Thoughts — an opinion, not a how-to.
I'll admit it: I don't find AI all that interesting right now. Not because the technology can't do anything — it can do a remarkable amount. But because we're all using it at a price that isn't the real price. And as long as that's true, we're watching a performance, not the business behind it.
Every prompt you send today is subsidized. The actual cost of compute, electricity, and the depreciation of enormous data centers isn't on your invoice — it's on the balance sheets of venture investors and hyperscalers. You're paying an opening-week ticket. Someone else is paying full price for now.
How wide that gap is, I've taken apart with the data elsewhere — see “A Tool, Not a Miracle”. The short version is enough here: by late 2025, OpenAI had signed compute commitments on the order of $1.4 trillion — against annual revenue of roughly $20 billion. That's about 70 to 1 between what's promised and what comes in. The customer isn't the one covering that difference today.
That sounds like a detail for investors. It isn't — it devalues almost everything we think we currently know about AI. Every ROI calculation, every "this saves us X per month" slide, every business case rests on a price that won't last. We're testing product-market fit at a rate that can't survive sustained operation. It's like measuring a restaurant's viability by how full it is while the food is free.
That's exactly why the interesting moment hasn't arrived. It comes once the bill reaches the customer — when providers need margins and have to pass the real costs through. That's when the genuinely interesting thing happens: most "let's sprinkle a bit of AI everywhere" applications won't survive full price. A few will — the ones where the value is so clear and so large that it carries the real price too. Which ones those are is the only question that really matters. And it simply hasn't been answered yet.
So my position isn't AI skepticism. I don't doubt what the models can do — I'm waiting for the price tag. Because I want to build on the version that survives contact with its own costs, not on a promotional rate. Anything that works at a subsidized tariff and tips over at full price isn't a foundation. It's a bet that someone else keeps paying the difference forever.
This is explicitly not doom-mongering. Whether all of this is a bubble is genuinely contested — and secondary to the question I care about. The capability is real. The unanswered part isn't "can it?", but "at what price, for what task, profitably?". As long as that's open, the euphoria is the providers' business model — not yours.
Until then, I'd rather build the boring stuff: clear processes, clean data, automation that still pays off when nobody is giving away half of it. AI goes to the place in the chain where its value beats the real price — not to the front, just because it shines.
AI gets interesting at the exact moment it has to pay for itself.
