Sarah Asks, Sam Interacts
The question of whether large language models would take hold in the home has been answered. In fact, our recent research with Parks Associates found that 59% of U.S. internet households now use AI tools, up from just 40% two years earlier. AI has firmly entered the mainstream.
Our own network telemetry tells the same story. In a study of Plume-enabled American homes, 86% touch an LLM¹ at least once a quarter, and two-thirds use one regularly. But adoption is no longer the most interesting question. The more important shift is how people are using AI, and what those changing usage patterns mean for the network. We think it marks the start of an Agentic Wave that will change the fabric of the internet.
Our research shows that 2 distinct subscriber profiles using LLMs in America are emerging. They’re reshaping broadband networks, one prompt at a time. Consider two fictional households. Sarah represents the growing number of people using AI as an everyday assistant, while Sam represents today’s early adopters and avid users of agentic AI. Together, they illustrate a broader shift that’s beginning to span generations, occupations and household types.
Sarah
Sarah uses LLMs the way most people do: on her phone. Six in ten of the regular LLM devices in our study are smartphones (“regular” means at least ten days of use in the quarter). Her style is the quick touch. She types a question into ChatGPT or Gemini, reads the answer, and closes the app. Similar to how someone would use a standard search engine. The typical active device logs about 14 minutes of LLM use over those three months, and that counts only LLM app usage. You can’t track individual website traffic beyond the browser level, so the real figure is certainly higher. Sarah is an experimenter. So are 15 to 20% of all homes. They dip in and out of LLMs without committing.
Juggling several LLMs is a power-user habit. For everyone else, the LLM of choice is most often ChatGPT. Sarah once tried a second LLM that came bundled with a device she already owned. Our data shows a burst of activity on her laptop. Within weeks, it fades. Only an occasional phone tap remains. We see this across the board: bundling buys trials, not habits.
Sam
Sam writes software, and it shows. A typical individual uses an LLM about once every three months. But Sam uses an LLM nearly every day (86 out of 91 days, to be precise). Sam runs AI agents too: we can see their signature on the network, flows where far more data leaves the home than enters it. And his hardware works as hard as he does. One workstation in our 1,000-home study produced over 10% of all measured LLM minutes. Sam is not going back. We see that homes that use an LLM ten or more days a month almost never stop or decrease.
Sam uses LLMs on his laptop. Laptops represent 35% of regular-use LLM devices but 49% of the measured minutes. He juggles LLM brands, too. A typical engineer’s device regularly uses nearly two of the four engineer-heavy LLMs we compared (ChatGPT, Gemini, Copilot, and Claude). Within this group, the market flips: the one with the highest user count isn’t the most used. The LLM engineers use least is the one they use on the most days, a pattern that points to coding work.
The Internet is Starting to Flow in Reverse
For decades, home broadband has rested on a simple assumption: people download far more than they upload. Streaming video became the modern internet’s defining workload, with services like Netflix averaging roughly 16 bytes downloaded for every byte uploaded. Even conversational AI, where users exchange prompts and responses with an LLM, is roughly balanced, generating about one byte uploaded for every byte downloaded.
Agentic AI flips the equation. When an AI agent is working on behalf of a user, it spends much of its time sending code, files, instructions and results back to cloud services. In our data, some of this traffic reaches 25 bytes uploaded for every byte downloaded. The total volume remains relatively small today, but the direction is what matters. For the first time, a fast-growing consumer workload isn’t primarily download-driven, but fundamentally the opposite.
Examine Sarah’s router and this shift is barely visible. LLM traffic still amounts to basically a rounding error, while streaming video and other traditional usage continues to dominate her household’s bandwidth. Examine Sam’s router, however, and a different pattern emerges. The applications placing the greatest demands on the network are no longer entertainment services, but AI workloads that depend on upload capacity, low latency and uninterrupted reliability. When a stream buffers, people wait. When an agent fails, work stops.
Why Every Persona Matters
It is tempting to build for Sam and forget Sarah. Resist it. Every Sam started as a Sarah, and the data shows exactly where one becomes the other.
What separates them isn’t how long a home uses an LLM, but how often. Count the days in a month on which a home is active and retention climbs a clear ladder: homes active one to three days a month have a 58% chance of still being active the next month; at four to nine days, 84%; past ten days, it never drops below 93%. Ten days is the threshold. Beyond that, usage may grow, but it doesn’t shrink. Those homes are 47% of all homes and 95% of all LLM minutes.
None of this shows up in public traffic data. The race between LLM providers is a race to move homes like Sarah’s up the ladder. Every home that crosses the ten-day mark brings the Agentic Wave closer.
Not All Upstream Traffic is Welcome
One more warning to keep in mind. LLM agents are not the only thing pushing data out of homes. Plume Security Labs recently found that popular streaming boxes quietly enroll homes in residential proxy networks without the owner’s knowledge. One network alone had signed up more than 60,000 home connections. A single infected device can make tens of thousands of connections a day, to thousands of destinations. Across our telemetry, these boxes touch more than 60,000 unique domains every day. None of that traffic belongs to the subscriber, but they almost never notice. Neither does the network, unless it can see each device. In some areas, the load was heavy enough to destabilize the neighborhood. There is an LLM twist too. Our researchers caught attackers routing ChatGPT and Google sign-ins through hijacked homes to look like real residential users. The Agentic Wave has a dark current, and it is already flowing. The full story is in our SuperProxy research paper.
Three Things for ISPs to Consider
First, it may be time for an LLM-friendly broadband plan. Sam’s agents need them.
Second, treat reliability as the core promise and move fast. A dropped stream costs a shrug. A dropped agent job costs real work. Homes like Sam’s will judge their provider on it.
Third, observe it early. Public data shows a flat market. The real movement happens home by home and device by device. Plume’s Application Intelligence and Behavior Intelligence show which homes are forming the habit, which LLMs they are settling on, and how close the Agentic Wave is to rolling through your network.
The Next Netflix Moment Has Already Arrived
Twenty years ago, broadband networks shifted because of a single change in consumer behavior: streaming video. It redefined what subscribers expected from their connection and what operators had to build to support it. Agentic AI may represent a similar turning point.
Not because it generates the most traffic today, but because it changes what “good broadband” means. Network quality is no longer measured only by how quickly content reaches the home. Increasingly, it is how reliably work leaves it.
The homes leading this transition are still a minority. But they are already showing us what tomorrow’s broadband workloads will demand. As AI moves from conversation to delegation, the networks that succeed won’t simply deliver more bandwidth. They’ll understand the applications, behaviors and outcomes that matter most to subscribers.
The Agentic Wave has begun. The question is no longer whether it will reshape broadband, but how quickly operators will recognize the shift.
¹Our study covers the major consumer LLMs: OpenAI’s ChatGPT, Google’s Gemini, Microsoft’s Copilot, Anthropic’s Claude, Meta AI, xAI’s Grok, and Perplexity. It also includes their developer and API channels, such as Claude Code and Codex.