Consumer AI had its best week in a long time at the end of September 2026. Meta’s personal assistant Muse, with its plush mascot Jolly, was a surprise hit. OpenAI released Dots, its own cartoon-friendly personal agent, on 29 September. And Instinct, an agent that books travel, makes reservations and cancels subscriptions, had just reached a $10 billion valuation. On 30 September, TechCrunch published a column with a deliberately cold title: “The ugly economics of consumer AI”.
Its argument is that the products are getting better faster than the business is. Only a small share of people pay for AI, the amount they pay is rising slowly, and running the models is unusually expensive. That is why, in TechCrunch’s words, the industry has shifted “toward the Anthropic model, focusing on enterprise contracts and vertical-by-vertical expansion”.
This article tests that argument against the underlying data. We compare the three surveys and payment studies behind the “who pays” figures, check one calculation in the column that does not add up, look at why costs are the real problem, and set out how Meta, OpenAI and Instinct each plan to make consumer AI pay. For the companies involved, see our reports on Instinct’s $10 billion valuation and on whether Meta’s Muse bet is working.
Table of contents
- The Bull Case for Consumer AI
- How Many People Actually Pay for Consumer AI
- The Spending Curve Barely Moves for Consumer AI
- The Netflix Test for Consumer AI
- Why Costs Are the Real Problem in Consumer AI
- Three Ways Companies Try to Make Consumer AI Pay
- What the Economics of Consumer AI Mean for Users
- Consumer AI FAQ
- References and Further Reading
The Bull Case for Consumer AI
TechCrunch starts by making the optimistic case fairly, and it deserves to be stated before it is tested.
Agents that finally work
The column’s point is that “agentic AI has finally gotten reliable enough to handle everyday tasks”. The new wave of consumer AI products does not just answer questions. It books tables, buys tickets, makes calls and runs errands. Companies are now pitching that service to ordinary people, who are getting real value from it.
Three launches in one week
The evidence is the release calendar. Muse topped Apple’s US App Store; our earlier analysis found estimates of between 2.3 million and 4.3 million downloads in under three weeks. OpenAI launched Dots at DevDay. Instinct raised $1 billion at a $10 billion valuation. For an investor, TechCrunch writes, that “looks an awful lot like the ChatGPT launch in 2022”.
Why the timing feels right
Consumers are also spending more. Menlo Ventures estimates that global consumer AI spending reached $40 billion in 2026, more than three times the $12 billion it measured in 2025. On the surface, that is a market taking off.
How Many People Actually Pay for Consumer AI
This is where the argument turns. The TechCrunch column relies on a chart in Andreessen Horowitz’s semiannual State of Markets report, which draws on research by PNC. It shows that, as of May, 2.2% of consumers were paying for AI, at an average of $31 a month.
Three sources, three answers
The trouble is that the main data sources on paying users disagree by a factor of more than ten. Each measures something different, so the headline figure depends on which one you read.
| Source | Date | Method | Who pays |
|---|---|---|---|
| PNC, via a16z State of Markets | May 2026 | Payment data | 2.2% of consumers, average $31 a month |
| Bank of America Institute | March 2026 | Bank of America card and account payments | About 3% of households; up 38% on the 2024 average |
| Menlo Ventures | Survey fielded July, published 16 September 2026 | Survey of 5,067 US adults | 55% of AI users pay for at least one AI product |
Putting them on one scale
Menlo reports that 64% of US adults use AI and 55% of those users pay. Multiplying the two gives roughly 35% of all US adults paying for something AI-related, by our arithmetic. Set beside the payment studies, the gap is enormous.
Share of US adults or households paying for AI, % (Menlo figure is our arithmetic: 64% use AI x 55% of users pay = 35.2%)
Why payment data and surveys diverge
Several things probably explain the gap. Bank payment data can only count a charge it can recognise as AI; a subscription bought through Apple’s or Google’s app store, or one paid by an employer, may not show up. Menlo itself finds that 34% of payers have a family member or friend covering the cost and about 20% have an employer or school paying. Surveys, in turn, rely on what people remember and may include AI features bundled into wider subscriptions.
Who the payers are
The sources agree more on who pays than on how many. Bank of America finds that households earning more than $125,000 and younger generations make up the largest share of AI spenders, but that median AI spending grew fastest among $75,000 to $125,000 households in February, a sign of spread into the middle of the market.
Menlo finds that 77% of people in households earning $100,000 or more use AI, against 56% below $50,000. Income predicts who adopts consumer AI; reliance predicts who pays. Menlo’s payers are nearly twice as likely as non-payers to use AI daily (50% against 26%) and five times as likely to use an AI agent (65% against 13%).
What every source agrees on
Whichever number is right, the direction is the same. Bank of America finds the number of paying households up 38% against the 2024 average. Menlo finds that 46% of payers spend more than a year ago, against 12% who spend less. Paid consumer AI is growing, but from a small base, and the column’s point about the slope of the line still stands.
The Spending Curve Barely Moves for Consumer AI
TechCrunch’s sharpest observation is about the shape of the charts. Both the share of people paying and the amount they pay rise slowly and steadily, “awfully linear”, even through huge jumps in model quality.
