When Execution Is Free
AI is collapsing the cost of building software. The same logic might extend to hardware, biology, everything. A speculative exploration of where execution approaching zero might lead.
I can describe an app and have it built in minutes.
Not a toy app. A real thing. Database, authentication, the works.
Five years ago this was science fiction. Now it’s a Tuesday.
So what if we extend that logic?
Software execution is nearly free. What happens when all execution is free? Hardware. Biology. Everything.
A wildly speculative exploration.
Software Is (Almost) Solved
We’re already living the early version of this.
The gap between “I want X” and “X exists” used to be defined by labor, cost, skill. You needed engineers. You needed time. You needed money.
Now? A good prompt and some patience.
The numbers are absurd. A Google engineer reported that Claude built in one hour what her team spent a year developing - a distributed orchestrator. Cursor’s agents ran for seven days straight and output over a million lines of Rust - a browser from scratch with a custom rendering engine.
NanoGPT training time: under 2 minutes and falling.
Approaching zero. So what does this world look like?
Hyper-personalization. You don’t download the same app as everyone else. Your AI generates a tool specifically for your workflow. Your interface looks nothing like mine.
Disposable software. Build a complex data-analysis tool to answer one specific question. Use it for five minutes. Delete it.
Most code being written today will run zero to one times.
That’s a complete inversion of how software has worked our entire lives.
The shift from “how” to “why.” When execution is free, the value of a person isn’t their ability to code. It’s their taste and their vision.
Wang Cong, a Linux kernel maintainer, put it well: “Correctness is becoming cheap. Competence is being commoditized. But there’s something AI cannot do: tell you whether something should exist. That requires taste.”
He rejected an AI-approved kernel patch. The code was correct. But the design was wrong in ways that only matter long-term. Bloat. Wrong layer. Symptom vs root cause.
The kind of judgment that comes from studying great systems and watching bad ideas fail.
Taste as currency.
If anyone can execute perfectly, the only differentiator is what you choose to make.
The Hardware Bottleneck
Atoms are heavy.
Software iterates at the speed of light. Physical stuff is slow. Energy and compute still cost real money. Materials have supply chains.
Your digital life feels magical. Your physical life is still physical.
US AI capex is now 1.9% of GDP. Three times the Apollo program. Five times the Manhattan Project. Just on infrastructure.
And it’s still not enough.
There’s a thing called the Memory Wall. Imagine the world’s fastest chef, but the ingredients arrive by carrier pigeon. You spend most of your time standing at the counter, waiting.
That’s modern AI hardware. The thinking happens in nanoseconds, but the data has to physically travel across the chip. That journey is the bottleneck.
DRAM prices jumped 55% this quarter. Meta is planning 6.6 GW of nuclear power by 2035 just to run their models.
This is the anchor. The constraint.
Software scales instantly. Building new chips, new memory, new power plants - that takes years.
Then again.
What If Hardware Gets Solved Too?
The most advanced version of this is called molecular nanotechnology.
Instead of carving objects or layering plastic, you have assemblers. Billions of tiny robotic arms that build things atom-by-atom.
Star Trek replicator vibes.
You don’t buy a phone. You download the hardware code and a box in your house organizes carbon, silicon, and metal atoms into a functioning device.
We’re nowhere near this. But the concept isn’t new. Eric Drexler wrote about it in Engines of Creation decades ago.
His argument: biology is the proof of concept.
Your body is already a molecular factory. Turns raw materials (food) into complex hardware (cells, muscle) using code (DNA).
If this works, the world becomes a sandbox game:
- No supply chains. You only ship raw elements and energy.
- Break a plate? Recycle the atoms, print a new one.
- The only limits left: energy and information.
Programmable matter takes it further. Instead of a static phone, you have a handful of smart dust that snaps into whatever shape you need.
Screen. Headset. Hammer.
The distinction between software and hardware disappears.
MIT’s Center for Bits and Atoms is building what they call “compilers for the physical world.”
That’s wicked.
The Bridge: Physical AI
We don’t have to wait for molecular assemblers.
There’s a bridge being built right now.
