The Robot Data Gold Rush Is Here, and I've Seen This Movie Before

That's interesting! That's genuinely interesting. Because these aren't venture capitalists pattern-matching to the last successful pitch deck. These are industrial companies with actual robot programs who presumably understand what they need.

Or maybe they're just hedging their bets. Call me old-fashioned, but I've learned to be skeptical when corporate venture arms start writing checks.

The timing isn't coincidental either

Here's what's changed in the last eighteen months: AI coding agents got good. Really good.

I spent a weekend recently reading about WIRED's experiment giving an AI agent a physical robot body, and the implications are sort of staggering. The argument is that AI models have gotten so capable at writing and debugging code that the software side of robotics, which used to be the hard part, is becoming almost trivially easy for certain applications.

Which means the bottleneck is shifting. If you can spin up robot control software in an afternoon, what's left? Hardware (still hard, still expensive) and data (increasingly the thing everyone needs and nobody has enough of).

This is the self-driving car hype cycle all over again, except maybe this time the companies have learned something. The autonomous vehicle industry spent a decade and roughly $100 billion learning that the last 1% of driving scenarios requires 99% of the data collection effort. Waymo won (sort of, in limited geographies) not because they had better algorithms but because they had more miles of real-world data than anyone else.

Config seems to be betting that robot manipulation will follow the same pattern. The first 80% of tasks will be easy. The last 20% will require data that's incredibly expensive to collect. And whoever has that data wins.

The part where I express some doubt

I should be clear that I don't actually know if this is going to work.

The TSMC comparison is flattering, but it glosses over some important differences. Semiconductor fabrication has massive economies of scale, you need billion-dollar fabs, so consolidation makes sense. Robot training data might not work the same way. Maybe every factory needs slightly different data. Maybe the data doesn't transfer well between robot platforms. Maybe the whole premise is wrong and robots will learn better from simulation than from real-world data collection.

There's also the question of defensibility. TSMC's moat is physical, you can't just copy a cutting-edge fab. Data moats are trickier. If Config collects a million hours of manipulation data, what stops a well-funded competitor from collecting two million hours? What stops the robot companies themselves from eventually bringing this capability in-house once they understand what they actually need?

I don't have answers to these questions. Neither does anyone else, as far as I can tell. The Config team believes that data collection requires specialized expertise that will be hard to replicate, and that being first will create network effects as more robots train on their data and more edge cases get captured. Maybe! But I've heard similar arguments from companies that no longer exist.

So what

Here's why I think this matters even if Config specifically doesn't pan out.

The robotics industry is starting to disaggregate. For decades, robot companies were vertically integrated by necessity, you built the hardware, wrote the software, collected your own training data, and sold a complete system. That model is breaking apart. You've got companies specializing in actuators, companies specializing in manipulation software, companies specializing in simulation, and now apparently companies specializing in data.

This is what happened to computers. It's what happened to smartphones. It's arguably what's happening to AI models right now, with the training compute layer separating from the inference layer separating from the application layer.

When an industry disaggregates, the companies that control the bottleneck layers tend to capture most of the value. Intel captured the PC era. TSMC captured the mobile era. Nvidia is capturing whatever era we're in now.

The bet Config is making, and the bet that Hyundai and Samsung and LG are apparently making alongside them, is that data will be the bottleneck layer for robotics. Not hardware, not software, but the messy, expensive, tedious work of teaching robots how the physical world actually behaves.

I've been wrong before about these things. I was skeptical about cloud computing in 2008 (why would anyone trust their data to someone else's servers?) and I was skeptical about smartphones replacing computers in 2010 (the screens are too small!). The kids building robots today will probably look back at this column and laugh at whatever I'm getting wrong.

But I've also been right about a few things, and my pattern-matching brain keeps telling me that data infrastructure companies tend to matter more than anyone expects. The boring picks-and-shovels plays often outlast the flashy gold miners.

Config might be the TSMC of robot data. Or they might be one of the hundred companies that tried to be the TSMC of something and ended up as a footnote. We genuinely don't know yet, and anyone who tells you they're certain is selling something.

What I do know is that the biggest manufacturers in Korea are taking this seriously enough to write checks. And when industrial companies start betting on infrastructure plays, I pay attention.

But what do I know. I'm just a guy who still prefers email to Slack, watching another tech cycle unfold and wondering if this time will be different.

Mark Kowalski · 5 hours ago · 6 min