Context
a16z publishes a Top 100 Consumer AI Apps list, and this is its seventh edition. Olivia Moore, who writes the report, and Josh Elman, an investor on a16z's consumer team, walk host Elena Burger through it. For the first time the list ranks products by what consumers pay, not just by web traffic. That matters to PMs because it shows who actually pays for AI today, which business models are working, and where consumer products are still missing.
The Big Idea
Consumer AI is a power-user game today: about half of Americans use AI, but only around 4.5% pay, and a small slice of them drives most of the revenue. The next big opening is in products that make people want to spend time, funded by ads and transactions instead of subscriptions.
The spend data shows a very different market from the traffic data. Most people are still using AI as a faster search box, so the biggest white space is in categories where AI is not yet used at all.
Key Insights
Spend data changed the rankings
- Earlier editions ranked only by web traffic, which worked when most AI products were a prompt box in a browser.
- This edition adds consumer card-spend data. It excludes enterprise and small business spend, and ranks products by what individuals pay.
- On traffic, the market looks settled: only 11 new products across the web and mobile lists, the fewest of any edition.
- Spend told a different story: of the 50 products ranked for spend, 29 were on neither traffic list.
Few people actually pay
- About half of Americans report using AI, and roughly 25% think they interact with it close to daily.
- Only around 4.5% of US consumers pay a subscription. Other sources say 2 to 2.5%. The figure has about doubled in a year.
- Inside that group, the top 10% of spenders generate more than half of revenue. The top 1% generate about 20%. The bottom 50% generate about 16%.
- The top 1% spend $903 per month on personal cards. The median payer spends $25.
Power users buy tools for making
Olivia said spend leans toward developer, productivity and creative tools, the things people use to "build, make, sell". She named n8n, Higgsfield and Manus as over-represented among top spenders. Josh read this as a good sign: people are using their own credit cards to code, make art and make videos they never could before. The reason is partly mindset. Developers instinctively ask how to get leverage over their time, and that worldview is still held by a small group.
Subscriptions are an unnatural fit
Josh argued he does not want the 4.5% who pay to simply grow, which he called maybe controversial for a consumer AI maximalist. His reasoning:
- About 85% of the web list monetizes by subscription, and another 62% by credits or extra-usage payments. Only about 13% use ads or other models.
- The biggest consumer internet companies earn most of their revenue from ads or transaction fees, not subscriptions.
- Most people don't want to pay for software. The top 20 consumer subscriptions globally are mostly media or big platforms with frequent purchase behavior.
- Subscriptions became the default because serving each user costs real money. In the past, companies built user density first, since serving cost was near zero.
Fast growth can hurt AI startups
Josh noted that high serving costs make companies afraid of growing too fast. Without meters on usage, spend can spiral out of control. This shapes the whole top 100: some products deliberately limit growth. He expects ads and transaction fees to start working as inference costs come down.
Ads are already working at OpenAI
- OpenAI's ads reached about $1 billion in annualized run rate (the number was from August). Olivia said this would normally take a company years, even after launching an ads product.
- Two reasons: ChatGPT has real density (the latest reported figure was 1.2 billion weekly active users), and it knows far more about each user than most ad platforms.
- Josh said the rollout has been careful. The ads he saw were clearly labeled and felt natural to the conversation.
- People now bring commercial questions to chat, such as shopping and trip planning. Those are discovery moments, where an ad can be a useful answer.
- Josh's caveat: the ad must not feel like an interruption inside a very personal conversation.
Vertical ads can work too
Olivia pointed to Open Evidence, an AI product for doctors. She said recent numbers showed 50 to 60% of US physicians using it. A dense, known audience of that kind lets advertisers target very well. She expects more focused products like this.
ChatGPT, Claude and Gemini are diverging
- ChatGPT is still the dominant consumer product by usage and revenue. It is about 6x ahead of Claude and about 2x ahead of Gemini on the web.
- The surprise: on a US panel of spenders, Claude has passed Gemini in paid subscribers, even though Gemini has a larger user base and Google's distribution.
- ChatGPT has about 3x more US paid consumer subscribers than either Gemini or Claude.
- Anthropic has said no to ads, so it leans harder on subscriptions. About 7.5% of its subscribers are on the $100-plus per month plan, versus about 1% for ChatGPT and Gemini.
- Olivia credited Claude's rise to a run of successful launches (she named Claude Design) and a lot of press.
Personal agents are the new trend
- Olivia said OpenClaw, which topped a retro ranking in the last edition, has dropped out of this one as its traffic fell. She noted the team was acquired by OpenAI.
