ChatGPT reached 800 million weekly users in three years.
That number is extraordinary. It may also reveal the limit.
In Altimeter’s chart, ChatGPT has the steepest adoption curve at equivalent age. Steeper than YouTube. Steeper than WhatsApp.
Then you notice where the curve begins to flatten. Well below the Core Utility tier.
That matters.
This is not a launch problem. It is not a distribution problem either. It may be a category ceiling.
Scrolling TikTok is a reflex. Opening WhatsApp is often a reflex too.
Using a chat-based AI is different. You have to decide to use it. You have to formulate intent. Then wait. Then evaluate the answer. Then refine the prompt if it misses.
That is a different loop.
Passive-consumption products rise fast because they ask almost nothing of the user. A chat box always asks for cognition.
That is why I think the race to build a better chat box may be the wrong race.
The AI products that win over the next decade may not be destinations at all. They may be capabilities embedded inside workflows that already exist.
Inside the email being written. Inside the document under review. Inside the handheld scanner a delivery driver already uses. Inside the enterprise software where the decision is actually made.
Pure-play chat products must pull the user back every time.
Operational software does not. It is already in the loop.
That difference may decide where enterprise AI compounds, and where it stalls.
I still do not think we have a clean answer on which software category gets there first.
But I am increasingly convinced the next decade of AI will be decided less by seat counts, and more by ambient presence.