Better models, flat willingness to pay
The column notes that “the enormous performance jump from GPT-5.2 to Astra, for instance, is barely visible on the chart”. That is a striking finding for an industry whose pitch rests on better models unlocking new demand. If a generational upgrade does not move the number of payers, product quality alone will not fix the economics of consumer AI.
Spending is concentrated in a few users
Menlo’s data shows where the money actually comes from. The typical payer spends $20 to $49 a month, but the 14% of payers who spend $100 or more account for 60% of all consumer AI spending. In other words, a small group of power users carries the market, while most paying users contribute modest, stable sums.
Who foots the bill
Menlo also asked who actually pays. Only 48% of payers cover the cost entirely themselves. Another 34% have a family member or friend paying, and roughly 20% have an employer or school picking up the tab. Those institutional payers spend more: a blended $95 a month, against $55 for people paying out of their own pocket. That matters for the economics of consumer AI, because a meaningful slice of “consumer” revenue is really business or household spending in disguise, which is exactly the bridge to enterprise sales that the labs are now building.
Users barely grew while spending tripled
The same report finds that the number of AI users worldwide grew only 11% in 2026, from 1.8 billion to 2.0 billion, while spending more than tripled. Growth is coming from existing users paying more, not from new people arriving. That is a healthy sign for engagement, but it also suggests the pool of potential payers is not expanding quickly.
Growth from 2025 to 2026, % (Menlo Ventures figures; our arithmetic: spend 40 / 12 = 3.33x, a 233% rise; users 2.0 / 1.8 = 1.11x, an 11% rise)
The a16z spin
Andreessen Horowitz puts a positive gloss on the same data. Its State of Markets II post says that, as of April, “barely ~2% of US households were paying for some AI service”, that “the number is higher now, and growing”, and that “it’s still so early when it comes to mature AI adoption and utilization”. Both readings can be true at once: the market is early, and the early numbers are small.
The Netflix Test for Consumer AI
To show how far consumer AI is from break-even, TechCrunch borrows a benchmark from streaming. Netflix, with about 325 million subscribers, is the standard for an online service that has reached saturation.
A calculation that does not add up
The column says that “$34 per customer only gets you to $11 billion in annual revenue, less than a third of OpenAI’s operating costs”. The multiplication works: 325 million times $34 is about $11 billion. But that only holds if $34 is an annual figure. The same column reports average spending of $31 a month, which is $372 a year. At that rate, 325 million payers would generate about $121 billion a year.
Why the error does not rescue the business
Correcting the arithmetic makes the revenue line look far better, but it does not change the conclusion as much as it seems. Reaching Netflix scale would mean persuading far more people to pay than any payment study shows today. And unlike Netflix, whose cost of streaming one more film is small, an AI service pays for computing power every time someone uses it, so revenue and cost rise together.
| Scenario at 325 million payers | Price assumed | Annual revenue |
|---|---|---|
| TechCrunch’s figure | $34 per customer, read as per year | About $11 billion |
| PNC average spend, per month | $31 a month = $372 a year | About $121 billion |
| Menlo typical payer, low end | $20 a month = $240 a year | About $78 billion |
Bank of America’s ceiling
For context, BofA Global Research expects the US consumer AI market could scale to $75 billion a year, helped by higher-priced plans and people paying for convenience. Even that optimistic figure sits well below the cost base of the largest labs, which spend tens of billions a year on computing power and staff.
Why Costs Are the Real Problem in Consumer AI
TechCrunch’s central claim is that “the problem with the consumer approach has less to do with revenue than with cost”. This is the part of the argument that explains why the major labs have turned to businesses.
Every answer costs money
Earlier consumer internet products were cheap to run. Once Facebook or Gmail was built, serving one more user cost very little. A language model runs on expensive chips for every request, and an agent that browses, plans and calls services may make dozens of model calls to finish one task. The heavier the use, the higher the bill.
Computing power is not getting cheaper fast
The State of Markets report adds a detail that cuts against the hope of falling costs. It finds that rental prices for AI chips have been climbing for the latest models and holding up even for older ones, with Nvidia’s A100 “pricing at-or-above what it was at the beginning of the year”. Demand for computing is still outpacing supply.
Agents multiply the bill
The products driving the current excitement are also the most expensive to run. Menlo finds that 41% of AI users have tried an AI agent, 24% use one regularly and 32% have let AI act on their behalf without a final approval step. Among agent users, 92% pay for AI. That is good news for revenue and bad news for cost, because an agent that books a restaurant may search, read menus, compare times and confirm, each step a separate model call.
Features that add computer vision, such as reading a photo of a receipt or a shop shelf, cost more again. The consumer AI products people love most are the ones with the thinnest margins.
Power users cost the most
The concentration of spending has a cost side too. The users who pay $100 or more a month are also the ones running the longest, most demanding tasks. Pricing has adapted: Menlo finds 15% of payers combine a subscription with usage charges. Our analysis of ChatGPT’s new $500 tier showed OpenAI charging roughly the same $20 for each unit of usage at every level from Plus upwards, which is how a business protects margins against heavy users.