- Josh said agents turned AI from a tool for productivity or search into something that gets things done. OpenClaw sparked that shift.
- Products now feel safe enough to hand to a family member. Olivia named Muse and Instinct as the biggest developments in pure consumer assistants.
- Instinct had announced about 100,000 users growing 10% day over day, with 40% connecting a credit card in the first three weeks and spending over $1,000 on average in the first month.
- Muse reported about 500,000 downloads and 250,000 active users in 12 days. In the US and Canada it had around 5 million downloads in its first 22 days, versus about 16 million for Threads.
Agent costs reflect who uses them
Founders post figures of hundreds or thousands of dollars a month to serve one agent user. Olivia traced this to David Pawlan's agent benchmark site, which tracks 170-plus agents and runs a community of 1,500-plus early adopters. The top use case there was coding and technical automation, even for consumer assistants. Assistants aimed at mainstream users see costs in the tens of dollars a month. So the high numbers reflect how technical today's users are.
A blank box does not spread
Josh said handing agents to a new user feels like a blank box. If you are not actively booking a flight, it is hard to know what to ask. On social media people copy what they see others do, and that creates cascading ideas. Because AI use is so personal, that sharing loop has barely started outside coding and creative work.
Mental Models & Frameworks
Save time versus spend time
- Eugenia, CEO of Wabi, has a line (Olivia said she might botch it): most people aren't looking to save time, they are looking for ways to spend it.
- Most consumer AI so far helps people do something a bit faster or easier.
- That is not a compelling daily value proposition for most people, which is why social media, entertainment and video lead consumer usage.
- Josh agreed and added that spending joy, not just saving time, shows up at the top of the data as people will ideas into reality.
Why startups beat labs on products
Olivia and Josh gave several reasons startups win on consumer products:
- Interface lock-in: big companies are unwilling or unable to cannibalize existing interfaces. Olivia said Google has not reinvented Docs, Gmail or Calendar, and OpenAI and Anthropic do best inside their own chat and coding interfaces.
- Speed: large companies move slower, with more checks and integrations. Olivia's example was a hardware device: OpenAI can build one, but everything takes longer.
- Rising tide: Josh said products like Whisper Flow and Granola were easy to start with, then got better as models improved, so utility grew without a rebuild.
- The test for startups: build a product that gets better as models get better but does not get absorbed as models improve.
The Great Expansion
This is a piece Olivia wrote on why consumer companies become enterprise companies so quickly. Products like Replit, ElevenLabs and Gamma grow through product-led growth, then get pulled into workplaces. Founders then add privacy, security and team plans. Canva took more than six years to reach real team and enterprise revenue, and these products get there much faster. It works best for tools that are personal but work-adjacent.
Compounding personal context as lock-in
Olivia's example was Tomo (the transcript spells it unclearly), a prosumer agent that connects to Slack, email and Drive. She said it holds playbooks of who she is and how she sounds that no other product has. The gap between an email that sounds 99.9% like her and one that is 85% like her is about 10 seconds of fixing versus 10-plus minutes. That gap is what stops her from switching. Josh pushed back on calling such products a "harness" or "wrapper". He said they are rich products where the model is one piece.
Trade-offs & Nuance
Usefulness versus privacy
- The more an assistant knows about you, the more useful it is.
- The more it knows, the less you want to bring it into contexts with other people.
- Olivia said no assistant has unlocked person-to-person network effects, partly for this reason.
- Josh added that trust in handing over email, life details and a credit card may be the biggest brake on mass adoption. People are good at judging what to share. Now they expect software to do the same.
Platform access cuts both ways
Agents depend on other platforms, and platforms decide whether to allow them. Olivia said Amazon won't let Muse browse and buy on its site, but Muse has hundreds of partnerships, including Shopify. Josh said that over time the stickiness will come from network effects, marketplaces and businesses that agents can deal with.
Labs versus specialized products
The labs are strong baselines, but specialized products often win a category:
- Audio: ElevenLabs and Suno rank near the top on traffic and spend. The labs have many other priorities, and music brings IP headaches.
- Images: OpenAI and Google took much of the casual traffic. Midjourney fell off the traffic ranking but is back in the revenue ranking, because power users want a model with taste they can tune.
- Video: Olivia called it a frontier where Chinese companies have an advantage, because they can train on any data.
- Josh's view is that value sits in bespoke interfaces and the software layer around the model, not just the model.
Practical Application
Rank your users by spend
- Do: split your paying users into deciles and see how much revenue the top 10% and top 1% drive.
- Then: study what those power users have in common, since the report found they cluster around building and making.
- Why it works: it shows whether you have a mass-market product or a power-user product, which changes pricing, support and roadmap.
