a16z's Olivia Moore: Ad-Supported AI Could Generate $152B/Year — Here's the Math
Olivia Moore at a16z calculated that ad-based AI ARPU matching Google's $460/user/year would dwarf subscription revenue. Here's the full model.
The $152 Billion Argument for Ad-Supported AI
Olivia Moore at a16z ran the numbers, and they’re hard to argue with: if ChatGPT’s ad-based ARPU matched Google’s $460 per user per year in the US, that’s $152 billion in annual revenue. By contrast, converting 5% of the US population to a $200/month subscription gets you $40 billion. The subscription model isn’t just smaller — it’s less than a third the size.
That gap is the entire argument for why consumer AI, despite being largely abandoned by the enterprise-focused labs right now, is probably not dead. It just needs a different revenue model.
You should care about this math if you’re building anything in the consumer AI space, or thinking about where the next wave of AI product opportunity actually lives.
Why the Subscription Model Has a Structural Ceiling
Start with the Bank of America data point: only 3% of their customers pay for AI. That’s not a temporary adoption lag — that’s a signal about the fundamental shape of consumer willingness to pay for a general-purpose tool.
Compare it to how people pay for other things. People pay for Netflix because it has specific content they want. They pay for Spotify because the music library is the product. They pay for ChatGPT because… they use it a lot and feel vaguely guilty about it? The value proposition for a general-purpose AI subscription is genuinely harder to articulate to a median consumer than “this is where the shows are.”
Jamie Dimon put it plainly: “It’s not clear to me how consumer is going to play out.” His framing was that enterprise use cases have found their niche — you make an investment, you measure the return, you keep paying. Consumer is murkier. A lot of people use Gemini for free and that’s sufficient for their requirements.
The subscription model also has a compounding problem: the people most willing to pay are power users, and power users are increasingly the ones whose usage patterns look more like enterprise than consumer. They’re running agents, burning tokens, building workflows. The casual user — the one who uses ChatGPT to draft an email once a week — is not going to pay $20/month for that.
So you end up with a barbell: heavy users who arguably should be on consumption-based pricing, and light users who won’t pay at all. The middle — the reliable $20/month subscriber who uses the product consistently but not intensely — is a smaller population than the subscription model needs.
The Google Comparison Is More Interesting Than It Looks
The $460/user/year figure for Google is worth unpacking. That’s not what Google charges users — users pay nothing. That’s what Google extracts from advertisers in exchange for user attention and intent signals.
The reason this number is so high is that Google sits at the moment of decision. When someone searches for “best running shoes,” they’re about to buy something. That intent signal is enormously valuable to advertisers. Google’s entire business is monetizing the gap between “I want something” and “I bought it.”
ChatGPT is increasingly sitting in the same position, and arguably a more powerful one. When someone asks ChatGPT “what’s the best running shoe for someone with flat feet who runs 30 miles a week,” that’s a richer intent signal than a keyword search. It’s a conversation. The model knows context. It could, in principle, make a recommendation that’s both genuinely useful and commercially relevant.
Moore’s argument is that ChatGPT’s ad-based ARPU could actually exceed Google’s $460 because of deeper and more frequent user engagement. That’s a reasonable hypothesis. ChatGPT’s engagement ratio — weekly to monthly active users — is already ahead of X, Spotify, and TikTok. Users are coming back habitually, not occasionally. Time per user has roughly tripled since early 2023. These are the metrics advertisers care about.
Meta makes around $250 per user per year on ads, for comparison. They have 3 billion daily active users and a highly tuned ad machine. ChatGPT at 900 million weekly active users is approaching comparable scale, without any ad infrastructure yet. The Anthropic compute shortage situation is a useful contrast here — when supply is constrained, you lose users and the audience thins. Meta’s advantage is that it built its ad machine on top of an audience it had already locked in.
What the Current Signals Actually Tell You
The industry’s current posture is instructive precisely because it’s so extreme.
How Remy works. You talk. Remy ships.
OpenAI canceled a billion-dollar Disney deal and shut down the Sora app to free up compute for enterprise and coding use cases. That’s not a subtle signal. That’s a company making an explicit choice about where its tokens go when supply is constrained. CEO of Applications Fiji Simo has been pushing hard to cut “side quests” and focus on the core coding and enterprise business. Consumer video generation was a side quest.
