If AI use really does track your internet connection, the digital divide just got a new floor

If you sit on a broadband funding committee, run a venue, or sign off on connectivity spending for a school, a city or a business, there is a research finding circulating right now that you will probably hear quoted at you. I want you to understand exactly what it says, what it does not say, and why the part that should worry you is not the part being marketed.

The short version: a firm called Recon Analytics surveys 6,000 people a week about how they use AI. They seem to have found a connection between the type of internet connection people have at home and how heavily they use AI tools. People on fiber optic connections are asking 28.5 questions per AI session. Fiber sessions are described as 50 percent more intensive than sessions on cable networks. Do that math backwards and it means someone on a cable connection is asking roughly 14 questions per session.

Two pieces of translation before we go further. "Fiber" means the internet line coming into a building is glass strands carrying light, which is the fastest and most responsive consumer option available today. "Cable" means the same coaxial line that historically carried television, which is still fast but generally not as responsive. An "AI session" here just means one sitting with a chatbot or similar tool, and the count is how many back-and-forth questions the person asked before they stopped.

My first caveat, up front

I don't know where this research comes from or how it correlates. That is not a rhetorical dodge, it is genuinely what I can and can't verify. What I do know is that the research came from the Fiber Broadband Association. There might be some advantage in why they are the ones mentioning this, rather than someone mentioning it at a telco or a cable expo. The fiber providers obviously win from this finding, and it is already starting to show up in broadband funding arguments. Treat it as an interesting read, and treat it as advocacy, because it is both.

Why a good number for fiber is a bad signal overall

The idea that if you've got a fiber connection you will do more with AI is actually a big red flag for me, because it starts to describe another digital divide.

We already know there is a digital divide in the United States and around the world between people who can afford high speed connectivity and those who can't. What this finding adds is worse than the old version of that gap. It is no longer just that some people can't get the same services as everyone else on a level playing field. It suggests that with advanced services like AI, if you're not on the right kind of network, you're at a significant disadvantage in using the tool at all.

Think about areas like South Texas and rural markets. Think about tribal lands, and places that have never traditionally had fiber optic connections. It puts the children there at a disadvantage. It puts the students at a disadvantage. The Fiber Broadband Association stands to win from this finding, because it is good material for pushing fiber as far as they can build it. But these are tragic conversations for people who can't get fiber, and for people who can't afford fiber even where it is available. This is another signal that the digital divide keeps growing every time a new technology or application arrives.

Does this change the math if you don't have fiber?

It has to. Anyone operating a network now has to think harder about what their customers are actually doing on those connections, not just how fast the connection tests.

The good news for wireless is that the equipment has moved. There are products like Tarana, a fixed wireless system that beams internet service to a home from a nearby tower rather than running a cable to it, that can deliver gigabit speeds over the air. Starlink, the satellite service, is pushing higher speeds, and other satellite providers are entering the market with high-end speeds too. On paper that should level the playing field for wireless operators, but we don't know what that looks like yet.

Here is the honest gap in the story. The research was done for the Fiber Broadband Association, so I wonder whether Recon Analytics looked at fixed wireless operators at all. For the operators running Tarana gear, what do their numbers look like? What do those AI sessions look like? I genuinely don't know what the current average utilization per user is on those networks. That would be a great question for someone in that specific space, and it is a fair question to put to anyone who quotes the fiber number at you.

Where a gigabit connection actually falls apart

Inside the building. It does not matter whether your internet connection is fiber, satellite or wireless if the Wi-Fi in the house is poor.

My son recently moved into a new apartment and called me because he needed to figure out how to get Ethernet cabling into the place. Ethernet is the physical network cable that runs from the internet equipment to a device, bypassing Wi-Fi entirely. He games a lot and streams what he plays, and that is his main source of income, so he needs the connection to be reliable, not just nominally fast.

Even though Wi-Fi is everywhere in people's homes now, there is still real demand for a wired copper or fiber connection inside the home. If the connectivity indoors is bad, the size of the pipe coming in from the street is irrelevant. You can't take advantage of speeds your indoor setup can't carry.

What changes for a venue when guests start using live AI

Latency becomes the number one concern, and honestly it already is. Latency is the delay between asking for something and the first response arriving, as distinct from throughput, which is how much data you can move at once. For a long time people cared almost entirely about throughput: how much capacity they had, how big the speed test number was. That misses the big picture, which is the whole experience.

It's not just about how much of something you can get. It's how quickly you can get it, where it's coming from, and how you can consume it. There are tools now, like Orb, that monitor overall network capability and give you a single experience score rather than a raw throughput figure. Being able to monitor that is essential, because that is what shows you the full picture.

