emits more telemetry per hour than a 2015 line emitted in a week. Paste height per pad. Placement offset in microns. Reflow profiles per board. The data exists, in volume, every second.

Protocols matured. OPC-UA, MQTT, SECS/GEM, REST — pulling data from a machine no longer requires a six-month custom integration. The plumbing is finally standard.

AI crossed a threshold. Anomaly detection on time-series data, vision models for defect classification, natural-language interfaces over operational data — all of this stopped being research and became commodity engineering in the last 36 months.

The hardware got cheap enough. Edge compute that can run inference at the line costs less than the conveyor it sits next to.

Every input the dark factory needed is now sitting on the table. The only thing missing in most factories is a system that can pick up the inputs and do something with them.

That system is the MES. Specifically, it’s an MES whose architecture was designed to consume real-time event streams, run AI models on them, and orchestrate the line based on the output.

Most MES platforms on the market were not designed this way. They were designed to be transactional databases with a workflow engine and a dashboard layer on top. They are very good at recording what happened. They are structurally incapable of deciding what to do next without a human in the loop.

This is the gap. And this gap is the difference between a factory that gets to the next decade and a factory that doesn’t.

The Buzz vs The Real Change

There’s a fair question hiding under all of this:

Is “AI in manufacturing” actually a shift, or is it just the latest buzzword cycle?

I’ll be direct. Most of what’s marketed as “AI in MES” right now is buzz. A chatbot bolted onto an old dashboard. A “predictive analytics module” that’s a regression model from 2018 with a new logo. A “smart” feature that’s three if-then rules in a trench coat.

But underneath the noise, three things are genuinely changing:

Decisions are moving from humans to systems. Not all decisions. Not the strategic ones. But the second-by-second operational decisions — route this board, hold this lot, adjust this parameter — are increasingly being made by software. This is real. This is shipping. This is what dark factory is.

The data architecture is collapsing. For years, factories had a database for production, a database for quality, a database for maintenance, a data warehouse for reporting, and an AI sandbox for experiments. The new architecture is one event stream that everything reads from. This is real. This is what makes the AI work.

The interface is leaving the screen. Operators in modern factories are talking to the MES, not clicking through it. Voice, gesture, hands-free. This sounds gimmicky until you’ve watched an operator on a live line try to navigate a 14-click menu while holding a board. This is real, and it’s where I’ve personally been spending a lot of time.

The first two are happening with or without your participation. The dark factory is not a vision. It’s a destination half the industry is already moving toward. The MES decision you make in the next 18 months decides whether your factory is on that road or watching it from the side.

What This Means for the Decision in Front of You

If you’re a manufacturing leader sitting on an MES decision right now — and a surprising number of you are — the framing has changed.

Five years ago, the question was: which MES gives me the best traceability and OEE reporting?

Today, the question is: which MES gives me the foundation for a factory that, three years from now, can run without me having to be in the building?

Those are different questions. They have different answers. And the vendors who win on the first question often lose on the second one — because the architecture that made them efficient at recording the past is the same architecture that makes them incapable of deciding the future.

The legacy MES you’re being shown was designed for the first question. It was a reasonable answer in 2010. It is the wrong answer in 2026.

Picking it doesn’t feel wrong on the day you sign. The board approves. Procurement is comfortable. The references are strong. The mistake is invisible.

The mistake announces itself three years later, when the factory across town runs three shifts on a system that learned how to run them, and yours is still asking the consultant for a change order.

What’s Next

In Part 2, I’ll get specific about what separates a real AI-native MES from one with AI on the brochure — the architectural test, in plain language, that any leader can apply.

In Part 3, I’ll give you the seven questions to ask any MES vendor sitting across the table from you. The questions are short. The answers are everything.

For now, sit with this:

The dark factory isn’t a future. It’s a sorting event. Some factories are walking toward it on purpose. Some are walking away from it without realizing.

The MES you pick decides which group you’re in.

Lava writes about manufacturing technology and the operating systems of modern factories. Find him at imlava.in. Part 2 drops next week.