AI is not mainstream yet. A small group of power users is spending a fortune on it.

Nearly half of US consumers say they use AI, but only 4.5% pay for a major assistant. Meanwhile the top 1% of AI spenders are averaging $903 a month, and what they buy says a lot about where AI is heading.

$903 a month. Not a year. A month. That is the average AI spend of the top 1% of AI spenders, according to new consumer data from Andreessen Horowitz.

My first reaction was that it sounded ridiculous. My second was that I could see exactly how it happens. Because the people spending the most on AI are not simply paying for better chatbots.

They are buying tools that help them build, automate, create and get work done. And what they buy tells us more about where AI adoption really is than any leaderboard of chatbot users.

AI use is wide. Deep use is not.

a16z has published its seventh Top 100 Gen AI Consumer Apps report, and for the first time it includes US consumer spending data. Traffic tells you what people try. Spending starts to tell you what they value enough to keep paying for. The gap between the two is large.

Nearly half of US consumers now say they use AI, but only around a quarter use it daily. As of August, just 4.5% of eligible US consumers in the data had a personal subscription to ChatGPT, Claude or Gemini. A year earlier it was 2.1%.

So the paying market is growing fast, but it is still a small slice of the population.

Then you reach the power users

Among people who do pay, spending is extraordinarily concentrated. The top 1% of payers accounted for 19.5% of all observed consumer AI spending. The bottom half, combined, accounted for 16.6%.

And that top 1% averaged $903 a month, against a median AI payer of about $25. That is not a small gap. It is almost two different markets. At one end, someone pays for a single assistant because it is occasionally useful.

At the other, someone is assembling an entire AI toolkit around the way they work.

When AI becomes infrastructure, one subscription is not enough

If you use AI casually, one general assistant may be all you need. If you build with it, the picture changes. A model subscription. A coding tool. An automation platform. An image or video generator. A research tool. An agent platform. API usage, hosting, databases, monitoring.

Before long AI is not a product you use. It is part of your working infrastructure. The data reflects it: the biggest spenders lean on automation and building tools like n8n and Manus, alongside creative products like HeyGen and Higgsfield. Spending seems to deepen when people move from asking AI things to making AI do things.

The real divide is not users versus non-users

We have talked about AI adoption as two groups: people who use AI and people who do not. I am less convinced that is the useful line. The bigger divide may be between people who use AI as a tool and people who build AI into how they work.

The first group opens ChatGPT when they need something. The second runs workflows, automations and agents across different parts of their work. Once you cross into the second group, your questions change. You start caring about integrations, reliability, permissions, data access, cost, and whether the things you automated actually happened.

The conversation stops being about prompting. It becomes operational.

A warning hidden in the $903

Spending $903 a month on AI does not mean getting $903 of value. One of the less glamorous parts of this boom is tool accumulation. You subscribe to solve one problem, then another tool does something slightly different, then a new model launches, then an agent platform, then a higher tier for more limits.

Suddenly you have a stack of overlapping products and no clear idea which ones earn their place. I have had a version of this myself. Once you are genuinely building with AI, it is easy to add another tool because it might remove another bottleneck. Adoption should not be measured by how many subscriptions you have collected.

The useful question is what each one is actually doing: saving time, producing work, improving a process, generating revenue, or removing repetitive tasks. If the answer is none of those, it is not transformation. It is just another direct debit.

AI still has a mainstream problem

There is another signal in the numbers. If AI is already this powerful, why are only 4.5% of consumers paying personally for a major assistant? Part of it is that the free versions are very good.

But the industry has also become excellent at selling AI to people who already know why they need it: developers, creators, founders, researchers, enthusiasts. The harder job is giving everyone else something valuable enough that they do not have to become an AI expert first. That is why the push towards personal AI agents is worth watching.

The next big consumer product may not ask people to learn prompting at all. It may just do something useful: book something, organise something, find something, finish something. That is a very different offer from paying for an empty chat box and being told the possibilities are endless.

Agents could change the economics

a16z points at another shift. Most AI products still make money through subscriptions or usage charges. Among the AI-native web products it ranks, 84% offer subscriptions, and few currently earn through advertising or transaction fees. Agents could change that.

If an agent actually completes transactions or does useful work, businesses may charge for the value of the action rather than access to the software. That looks less like traditional software and more like paying for an outcome. We are not there yet, but it is a direction worth watching.

AI adoption is entering its second phase

The first phase of consumer AI was about access: give everyone a chatbot and see what happens. The second phase looks different. People are building AI into workflows, specialist apps are forming around the foundation models, agents are starting to take actions, and a small group is already spending serious money assembling AI around their work.

That does not mean everyone will spend $903 a month. They almost certainly will not. It means we are seeing what happens when AI stops being something you occasionally visit and becomes something you work through. For most people, AI is still a chatbot. For its heaviest users, it is already becoming infrastructure.

That gap may tell us more about where AI is heading than another count of who has the most users.

Where to go next

• How many AI tools does your business actually need? The tool-accumulation trap, and how to tell which subscriptions earn their place.
• AI productivity is not the same as business value Busy is easy. The question is whether the work actually moved.
• When AI quietly becomes business infrastructure What changes once AI stops being a product you open and becomes part of how you work.

Related: How many AI tools does your business actually need?.