Imagine opening a coffee shop where every cup costs you four dollars to make, you give it away free to build a habit, and your plan for profitability is that one day a law firm down the street will buy espresso machines from you in bulk. That’s roughly the shape of consumer AI right now. The product is delightful. The spreadsheet behind it is not.
I build bots for a living. That means I spend a lot of time staring at the least glamorous number in this entire industry: cost per interaction. And after reading TechCrunch’s piece on the ugly economics of consumer AI, I keep coming back to the same thought — the fun part and the sustainable part are two different businesses wearing one hoodie.
Consumer AI is having a moment again
TechCrunch argues consumer AI is making a comeback, and the evidence is hard to wave off. Meta’s personal assistant Muse, along with its plush-like mascot Jolly, turned into a surprise hit. OpenAI shipped Dots. People are using these things, talking about them, and forming the kind of daily habits product teams dream about.
That’s genuinely good news for anyone building in this space. Consumer traction means the interface problem is getting solved. Normal people now understand what to do with a chat box, an assistant, a character that remembers them. Five years of awkward onboarding flows finally paid off.
The problem is that traction and sustainability aren’t the same metric, and consumer AI keeps confusing one for the other.
Why the model strains
The economic challenge TechCrunch highlights is blunt: consumer AI struggles to grow without enterprise revenue underneath it. Enterprise money is the thing that makes the consumer side survivable. That’s not a footnote. That’s the business model.
Over on Hacker News, discussion of Ed Zitron’s argument lands in similar territory — that AI currently requires enormous financial complexity and volume to work at all, and that this isn’t sustainable in the medium term. You can disagree with how hard he pushes it. The underlying observation is just arithmetic.
And the pressure doesn’t stop at inference bills. There’s a parallel conversation about data centers and their impacts, plus a regulatory layer forming around them. A Colorado bill that took effect in February 2026 regulates “high risk” AI in sensitive sectors like hiring, housing, lending, and healthcare. Compliance is a cost. Infrastructure is a cost. Neither shows up in a demo video.
Reuters Breakingviews walks through the good, bad, and ugly of AI’s economic impact, flagging a risk raised by Oxford economist Carl-Benedikt Frey that overshadows the rest. Stanford’s Chad Jones, in a June 2026 conversation published by Stanford Graduate School of Business, maps possible economic futures ranging from abundance to apocalypse. When serious economists describe a range that wide, it means nobody has the unit economics nailed down yet.
What this means if you’re the one building
Here’s where I get practical, because this is the part I actually live in. If the biggest, best-funded consumer AI products can’t make the math work on free users, a solo builder or small team absolutely can’t. So design like cost is a first-class constraint, not a thing you optimize later.
- Know your cost per conversation. Not per token. Per conversation, averaged across real usage, including retries and the user who sends forty messages at 2am. If you can’t say that number out loud, you don’t have a product, you have a demo.
- Route by difficulty. Most requests in a typical bot don’t need your most expensive model. Classification, routing, and simple lookups can run on something small. Reserve the heavy model for the queries that earn it.
- Cache aggressively. Users ask the same things. Embeddings, retrieval results, common responses — all of it can be cached. This is boring engineering that directly improves your margin.
- Build the business-facing version early. If enterprise revenue is what subsidizes consumer delight at the top of the market, that pattern scales down. The B2B version of your bot — the one with audit logs, admin controls, and a contract — is what funds the fun one.
- Treat regulated domains carefully. If your bot touches hiring, housing, lending, or health, the Colorado rules are a signal about where this is heading. Architecture decisions made now determine whether compliance is a weekend or a rewrite.
The uncomfortable read
Muse, Jolly, and Dots prove consumer appetite is real. That was the open question two years ago, and it’s basically answered. What’s still open is whether anyone can serve that appetite at a price that works without a much larger business standing behind it.
For builders, that’s not a reason to stop. It’s a reason to be specific. Pick a narrow problem where the value per interaction clearly exceeds the cost per interaction, charge for it, and let the companies with infinite balance sheets fight over free assistants. The economics are ugly at scale. At small scale, with discipline, they can be perfectly reasonable.
🕒 Published: