Remember the smart speaker era? Hardware sold near cost, voice assistants given away, every company convinced that owning the kitchen counter would eventually print money. The assistants got better. The business never arrived. We talked about engagement because engagement was the only number that looked good.
I thought about that this week reading TechCrunch’s piece on the ugly economics of consumer AI. Meta’s personal assistant Muse and its plush-like mascot Jolly have been a surprise hit. OpenAI’s Dots just shipped and is picking up users fast. By the usual measures, consumer AI is back. And the usual measures are exactly the problem.
Why consumer agents cost more the better they get
As someone who spends most of her time looking at agent architectures rather than growth charts, the thing that bothers me about this comeback is structural, not financial. In classic consumer software, marginal cost per user trends toward zero. You write the code once, serve it a billion times, and your infrastructure bill grows far slower than your user count. That asymmetry is the entire reason consumer internet businesses work.
Agent architectures invert it. Every meaningful improvement in an assistant’s behavior adds compute per interaction:
- Longer context. An assistant that remembers you is an assistant carrying more tokens into every forward pass. Memory is not free storage; it is recurring inference cost on every turn.
- Tool use and planning loops. One user request becomes several model calls — plan, call, observe, revise. A single sentence from a user can fan out into a dozen inferences.
- Multimodality. Voice and images are orders of magnitude more expensive per unit of meaning than text. A cute plush interface is, computationally, an expensive interface.
- Reliability work. Verification passes, self-critique, retries on failure. The techniques that make agents trustworthy are the techniques that multiply token spend.
So the better the product gets at the thing users actually want — persistence, autonomy, initiative — the worse the per-user cost curve looks. Engagement, the north star of consumer products, is also the cost driver. A power user is a liability in a way they never were at a social network.
The enterprise crutch
The industry’s answer, as TechCrunch notes, is enterprise revenue. Consumer products struggle to grow sustainably without it. That is true, and it is also a quiet admission about where the value density sits.
Enterprise buyers tolerate per-seat pricing that reflects real inference cost because the output replaces labor with a known price. Consumers do not. Consumer willingness to pay for software has been anchored for fifteen years at somewhere between free and the price of a coffee. You cannot run a multi-step agentic loop with a long memory window on a subscription priced against that anchor, not without heavy quantization, aggressive caching, smaller routed models, and throttling — all of which degrade exactly the behaviors that made the product a hit.
Which gives us an uncomfortable shape: consumer AI as a demand-generation and data-gathering layer, subsidized by enterprise contracts. That is not a business model. That is a loss leader with very good PR.
The sustainability argument deserves engineering attention, not just skepticism
Ed Zitron’s broader critique — that the financial complexity and volume AI requires is not sustainable in the medium term — gets dismissed in my circles as doom-posting. I think that is lazy. You can disagree with his conclusions and still accept his framing as an engineering constraint worth designing against. Chad Jones at Stanford has been mapping the range of economic futures AI might produce, from abundance to something much darker, and the spread in those scenarios is wide enough that confidence in any one outcome is unearned. Reuters, covering the good, bad, and ugly of AI’s economic impact, flags Carl-Benedikt Frey’s point that the risks around how a technology diffuses can overshadow the more obvious drawbacks.
Meanwhile the physical layer is getting its own scrutiny. Data center siting is becoming a political question, and regulation is arriving in specific sectors — a Colorado law in effect since February 2026 governs high-risk AI in hiring, housing, lending, and healthcare. Compute is no longer an abstraction you provision with a credit card. It has neighbors, permits, and now statutes.
What I would actually watch
Not daily active users. Watch inference per session, and whether it is falling as these products mature. Watch whether anyone ships an agent that gets cheaper per useful action over successive releases — through better routing, smaller specialized models, cached reasoning, or on-device execution. That curve, not the download chart, tells you whether Muse and Dots are businesses or demos with excellent retention.
The constraint is not a reason to stop building consumer agents. It is the most interesting design problem in the field right now, and almost nobody is treating it as one.
🕒 Published: