There's a pattern running through the most interesting technology news right now and most people are missing it because they're looking at the shiny part. Look at the substrate instead. Look at what the new thing is actually made of, or what it has to pretend doesn't exist to function. That's where the real risk lives. That's also where the real opportunity hides.
Recycled junk is the new ore
Start here: OnePlanet says it can extract silicon above 95% purity from end-of-life solar panels, recovering ore-grade material from what was headed for a landfill. The panels were considered waste. Turns out they're feedstock. The raw material for the next generation was buried in the previous generation's garbage.
That's not just a materials science story. That is the story of this moment in technology, told in silicon and copper and aluminum.
Meanwhile, Sungho Bae is doing something similar in sculpture, pulling apart stuffed animals and found objects to reconfigure them into something uncanny and alive. The source material is discarded. Familiar to the point of invisibility. But the recombination reveals something the original parts never had. The junk had structure in it the whole time. Nobody looked hard enough.
That's two stories from completely different domains pointing at the same idea. Pay attention.
The substrate problem in AI
Now look at what's happening inside the AI systems everyone is selling you right now. Birgitta Böckeler's experiments at Thoughtworks on TDD inside AI agent loops reveal something uncomfortable: the discipline that actually keeps software from becoming a pile of unverifiable garbage doesn't automatically get better just because you wrap an LLM around it. The substrate still matters. The tests still matter. The process still matters. Telling an AI agent to "use TDD" doesn't magically produce the benefits of TDD any more than labeling solar panel waste as "ore" produces ore. You have to actually do the work of extracting value from the underlying material.
We've watched plenty of teams get seduced by the surface behavior of AI coding agents, the confident output, the fast iteration, the apparent competence, while the foundation underneath is quietly rotting. No coverage. No assertions that catch the edge cases. Just vibes and speed. The shiny system is built on junk it hasn't processed yet.
The platform control problem is the same problem
Suno just capped Pro subscribers at 20 song downloads a month while simultaneously announcing future models developed "in partnership with the music industry." That's a polite way of saying: the thing you were using freely is now under someone else's constraint structure, and the terms will keep changing as the interests of larger parties assert themselves. The substrate of Suno's outputs, the training data, the distribution rights, the commercial agreements, was always built on top of a negotiation the platform had to have eventually. Subscribers were using a service priced as if that negotiation would never arrive.
If you're building a music business, a creative studio, an AI-assisted production workflow, or anything that touches these platforms, you already know where this goes. The substrate doesn't stay invisible forever. It resurfaces, usually in a terms-of-service update, usually at the worst possible time.
And if you think the cybersecurity layer of AI is any different, look at OpenAI and AWS moving Daybreak's cybersecurity capabilities into Amazon Bedrock for enterprise workflows. The capability is real. But it's now layered inside a cloud provider relationship, an enterprise licensing arrangement, a platform dependency. The thing securing your infrastructure is itself dependent on infrastructure you don't control. Substrates all the way down.
What this means for you, specifically
If you're running a real business, not a VC narrative but an actual operation with customers and revenue, the pattern here is the thing that has gotten founders in trouble since before we started paying attention to this industry. You adopt a platform at the price point where it feels free or nearly free. You build process around it. You let the substrate of your own operation become dependent on a layer you don't own. Then the platform matures, the negotiation arrives, the terms change, and you're stuck.
The OnePlanet insight is that waste has value if you're willing to do the work of extraction. The Sungho Bae insight is that familiar materials recombined deliberately produce something new. The TDD-in-AI-agents insight is that process discipline still matters even when the surface looks different. The Suno story is a warning about what happens when you mistake platform access for a durable foundation.
Put those together and you get an operating principle: the founders who win in this environment are the ones who are clear about what they actually own versus what they're renting access to, and who apply real discipline to the layer they control instead of outsourcing that discipline to a platform and hoping for the best.
The move right now
Audit your stack. Not your tech stack specifically, though that too. Your dependency stack. Where is your business built on top of something you don't control? Where are you assuming that the terms, the pricing, the access, the model behavior, the download limits will stay the same? Every one of those assumptions is a solar panel that needs to go through OnePlanet's process before you find out what's actually in there.
Build the tests. Own the output. Control the distribution where you can. Not because platforms are evil but because the substrate always resurfaces. Every system is eventually made visible by the negotiation it was trying to avoid.
The junk you're sitting on might be ore. But you have to actually process it. Nobody's coming to do that part for you.