Consumer AI: Why Now, and What It Takes

Consumer AI is artificial intelligence built for individual people rather than for companies. Same models as the enterprise tools, entirely different job: nobody signs a contract, nobody is told by a manager to log in, and anyone who gets bored on a Thursday just stops opening the app.

Y Combinator recently put out a request for startups arguing that this is the category to build in right now. The pitch is short and it is good. It is also the easy half of the problem. The head of ChatGPT, the CEO of Palo Alto Networks, and the creator of OpenClaw have each said something on the record that the pitch leaves out. Here it is with that attached.

What consumer AI means, and how consumer AI apps differ

The clean line is who pays and why. Enterprise AI is bought by an organization for its staff, through a procurement process, against a business case. Consumer AI is chosen by one person, with their own money or their own attention, for reasons they mostly do not articulate.

That distinction is not academic, because it changes which numbers matter:

• Retention is the honest scoreboard. There is no contract to hide behind and no admin rolling out a mandate. A consumer product that people do not come back to is not a product with a churn problem, it is not a product.
• Distribution is not a go-to-market function, it is the constraint. An enterprise startup can win on a slow, expensive, high-touch motion. A consumer startup that cannot get noticed cheaply does not get a second act.
• The buyer and the user are the same person. That removes an entire class of enterprise pain (the champion who loves it, the CFO who does not) and replaces it with a harder one: the product has to be good enough to survive a bad mood.

YC's case: the window is opening, and you can compute when

The argument starts with pattern matching, and the pattern is real. "Every platform shift mints consumer giants," the video says. "The web gave us Google and Airbnb." Mobile produced Instagram and DoorDash. Then the observation that makes the whole thing sting: "And 3 years in, the only new icon on your home screen is ChatGPT."

That is an unusually specific claim about the state of a boom. Enormous capital and attention have gone into this platform shift, and the consumer layer has produced roughly one durable new habit. Either consumer AI is structurally hard, or it is early. YC argues it is early, and names two conditions that just changed.

The first is capability. "You can treat an agent like a person," which is a low bar to state and a high bar to clear: it means the thing holds context, takes an ambiguous instruction, and comes back with something you did not have to correct.

The second is price, and this is the part worth writing on a whiteboard. "Today, the magic can run a thousand dollars a month in tokens for each user. But that is falling 10x a year." The conclusion drawn from it is the useful bit: "If you follow the curve, you can predict the consumer moment."

That is a planning instrument, not a slogan. Build the thing you actually want to ship with no cost discipline at all, measure what it costs per active user per month, and divide by ten for each year out. If your dream product costs $400 a month per user today and you can charge $20, you are a little over a year from viable at that rate, which is roughly the time it takes to build it properly anyway. That is a real answer to why now, and you can compute it this afternoon.

Near the end, the video lists the categories it expects to reopen: "How we get things done, get around, learn, stay healthy, manage our money, play, connect with friends."

Now for the parts a 57-second pitch does not have room for.

The price you pay is not the same as the cost curve

The 10x-a-year assumption is doing a lot of work in that plan, so it deserves a skeptic. Nikesh Arora, chairman and CEO of Palo Alto Networks and previously president and COO of SoftBank after a decade at Google, gave one on 20VC.

Asked why frontier model prices have not fallen the way everyone expected, his answer was not about silicon. It was about who is paying for whom. Free consumer usage, he argues, is "sucking away half the compute which is making no return", which pushes cost recovery onto the half that does pay, meaning enterprise and coding. "So that's forcing token prices to go up."

Behind that sit the labs' own balance sheets: "your frontier model companies are value maxing, not token maxing." They are raising at enormous valuations, they need the next hundred billion of compute, and at some point the financial markets stop funding that without a credible path to gross margin. The only lever they have is to charge more for the fastest growing thing they sell. "That's where you get the price of tokens from."

He is not arguing the curve is fake. He thinks long-term token pricing should land at about a tenth of today's, and he is explicit about the pace: "I think in the next 3 to 5 years we will see reduction in token pricing." He also notes he does not need the newest frontier model to do 90 percent of what people actually use AI for, which is where a lot of the practical saving comes from.

Now put the two forecasts side by side, because a consumer AI plan lives or dies in the gap between them. YC says roughly tenfold a year, which compounds to something like a thousandfold over three years. Arora says roughly tenfold in total over three to five years. Same direction, completely different arrival date for your product. If your unit economics only work at the first rate, you are betting on the more aggressive of two serious forecasts, and you should at least know that is the bet. That arrival date is not evenly distributed either: for AI startups in India building for the next billion users, the price point their customers can reach is the product spec, not a line in the model.

There is one more thing in his answer that a consumer founder should not skip, because it points straight at this category. He expects that "the consumer use of AI will get constrained by these frontier AI companies", on the grounds that they now have more post-training data than they need and that "each user is inherently unprofitable in their activities they do on Frontier AI models". Read plainly: a lot of what makes consumer AI feel possible today is running on somebody else's subsidy, and the people paying for it have started saying out loud when they expect to stop.

