When Execution Stops Being the Hard Part

A founder on a recent Invest Like the Best podcast said his firm’s annual AI spend went from tens of thousands of dollars to seven million in a few months. Not over years. Months.

That number is easy to flatten into shock. What made it interesting was his interpretation. He did not treat it like a cost blowout. He treated it like a signal.

The signal was simple: value creation is changing. Execution is no longer the main bottleneck.

The inversion nobody is saying loudly enough

For most of economic history, ideas were the cheap part. Everyone had them. The scarce part was turning them into something real. That meant hiring engineers, building systems, shipping product, managing operations, and scaling without everything breaking.

That constraint is collapsing.

The podcast gave a few examples that made the shift feel concrete. One person on the team spent a few thousand dollars in AI tokens and built a chip analysis system, overlaying materials across an entire semiconductor stack from microscope images. A former Intel engineer told him that would have been a full team’s job to build and maintain.

His economist, working alone, piped employment data, regression models, and 2,000 Bureau of Labor Statistics task categories into an analysis that he said would have taken a team of 200 economists a year.

Then there was the US power grid map. One person. Three weeks. Every plant, every transmission line above a certain voltage, micro supply and demand by region. When shown to energy traders, they said it was better than what companies with 100 people and a decade of work had produced.

The pattern is hard to miss. Execution is getting cheaper. The bottleneck is moving upstream to deciding which idea is worth pointing AI at.

Tokens are becoming the new electricity

The spend numbers from that conversation are worth sitting with. Seven million dollars a year on AI tokens. Twenty-five million in salary costs. At the current trajectory, AI spend could exceed salary spend by the end of the year if the curve holds.

That sounds alarming. It is also rational.

The output per token is rising faster than the cost per token. Models get more capable. Prices for similar capability keep falling. Yet spending does not go down, because each new model unlocks work that was previously impossible or too expensive. That new work creates more value than the token bill.

That is what token economics looks like on the ground. It is not just a pricing model. It is a new kind of infrastructure. If you want the output, you need access. If you want the best output, you need good judgment about where to spend it.

I wrote about this in Running Claude Code on Your Own GPU. Local inference can make sense when you care about cost control, but the tradeoff is real. Frontier models still do things local models cannot, especially on the kind of complex tasks that matter for real business value.

The access problem nobody wants to say out loud

One part of the podcast was uncomfortable in a useful way. The best models are not equally available to everyone. Rate limits exist. Enterprise contracts matter. The speaker described being on his knees in front of an Anthropic co-founder asking for access to a model that officially did not exist yet. He eventually got it. Most people will never have that conversation.

That is the new moat.

It is not just talent. It is not just capital in the old sense. It is token access. Whoever gets the best model earliest, with the highest throughput and the fewest constraints, gets a compounding advantage over everyone else working under standard limits.

This connects to something I wrote about Tegmark's 12 AI futures. The real risk in every bad scenario is a control layer, the place where someone sits between people and their ability to act, earn, and participate. Token access is becoming one of those layers.

A friendly interface can hide a cage. Generous pricing today does not guarantee generous pricing when you are dependent.

The three things that matter now

If execution is cheap, value shifts to three things.

The first is idea selection. You can build almost anything now. You cannot build everything. Tokens cost money. Time is finite. Bad ideas executed cheaply are still bad ideas. The winner is the person who knows which problem is worth solving and which market will pay for it.

The second is direction. Knowing what to build is not the same as knowing how to work with AI to build it well. The people in that podcast who got extraordinary results were not just typing prompts. They understood the problem deeply enough to break it into pieces AI could actually solve. They knew when the output was good. Domain knowledge still matters. It just compounds differently now.

The third is capture. Building something is not enough. You have to sell it, defend it, and improve it faster than the person who will copy it next month with the same tools you used. The speaker made that point clearly. The way they did their first data product in 2023 is now what everyone else is doing. Stand still and you get commoditised.

That is also why local AI infrastructure matters more than most people think. Not because it replaces frontier models, but because it gives you a cheap place to iterate on direction and structure before you spend serious tokens on execution.

The permanent underclass line

The line that landed hardest in the podcast was blunt. If you do not use more tokens, generate value from those tokens, and capture that value, you will never escape the permanent underclass.

He was not being dramatic. He was describing compounding. The people who learn how to use AI to produce outsized economic output now are building a lead that gets harder to close every month.

The lazy version of AI adoption is using it to do less. The valuable version is using it to do more, faster, and then capturing the difference.

That gap is widening. The window to move from one mode to the other is still open, but not for long.

What this means if you are building

Execution used to be the hard part. That world is fading.

The people with use now are the ones who can decide, correctly and quickly, what is worth building. The next layer of use goes to the people who can access the best tools to execute those decisions.

The question is not which AI future is coming. These scenarios are not forecasts. They are design consequences of choices being made now.

The question is what you are pointing the tokens at, and whether you are capturing what comes out.