Free the Data

On a factory floor, connecting a 15-year-old machine is the easy half. Getting it to speak is the last mile nobody budgets for.

There’s a comforting story about AI on the factory floor: connect the machines, pipe the data to a model, and let the intelligence do the rest. Spend time in an actual plant and you learn how much hides inside that first word - connect - and how much more hides after it.

Walk the manufacturing floor at Hypertherm, a company that has been building precision cutting systems in New Hampshire for decades, and you find equipment that has run reliably far longer than most software has existed. Rob Kay, who runs IT operations there, describes the reality plainly: “Some of these devices are really old. I recall showing one of the Ramen engineers equipment connecting through a serial port - serial ports haven’t been around for 15 years. We worked through that very successfully, got the devices connected, and showed a consistent quality of the network connection, which is something we’d really struggled with.”

That first problem - physical connectivity - is real, and it is solvable. With the right gateway you can bridge a serial port, an ancient controller, or a one-off interface onto a modern network. It takes work, but it is a known kind of work.

Then comes the problem almost no one budgets for.

Once the machine is connected, it is still speaking a language you cannot read. Industrial equipment built over the last several decades communicates in old, proprietary protocols - data formats defined by the original manufacturer, often undocumented, sometimes understood only by the vendor who built the machine. The result is a strange and expensive kind of failure: the machine is online, the packets are flowing, and the plant still cannot tell what any of it means. The data you connected the machine to get - cycle times, fault codes, quality signals, utilization - is trapped inside, encoded in a dialect no one on site can interpret.

Networking the machine takes hours. Learning its language can take weeks.

This is the part of the last mile that gets missed, because it doesn’t look like an infrastructure problem. It looks like a data problem, or a vendor problem, or something to be quietly worked around. But it is exactly an infrastructure problem, and it is the one standing between a factory and all the AI it has been promised. Physical AI is worthless without data. A model that cannot see the floor cannot run it. And in the real world, most of that data isn’t missing - it is stranded, locked inside equipment that was never designed to give it up.

Getting it out is where AI finally earns its place - not the AI of the demo reel, but agentic AI aimed at the unglamorous, high-value job of interpreting these proprietary machine languages and translating them into something a business can use. An agent that can learn a machine’s dialect, make sense of its stream, and hand back clean, structured data does something no amount of raw connectivity can: it makes a silent machine speak.

That reframes the last mile of Physical AI. It was never only physical. Connectivity gets a machine online; interpretation makes it legible. Both are infrastructure, and both have to be solved before a single useful model can run on the floor. A plant can be fully wired and still completely dark.

This is the layer we build at Ramen: the infrastructure that connects the un-connectable and frees the data locked inside it, so the physical world becomes legible to AI, machine by machine, protocol by protocol. The intelligence everyone is excited about only pays off if it can actually see what is happening on the ground.

The factory floor is full of machines that have quietly done their jobs for fifteen years, holding decades of data nobody can read. The opportunity isn’t to rip them out. It’s to free the data.

^{Field Notes from the Last Mile is a running series on what Physical AI actually takes to deploy in the real world - from the people building it. Subscribe to get the next one.}