For a founder, that turns one plan into two:

1. Plan A, the curve holds. Your unit economics fix themselves on schedule. Great, but do not spend against it before it arrives.
2. Plan B, prices stay sticky for eighteen months longer than you assumed. What do you cut? Which requests can go to a small model, what can you cache, which single moment in the product genuinely earns the frontier model?

The teams that survive a sticky year are the ones that built Plan B while the deck said Plan A. Our post on AI pricing works from the assumption that the cost floor keeps falling, which is the right long-run planning assumption. Arora's point is about the short run: set your price against the margin you can defend on today's invoice, not the one the curve has promised you.

Model quality is about a third of the story

The most useful correction to the idea that the models finally got good enough comes from the person with the most to gain from that framing being true.

Nick Turley is VP and head of ChatGPT at OpenAI, and before that a product lead at Instacart and Dropbox. Asked on the BG2 podcast what actually drove ChatGPT's growth from zero to the scale it is now, he broke it into three roughly equal parts.

The first was "classic friction removal type of work". His example: one of the single biggest impact moments was "removing the authentication wall", the decision to stop making people log in before they could try it. He notes that this had been Sam Altman's feedback from day one. This is not AI work. This is the same conversion work every consumer product has done for twenty years.

The second was core product investment, built jointly by research and product rather than by either alone: search, personalization, the writing surfaces that render when you are drafting with the model. His framing is that the wins came specifically from the moments where research and product built the thing together, which is a staffing observation as much as a product one.

Only the third was model improvements, and even there he splits it: the step changes everyone remembers, plus a lot of unglamorous iteration that never earned a named release and shows up in retention anyway.

Sit with the ratio. At the company with the best models in the world, roughly two thirds of the growth of the flagship consumer product came from work that has nothing to do with model quality. That is either bad news or the best news in this post, depending on what you were planning to differentiate on. If your consumer AI thesis is that the model is finally good enough, you have described the market's entry ticket, not your advantage.

Turley is also blunt about how much room is left, which is the actual answer to the billion-user question: "We've got about 10% of the world coming to us now. 90% left to go". OpenAI reported 900 million weekly active users in February 2026, per TechCrunch. The incumbent's own read is that nine tenths of the world's population has not shown up yet.

One more thing from him worth keeping, because it is the honest version of shipping too early. On an earlier attempt at agentic capability: "The models weren't quite good enough to hit real escape velocity." And the consequence, which is the part founders underrate, is that "users don't learn to trust it. They don't even try." Shipping a consumer AI feature that mostly works but cannot be relied on does not buy you a head start. It teaches people the thing does not work, and that lesson is expensive to unteach.

The magic is how it feels, not what it can do

Peter Steinberger spent more than a decade building PSPDFKit, the PDF framework used inside a very large number of shipped mobile and web apps, and then created OpenClaw, which became one of the most talked about open-source projects of the year. In a talk at Y Combinator he described the moment the thing clicked, and it is the clearest description of consumer AI product design I have found.

He was annoyed that he could not check on his coding agents from his phone, so he had a model build him a WhatsApp relay in about an hour. Then this: "But the magic is how it felt. It wasn't a terminal. It wrote concise answers." It was proactive enough to check in on him during the day. And critically: "You don't have to think about which model, which context size or when you start a new session."

Nothing in that list is a capability. Every one of them is a decision the product made so the user would not have to. The underlying thing (send text to a model, get text back) is what everyone already had, in a terminal, for months.

That is the whole game in consumer AI. The capability is on tap for you and for every competitor at the same price. What is not on tap is the judgment about what to hide, what to decide on the user's behalf, and when to speak first. Deciding when a product should speak first is its own discipline, and we cover it in proactive AI. The order matters: decide what it should feel like, then pick the architecture that allows it.

Your demand test is an emotional reaction, not a survey

The second half of Steinberger's story is the cheapest consumer research method there is, and it is the one founders skip while they are still trying to explain the product.

He could not get people to understand why this mattered by describing it. "Week after week I failed." So he stopped describing it. He put it in group chats with friends and let them talk to it directly. (The same failure to explain shows up as a distribution cost in our post on running an agent on your own machine; here it is the test itself.) "Every time I got a strong emotional reaction." Some were amazed, some were scared or freaked out. Every time, the emotion was strong. And the signal that mattered most: he told his non-technical friends this was not for them yet, "and they got mad". His conclusion: "So if that's not an indicator to have product market fit I don't know what is."

Note what he is measuring. Not a rating, not an intent-to-purchase score, not whether people said they liked it. Intensity. In consumer, the failure mode is not people hating your product, it is people feeling nothing and being polite about it. A five-out-of-seven satisfaction score from someone who will never open it again looks identical to real demand on a dashboard, and nothing like it in a group chat.

Two practical rules fall out of this:

• Put the thing in front of people, do not explain it. If it needs the explanation, the explanation is the product and you have not built it yet.
• Score the reaction, not the answer. Did anyone get annoyed that they could not have it? Did anyone show it to someone else without being asked? Before you have users, those two beat any number you can put on a dashboard. If nobody reacts, you may be looking at a real need with no felt urgency, which is a different problem worth reading up on in our post on latent demand.