Meanwhile, 159 of 175 companies in the latest Y Combinator batch were focused on enterprise. Brian Chesky’s read on this is interesting: he thinks part of the reason is that consumer is just harder. You need design, marketing, culture, press — not just technology and sales. But his prediction is that we’re in the age of enterprise AI now, and in 12-24 months you’ll see the beginning of a consumer AI renaissance.
The contrarian bet, if you’re inclined to make it, is that the current enterprise focus is creating a gap. Consumer AI infrastructure is being underinvested relative to its eventual monetization potential. The companies that figure out the ad model — or some other non-subscription revenue model — before the renaissance arrives will have a significant head start.
Meta is the most obvious company making this bet explicitly. They’re forecasting $125-145 billion in infrastructure spend in 2026, and Zuckerberg has been clear that consumer AI is their primary focus, not coding agents or enterprise APIs. “I’m not against having an API or coding tools, but it’s not our primary focus,” he said on the earnings call. For a company that makes $250/user/year on ads, the math of consumer AI is obvious.
The Agentic Commerce Problem
There’s a third revenue model that gets discussed alongside ads: agentic commerce. If AI agents are doing your shopping, they could take a cut of transactions, or at least be paid for referrals.
The skepticism here is warranted. Andy Jassy’s framing is useful: agentic commerce is “a small fraction of search engine referrals,” and the experience hasn’t gotten great yet. Third-party agents don’t have personalization data or shopping history. They can’t always get pricing right. They lack the context that makes a recommendation actually trustworthy.
There’s also a behavioral question. Shopping is two different things. Sometimes you want the thing and you don’t want to think about it — you want the agent to just handle it. But a lot of shopping involves browsing as the experience. People like the quest. Discovering options is part of the value. It’s not obvious that agents fit that mode at all.
And even in the “just get me the thing” scenario, there’s a cognitive cost to offloading that decision. You have to tell the agent your criteria, your constraints, your preferences. If you leave something out — which you will, because you don’t know what you don’t know until you see the options — you get results that don’t quite fit. At that point, you might as well have done it yourself.
Ads have none of these problems. They work with existing behavior rather than trying to change it.
Building the Infrastructure for Ad-Supported Consumer AI
If you accept the thesis that ad-supported AI is the likely end state for consumer, the question becomes: what does that infrastructure look like, and who’s building it?
Not a coding agent. A product manager.
Remy doesn't type the next file. Remy runs the project — manages the agents, coordinates the layers, ships the app.
The technical requirements are different from subscription AI. You need user modeling at scale — understanding preferences, history, context — to make ad targeting valuable. You need measurement infrastructure so advertisers can verify that their spend is working. You need a policy layer that prevents the model from making recommendations that are purely commercially motivated at the expense of user trust.
That last one is the hard part. The reason Google’s ad model works is that users trust the search results enough to keep using Google. The moment users feel like the AI is steering them toward paid results rather than good results, the engagement metrics collapse and the ad revenue follows. The alignment problem for ad-supported AI is not just technical — it’s economic.
OpenAI has been quietly building ad platform capabilities, even if those announcements get buried relative to model releases and enterprise deals. The foundation is being laid.
For builders thinking about this space, the interesting question is what you can build on top of this infrastructure before it exists at scale. The consumer AI renaissance Chesky is predicting will need applications — things that live on people’s home screens and change how they interact with the world. Almost every app on his home screen, he noted, hasn’t changed since AI arrived. That’s a lot of surface area.
If you’re prototyping consumer AI applications and need to connect multiple models and data sources, MindStudio handles the orchestration layer — 200+ models, 1,000+ integrations, and a visual builder for chaining agents and workflows — which lets you test product hypotheses without rebuilding infrastructure for each experiment. That kind of flexibility matters when you’re iterating fast on consumer product ideas before the ad model infrastructure fully matures.
The Free Tier as Distribution
One underappreciated angle in the ad model argument is what the free tier is actually doing right now.
