When something gets cheap, there's a lot of value right next to it
Where value seems to go when technology compresses a layer of your business. I've watched it from inside the early internet, the cloud at Google, and now AI. If you're building with AI and wondering where you still matter, I hope this is a useful map.
Every time a technology makes something cheaper, something else gets more valuable. Usually not the thing that got cheap, but the thing next to it.
I call it Compression Theory, and the shape of it is simple enough. A wave compresses a layer of business and makes it cheap or free, but the value doesn't vanish. It migrates to the adjacent layer that didn't get compressed: expertise, judgment, design, relationships, taste, trust. There are still huge winners inside the compressed layer, the few players who end up owning it as infrastructure, but that's a small club with deep moats. For most people, the room to win is in the layer next to it. And it tends to cascade, because each compressed layer becomes the engine for the next one, usually faster than the last.
Right now AI is compressing the creation of software. Writing code, the part that used to be the barrier, is being commoditized in real time. So the value seems to be moving up, toward the stuff AI can't really do for you: figuring out what to build, who it's for, and why. That part doesn't get easier just because the code does.
That's the conclusion, or at least the pattern as I've seen it.
I've watched it happen three times
The internet compressed distribution. The moat used to be presses, towers, stores, a sales force, and the internet drove all of it toward zero, so the value moved onto whatever was worth distributing: content, products, brands, commerce. (I caught a piece of this at CERFnet, helping wire up some of the first commercial peering between networks.)
Open source compressed software. With the LAMP stack, the price of that whole layer fell to near zero just as the dotcom crash made everyone cost-conscious, and the value moved to trust: the code was free, but Red Hat got paid to vouch for it. A preview, I think, of what AI is doing now.
Cloud compressed the operation of infrastructure. It got so cheap to run that everyone used more of everything, and the value moved to focus: you could put your attention on your actual business instead of running data centers. One layer didn't really compress though, which is the interesting part: IP addresses. They're a finite resource, so while everything else got cheaper, their price climbed for years. IPv6 will fix that, but it's not a simple move.
AI is compressing software creation
LLMs are compressing language. The killer thing is that code is language that runs, which is why these models are so good at it. Anyone can vibe-code a small app now, and skilled people can spec a serious agentic system in a fraction of the time a strong team used to need.
So the value moves to everything that happens before the first prompt: domain expertise, market understanding, customer relationships, design, pricing, sales. AI can help with all of it, but it can't really do it for you, because none of it is a language problem. That's what I've been calling the thought layer. And that's product management. A plan rarely survives contact with the enemy, and a business plan doesn't survive contact with real customers either. The customer comes first. That was always true, it just matters more now.
None of this means competing in the compressed layer is a dead end. Whoever ends up owning it does great, the way the ISPs and the data centers did in the last two waves. OpenAI, Anthropic, and Google are going to do very, very well. It's only a dead end if your name isn't Sam, Dario, or Sundar. For the rest of us, that layer is already crowded and the moats are already dug, which is the whole reason I keep pointing next to it.
There's a wildcard. People like Yann LeCun, working in very deeply technical arenas, may yet disrupt this industry again. But that's more evidence of cascade than a counter-argument, and it's kind of the point. Does a world model compress the thought layer? I still think not, but time will tell.
Shipping software is shipping trust
Something tends to get skipped in all the "idea to demo in a weekend" talk: shipping software is shipping trust. An untrained person can produce a working product now, but that doesn't make it wise to ship something complex, with real consequences, when you don't really understand how it works. Would you trust vibe-coded tax filing software? Air traffic control? The system that holds your money? I'm just saying, don't confuse the ability to produce a thing with understanding what it does, and don't mistake either one for a real dev team. The build became cheap. You're still responsible for what you ship.
The prediction
If the pattern holds, and it has for thirty years, AI-generated code becomes infrastructure. It mostly already is, and for the people building on top of it, the tool itself isn't the moat. AI-as-infrastructure then enables the next cascade: professional services, analysis, education, creative production, all probably about to get cheaper, each with an adjacent layer where the value ends up. The thought layer is the one I focus on, because that's my background, but I doubt it's the only uncompressed ground left. Trust, judgment, creativity, relationships, accountability, physical presence. Maybe they all collapse into the thought layer. Maybe they don't. That's up to you to decide.
The question is never whether you can do the thing that just got cheap. It's whether you own the thing that just got valuable.
One last thing
This isn't about intelligence. This is about language. It's easy to confuse, because language is how we experience intelligence externally. These AI tools are language experts, and that's amazing enough on its own. It's not super-intelligence...not yet at least. And until then, you still own the thought layer. That's the point.