Why scale does not guarantee profit
In software, more customers usually mean higher margins. In consumer AI, TechCrunch warns, “even hundreds of millions of paying customers doesn’t guarantee you’ll break even”. That single line explains much of the industry’s recent strategy.
Three Ways Companies Try to Make Consumer AI Pay
The companies behind last week’s launches are each taking a different route around the problem.
| Company | Consumer product | How it plans to make money | Main risk |
|---|---|---|---|
| Meta | Muse and Jolly | Personalised advertising; an enterprise platform | Users may resist ads inside an assistant |
| OpenAI | ChatGPT and Dots | Subscriptions plus selling the same tools to businesses | Consumer use cross-subsidised by enterprise |
| Instinct | Errand-running agent | A cut of purchases made through the agent | Thin margins on each transaction |
| Anthropic | Claude apps | Enterprise contracts, sector by sector | Smaller consumer brand |
OpenAI’s enterprise pivot
TechCrunch says OpenAI “seems to have adapted well”, with enterprise bookings reportedly doubling since July. Even the Dots launch had an enterprise angle, with demonstrations aimed at software engineers and agency creatives. The column calls this an old playbook: build a popular, cheap consumer service, then sell it to businesses at a markup.
Meta’s advertising engine
Muse has “the juggernaut of Meta’s personalized ad targeting behind it”, which buys time before monetisation becomes urgent. Meta is also exploring business sales; it recently launched an enterprise AI platform and hired MongoDB’s chief executive to run it.
The subscription stack
Most consumer AI revenue still comes from subscriptions. Menlo counts 50% of payers on flat monthly plans and 26% on annual plans. The risk for providers is subscription fatigue: a household that already pays for streaming, music, cloud storage and a phone plan has a limited budget for another $20 a month, let alone several. That pushes consumer AI towards bundles, such as AI features folded into existing phone or software subscriptions, where the AI revenue is harder to see and harder to grow on its own.
Instinct’s commission model
Instinct plans to take a cut of purchases made through its agent, which ties revenue to activity rather than to a fixed subscription. It also avoids the cost of training its own frontier model. Still, the column argues that consumer AI economics put “a hard cap on how large the company can plausibly grow without tapping into enterprise revenue”.
What the Economics of Consumer AI Mean for Users
If the economics are as tight as the data suggests, users should expect the products to change in predictable ways.
Free tiers will tighten
Menlo’s advice to founders is to keep free tiers, because that is where people discover value before they pay. But free use is pure cost. Expect tighter limits, slower models for free users and more prompts to upgrade.
More tiers, more metering
The move to usage-based pricing will continue. Menlo already counts 17% of payers on usage-based plans and 15% on a mix of subscription and usage charges. Heavy users of consumer AI will increasingly pay by the task.
Advertising and commerce
Where subscriptions fall short, advertising and commissions fill the gap. That raises questions about whether an assistant recommends what is best for the user or what pays the platform, a concern that regulators in the UK and EU are already watching in search and online shopping.
Trust becomes part of the price
Menlo’s survey suggests that what users value is shifting. Accuracy (45%), trustworthiness (40%) and security and privacy (36%) now rank ahead of ease of use, which fell from 38% to 32%. Among people who do not use AI at all, 70% distrust the information it provides, up from 58%, 76% cite privacy and 83% would rather deal with a person. For providers, that means the cheapest route to more payers may be proving reliability rather than adding features, which is a slower and costlier kind of product work.
For UK businesses
UK firms building consumer AI products should model their cost per active user before they model revenue, and plan for a business-customer route from the start. Firms buying AI tools should expect vendors to push enterprise plans, and should negotiate on usage limits as much as on price.
Consumer AI FAQ
How many people pay for consumer AI?
Payment data from PNC and Bank of America suggests only 2% to 3% of US consumers or households pay. A Menlo Ventures survey suggests a much higher share, about 55% of AI users. The gap reflects different methods.
How much do people spend on consumer AI?
PNC’s data shows an average of $31 a month among payers. Menlo finds most payers spend $20 to $49 a month, while the 14% who spend $100 or more account for 60% of spending.
Why is consumer AI hard to make profitable?
Because each request costs real computing power. Unlike social networks, an AI service’s costs rise with use, so more customers do not guarantee a profit.
Why are AI labs focusing on enterprise customers?
Businesses pay more per user and sign longer contracts. TechCrunch describes an industry-wide shift toward enterprise deals and sector-by-sector expansion.
Will consumer AI products get more expensive?
Expect more pricing tiers, more usage-based charges and tighter free plans, alongside advertising and commission models.
References and Further Reading
The ugly economics of consumer AI (TechCrunch)
State of Markets II (Andreessen Horowitz)
2026: The State of Consumer AI (Menlo Ventures)
Menlo Ventures report: consumer AI spend tripled to $40B this year (GlobeNewswire)
Not quite mainstream: a consumer AI profile (Bank of America Institute)
Not quite mAInstream: full analysis (Bank of America Institute, PDF)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