Test a non-subscription model early
If your product has high serving costs, don't assume a subscription is the only way. Review which parts of the experience could be funded by ads or transaction fees, such as moments where users are in discovery mode. Check first how much of your usage is commercial intent, since the OpenAI example shows those queries are where ads fit.
Put usage meters in place
Josh warned that fast growth without meters on usage can create uncontrolled spend. Add per-user usage tracking and a credits or extra-usage option before a growth push, so heavy users pay more and light users don't subsidize them.
Show new users what to ask
- Do: add examples, templates or shared outputs so a new user sees what others have done.
- Why it works: Josh said the blank-box problem is real and sharing loops are how social products spread ideas.
- Then: measure how many users reach a first useful outcome after seeing an example.
Design for an agent caller
Olivia said creative tools that are friendly and accessible to agents will see strong tailwinds. Check whether an agent can use your product through an API or tool call, since she saw models pick up outside tools on their own.
Questions to Consider
- If the top 1% of your paying users each spend roughly 36 times what your median paying user spends ($903 versus $25 a month in this report's data), would your product and pricing change?
- Which part of your product helps people spend their time well, not just save time?
- If your costs per user fall by half next year, would ads or transaction fees replace subscriptions, and what would you need to build first?
- What would a first-time user ask your product on day one if they were not already a power user?
Bottom Line
Consumer AI is paid for by a small group of power users who build, make and sell, while most people still use it as a search replacement. The next wave depends on products that people want to spend time in, backed by ads and transactions and by interfaces that models alone don't provide.
Resources Mentioned
| Resource | Type | Why it was mentioned |
|---|---|---|
| Top 100 Consumer AI Apps, Seventh Edition (a16z) | Report | The subject of the episode, with spend data added for the first time. |
| Olivia Moore's piece "The Great Expansion" | Article | Explains why consumer AI products so quickly become enterprise products. |
| State of Markets (a16z) | Report | Sarah Wang's comment on power-law dynamics in enterprise and consumer AI. |
| David Pawlan's agent benchmark site and podcast episode | Website | Source of data on what early adopters do with consumer agents. |
Case Studies Mentioned
OpenAI's advertising run rate
OpenAI rolled out ads slowly through a network of partners and reached about a billion dollars in annualized run rate, based on an August figure. It had 1.2 billion weekly active users, plus features like login with ChatGPT and a wallet on the way. The lesson is that density plus personal context can make ads worth more to advertisers, if the ads stay clearly labeled and natural. Josh described seeing useful ads while researching a New York trip.
Muse versus Threads launch
Muse launched with strong buzz in tech, with around 500,000 downloads and 250,000 active users in 12 days. In the US and Canada, it reached about 5 million downloads in 22 days, while Threads had about 16 million. Josh said Muse is not pushing as hard, partly because serving each user costs more. Threads also needed critical mass for its network to form, while Muse is a personal tool. The lesson is that launch metrics must be read against how a product's economics and network effects work.
Plot, the AI note taker device
Plot is a note-taking device that started in China and came to the US. Users buy the device and a subscription. Olivia used it to show that big labs can build hardware, but they move slower than a focused startup.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| OpenClaw | Open source personal agent | Sparked the personal agent wave, then fell out of the rankings as traffic declined. |
| Muse | Consumer AI assistant | Big launch compared to Threads, and partnerships with hundreds of products. |
| Instinct | Consumer agent in messaging | Early growth stats and users spending over $1,000 in their first month. |
| Open Evidence | AI product for doctors | A dense, targeted audience that suits advertising. |
| Suno and ElevenLabs | Music and voice generation | Specialized products that lead in audio, a modality the labs have not focused on. |
| Midjourney | Image generation | Dropped off the traffic ranking but is back on the revenue ranking. |
| Whisper Flow, Granola, Superhuman | Work-adjacent AI tools | Examples of product-led growth that gets pulled into enterprises. |
People to Follow
Olivia Moore
Partner on a16z's consumer team and the author of the Top 100 Consumer AI Apps series, now in its seventh edition.
Josh Elman
Investor on a16z's consumer team who joined a few months before the episode. He argues that value in AI is moving back to the software layer and to network effects.
Elena Burger
a16z editorial partner who hosts the conversation.
Notable Quotes
"The top one percent user is spending $903 per month personally on their personal credit cards on AI." (Olivia Moore)
"The reality is people are now building software and they're building these really rich products." (Josh Elman)
"Most people aren't looking to save time. They're looking for ways to spend their time." (Olivia Moore, quoting Eugenia of Wabi from memory)