Think of ChatGPT as a brain in a jar. Physical AI is that brain getting a body.
Two pieces matter:
World models. AI that doesn’t just predict text. It has an internal 3D simulation of physics. It understands that if you drop a glass, it breaks. If you push a door, it swings.
The AI can rehearse physical tasks millions of times in its head before moving a finger.
VLA - Vision-Language-Action models. Think about how you catch a ball. Your eyes see it, your brain predicts where it’s going, your arm moves to meet it. All in one smooth flow.
VLA models are trying to do the same thing for robots. Connect seeing directly to doing, without clunky translation steps.
NVIDIA’s GR00T. Figure. Boston Dynamics. Tesla Optimus.
The field went from “let’s try LLMs on robotics” to a whole taxonomy of approaches in 18 months.
And it’s already deploying. Harmattan AI in France is building capacity for 10,000 defense drones per month. Wing is expanding drone delivery to 150 Walmart locations, reaching 40 million Americans. Tesla FSD navigated San Francisco during power outages. No traffic lights, no problem.
One example captures this.
Researchers at MIT wanted to let AI stylize 3D models. Add textures, decorative patterns, artistic flourishes.
The problem: only 25% of the stylized objects actually survived printing.
The AI was designing gorgeous things that couldn’t exist in reality. Like drawing a staircase that collapses under its own weight.
So they gave the AI a physics model. Let it simulate stress and strain as it designed. Let it check its own work against reality.
Result: 100% structural viability. No loss in aesthetics.
The AI learned to dream within constraints.
Give AI a world model. Let it iterate in simulation. Output works in reality.
The path from “software is solved” to “hardware is solved” runs through here.
The Wild Card: Health
If we’re playing out the logic, biology is just another execution problem.
But to understand what changes, you have to feel what exists now.
Let’s take chemotherapy.
You sit in a chair while poison drips into your arm. The poison kills fast-dividing cells. Cancer cells, yes. But also your hair follicles. Your gut lining. Your immune system.
You count the days between treatments. You watch your hair collect in the shower drain.
The treatment works, sometimes, by being slightly more lethal to the disease than to you.
Now imagine the alternative.
The AI has a simulation of your specific body. Not a generic human. You. It runs the tumour forward in time, finds the molecular signature that makes those cells different from healthy ones, and designs a targeted intervention that goes to one specific coordinate and deconstructs the problem.
No collateral damage. No hair in the drain.
We’re not there. The gap between current medical AI and a “world model of your body” is enormous.
But the logic extends.
If execution is free for software, then for hardware, then for biology, healthcare stops being about treatment and becomes optimization.
What’s Already Happening
All of that sounds far off.
But the shift has already started.
Developers are reporting significant increases in output. People are vibe-coding entire applications in weekends. The shift from writing code to supervising AI writing code is happening now.
Jobs don’t get “replaced” dramatically. They evaporate quietly.
You get fewer emails. Fewer tasks. Fewer requests. At first it feels like a slow week. Then a slow month.
Until one day, the role still exists, but only on paper.
244,851 tech jobs cut in 2025.
The success rate for cold applications is now under 1%.
Stack Overflow volume decayed to pre-public levels. The questions stopped because people stopped needing to ask.
The AGI doesn’t arrive like a movie villain.
It arrives like a software update.
The Video Game Question
If execution is free - both digital and physical - life becomes creative mode.
The danger is loss of resistance.
If every problem can be solved by prompting, where does purpose come from?
The mountain you spent years climbing is now a button.
The skill you mastered is now a commodity.
The thing that made you you. The struggle that shaped you.
What happens when struggle becomes optional?
Maybe we find new walls to climb.
Maybe the shift from “how” to “why” applies to life itself. When you can do anything, the only question that matters is what’s worth doing.
Or maybe it’s a crisis without a clear resolution.
No one knows where this goes.
The speculation about hardware and biology might never pan out. The constraints of physics might hold forever.
But the software piece isn’t speculation anymore. That’s happening.
The question is just how far the logic extends.
If you could print anything tomorrow - what would it be?