For venues specifically, you have to consider your access point density. An access point is the box on the ceiling that broadcasts Wi-Fi, and each one covers an area called a cell. In large spaces carrying delay-sensitive traffic, you want more access points with smaller cells, because that gives every connected person a better quality connection instead of crowding everyone onto a handful of overloaded radios.

Capacity is not the only reason to build

People look at capacity and assume that means they are doing more. But most AI sessions are really about latency. You are talking about exchanges of a few kilobytes, text going back and forth between someone's device and the service. That's the typical large language model use, meaning the text chatbots most people have tried.

Then there are people moving larger files, using AI to generate images or video, who want an immediate result. Someone wants to say "make me a picture of a cat playing a piano" and see it right away. For that you need the low latency to handle the conversation and a big enough pipe so the high resolution image or generated video lands on the device immediately. It's a combination of the two. In an age where people expect everything instantly, the network has to deliver both.

Speed tests can look great while applications fail

That describes just about any enterprise network. Sometimes the connectivity is genuinely excellent, but the applications are hosted somewhere else entirely and people still can't use the tools they need. In large venues you'll see people getting online and offline perfectly happily while the applications run poorly. The network has to be built to handle that, not just to pass a speed test.

This goes back to the methodology I use when designing a network. I think about the devices, the people using the devices, the services that make those devices work, and the applications running on them. Miss any leg of that survey and you're leaving yourself open to a poor quality network.

What you need is a tool that monitors as much of that as possible. You're not only looking at throughput, and you're not only looking at jitter or latency. Jitter is the inconsistency in that delay, the difference between a steady response and a stuttering one. What you're really looking for is true application quality. Once you find something running slowly, there are options, including SD-WAN, which is software that intelligently steers traffic across multiple internet providers to get data where it needs to go by the fastest available path.

Why proximity to the processing matters as much as the pipe

This whole conversation about fiber users doing more with AI comes back to delivery. It isn't only the immediacy of the connection and getting the questions answered, it's the delivery of whatever service the person wanted, whether that's travel plans, weather updates, or pictures of cats playing the piano. It really doesn't matter what the end user is doing, because they are going to use it for everything. We've already seen that as adoption grows, people apply AI to just about anything they can think of.

Without a fiber connection providing low latency, or a wireless connection from something like Tarana providing low latency and high throughput, you really are at a disadvantage. Enormous numbers of people in this country and around the world can't access either, and that drives the gap further.

This goes hand in hand with the proliferation of data centers, the buildings where the AI computation actually happens. If there are more data centers closer to people, there's less delay in getting that data processed. There are groups providing power along with on-premises processing for individuals, so the AI is running in your own backyard, and Starlink is talking about building data centers in space, where as a customer you'd have almost direct access to the data center.

So the latency conversation isn't just about the pipe. It's about how close you are to where the data is being analyzed and where the AI is actually functioning. That's one part of it. The other part is what we do to make sure people on low speed or mediocre connections still have the ability to process AI quickly enough to take advantage of it.

What to do with this

• Monitor what's actually happening on the client device, not only at the network core. That applies to every operator, from a wireless internet service provider serving a rural county, to the airport Wi-Fi you sit down on, to the city park.
• Stop optimizing only for throughput. Measure latency, jitter and true application quality instead.
• Use tools like Orb that give you an overall network experience score rather than a raw speed test number.
• Increase access point density and shrink cell sizes in large venues carrying delay-sensitive traffic.
• Run Ethernet inside the home or the venue, because a gigabit pipe is wasted behind poor Wi-Fi.
• Put proximity to data centers and on-premises or edge AI processing into your latency equation, not just the size of the circuit.

Questions worth asking

• Who funded the study being quoted at me, and does the funder sell the thing the study recommends?
• Did the researchers look at fixed wireless or satellite users at all, or only fiber and cable?
• Are we measuring our network by speed test results, or by whether the applications people actually need are working?
• Have we surveyed all four legs: the devices, the people using them, the services behind them, and the applications running on them?
• For any building we fund or operate, is the indoor wiring and Wi-Fi good enough to use the connection we're paying for?
• How far away is the processing our users depend on, and does anything closer exist?

I'd encourage anyone operating a service provider network to make sure they're offering a robust service and giving people the opportunity to do what they want and need to do on it. If your network isn't optimized for handling AI transactions, that's not just latency and delay, it's also the throughput needed to deliver the immediacy people are looking for.

Want this kind of thing weekly, in audio? That's the show.