GPT 5.5 Instant just replaced GPT 5.3 Instant as the default model for free and $8 Go plan users. The benchmark jump is significant: 81.2 on the AIM 2025 math test versus 65.4 for its predecessor. MMLU Pro went from 69.2 to 76. The model now has memory access, a Gmail connector, and better context management. Ethan Mollick’s read is that the free model is now at a similar level to frontier models from late 2025.
This matters for the ad model because the free tier is the audience. OpenAI removed the model selector for free and Go users when GPT 5.3 Instant launched in March — they’re not trying to upsell you to a better model, they’re trying to give you a good enough model that you keep coming back. At 900 million weekly active users, they have an audience that rivals TikTok and is approaching WhatsApp.
That audience is the asset. The question is how to monetize it without destroying the engagement that makes it valuable. When you can’t serve demand, you lose users — the free tier has to be good enough to retain the audience that the ad model eventually monetizes. The model choices available to free users also matter competitively; a comparison of GPT-5.4 vs Claude Opus 4.6 on specific workflows illustrates how much the quality gap between tiers affects retention — if the free model is too far behind the paid one, users churn rather than upgrade.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
The GPT Images example is instructive here too. GPT Images 1 drove 12 million incremental downloads in 2025 — a genuine viral moment. GPT Images 2 this year generated far less consumer hype. The only meme that landed was replicating a 5-year-old’s MS Paint drawing. The viral moment was already faded within a week. Consumer attention is hard to capture and easy to lose.
It’s also worth watching what open-weight models are doing to this dynamic. Qwen 3.5 and similar releases are compressing the cost of serving capable models, which changes the economics of the free tier — if inference gets cheap enough, the gap between free and paid narrows, and the ad model becomes the only lever left for meaningful revenue differentiation.
The Spec for a Consumer AI Business
If you were writing the spec for a consumer AI business that works in an ad-supported world, what would it say?
Deep engagement over broad reach. The ad model rewards habitual users, not occasional ones. ChatGPT’s engagement ratio exceeding TikTok is the right direction. You want users who come back daily, not weekly.
Intent-rich interactions. The more the AI understands what users actually want — not just what they typed — the more valuable the attention signal is to advertisers. Conversational AI has a structural advantage over keyword search here.
Trust preservation. The moment users feel manipulated, engagement collapses. The ad model only works if users trust the recommendations enough to act on them. This is a harder constraint than it sounds.
Personalization infrastructure. Google’s $460/user/year is built on two decades of user data. A new entrant starts from zero. The companies that build personalization infrastructure now — even before the ad model is live — will have a significant advantage when it is.
Tools like Remy take a related approach to the infrastructure question: you write a spec — annotated markdown — and the full-stack app gets compiled from it, including backend, database, auth, and deployment. The spec is the source of truth; the code is derived output. For consumer AI builders who need to iterate fast on product hypotheses, that kind of abstraction matters — you can validate whether a product concept has legs before committing to a full engineering build.
Where This Leaves the Math
Moore’s $152 billion figure is not a prediction — it’s a ceiling calculation. It says: if you match Google’s ad efficiency at ChatGPT’s current scale, this is what the revenue looks like. The actual number will be lower, at least initially, because ad infrastructure takes time to build and the targeting capabilities aren’t there yet.
But the comparison to the subscription model is the point. $40 billion from 5% subscription conversion is not a bad business. $152 billion from ads is a different category of business entirely. And the subscription model has a ceiling that the ad model doesn’t — you can only convert so many people to paid tiers before you’ve exhausted the willing population.
One coffee. One working app.
You bring the idea. Remy manages the project.
The current moment — where enterprise AI is getting all the attention and consumer AI is being treated as a side quest — is probably not the permanent equilibrium. The economics of consumer AI at scale are too good to ignore indefinitely. The labs that are currently deprioritizing consumer are doing so because compute is constrained and enterprise users consume more tokens per dollar of revenue. When compute supply catches up with demand, that calculus changes.
The question for builders is whether to wait for that shift or to position ahead of it. The infrastructure for ad-supported consumer AI — the user modeling, the measurement, the trust layer — takes time to build. The companies that start now will be ready when the renaissance Chesky is predicting actually arrives.
The math says it’s coming. The only question is when, and who’s ready for it.