The hardest problem is not the technology

Asked how he would approach his next startup, Steinberger's answer was almost entirely about attention: "your hardest problem is not the tech, it's not the software, not even the people. The hardest problem now is like getting eyeballs."

This is the structural reason consumer AI is harder than the YC video makes it sound, and it gets worse as the two conditions in that video get better. Cheap intelligence means your competitors also ship fast. When anyone can prompt a working version of your idea into existence, the scarce inputs are the ones that were always scarce in consumer: taste, trust, and reach.

His other suggestion is worth repeating because it cuts against the instinct of most AI founders: he would pick something in the category of hard and boring, because it is easier to find people who genuinely appreciate a solved problem there. Something fun, he argues, gives you a very tough time even when it is technically hard, in a period when people can prompt things into existence.

Practically, that means treating distribution as a product surface with an owner and a budget from the first week, not a phase you enter after launch. It also means being honest about which one you are actually good at. Consumer AI rewards founders who can build and can get noticed. If the second is not in the founding team, it is the first hire, not a later problem. The engineering and go-to-market sections of our AI pillar go deeper on how those two halves fit together.

What a billion people are actually handing over

The last piece is the one my security background will not let me skip, and it is not a compliance footnote. At consumer scale it is a growth constraint.

Mike Mignano, now a general partner at Union Square Ventures and previously co-founder of Anchor, the podcasting company Spotify acquired in 2019, framed it on 20VC about as starkly as it can be framed: "We have never handed over so much of ourselves to a technology before than we're about to do with agents."

His concern is incentive alignment, and it is a product question rather than a legal one. If a user hands an agent their personal information, their goals, and their payment methods, someone is going to ask who that agent is working for. His view is that products, agents, and harnesses that are, in his words, "human aligned and aligned with your own goals and incentives" will be important. He is careful about how far that goes: he expects most people to get comfortable handing over the keys within about five years, the way they got comfortable putting card numbers online, and he does not think every user needs to care. His bet is narrower, that enough of them care to keep a few honest players in the market holding the rest in check.

Pair that with Turley's point about escape velocity and you get the real trust dynamic in consumer AI: users are not evaluating your privacy policy, they are deciding, fast, whether this thing is on their side. Every unexplained action, every piece of data that turns up somewhere they did not expect, every moment the product acts on an incentive that is obviously yours and not theirs, spends that down.

So the security work in a consumer AI product is mostly product work:

1. Say what leaves the device, in the product, in plain language. Not in a settings page nobody opens. At the moment it happens.
2. Put a hard ceiling on unsupervised action, especially anything involving money or messages sent as the user. The blast radius of a confused agent is the whole relationship.
3. Build the undo before you build the automation. A visible, one-tap reversal of anything the agent did on its own is worth more trust than any amount of accuracy marketing.
4. Assume the incident becomes a headline. Steinberger watched the press report that 20 percent of OpenClaw's skills were malicious; he says they scanned all 67,000 and published a figure closer to 0.3 percent, and as he put it, "a correction never travels as far as a scare." At consumer scale you are not managing a bug, you are managing a story about your product that you do not control.

None of this costs you speed. All of it is far cheaper before launch than after.

What to do this week

1. Compute your consumer moment twice. Build the ungated version of your product and measure real cost per active user per month. Then run the number down two curves: tenfold a year, which is YC's, and tenfold over three to five years, which is Arora's. The gap between those two dates is your actual risk, and it is worth one page of writing down.
2. Write Plan B for a sticky price year. Which requests move to a smaller model, what gets cached, which single moment keeps the frontier model. Do not ship it yet, just know it.
3. Audit your differentiation against the one-third rule. List what makes your product better. Cross out everything that is really just a good model underneath. If the list is now empty, your roadmap for the next quarter is friction removal and product craft.
4. Run the group-chat test on ten people, five of them non-technical. Show, do not explain. Record who reacts strongly and who reacts politely. Ship nothing else until at least three people get annoyed that they cannot keep using it.
5. Name a distribution owner. One founder, this week, with a budget and a weekly number. If nobody owns getting noticed, nobody is doing it.
6. Ship the undo before the automation. Take the one action your product already performs on its own and put a visible, one-tap reversal on it this week. If it cannot be reversed, it should not be unsupervised.

If you want the full operating system for running this, from the model choices to the security posture to the go-to-market, that is what we teach in AI Operating System for Startups.

Sources

• AI-Powered Consumer Products for 1 Billion People, Y Combinator's request for startups, the source this article distills.
• Nick Turley on the BG2 podcast for the one-third breakdown of ChatGPT's growth, the escape-velocity and trust point, and the 10 percent of the world figure. Profile: Nick Turley.
• Peter Steinberger's talk at Y Combinator for the WhatsApp relay story, the group-chat demand test, the getting-eyeballs point, and the malicious-skills correction. Background: Wikipedia and Fortune.
• Nikesh Arora on 20VC for the argument that token prices reflect the labs' need for gross margin rather than the cost of serving. Background: Wikipedia and Fortune.
• Mike Mignano on 20VC for the incentive-alignment argument about agents. Profile: Union Square Ventures.
• ChatGPT reaches 900M weekly active users, TechCrunch, for the reported scale figure.