For years, wireless policymakers have debated a familiar question: How much spectrum will America need to support future demand?

For years, the answers have generally followed a predictable pattern. More people connect. More devices come online. Traffic grows. Networks expand. Capacity increases. Demand rises. And 90% of that wireless traffic travels over WiFi –not cellular networks.

A new analysis from Cisco explores how AI is poised to shake up this model. The next spectrum debate will not be about connecting smartphones. It will be about connecting intelligence—and much of it, again, over Wi-Fi networks.

Examining AI inference and agentic AI traffic, the report suggests that AI is not simply another application riding on top of existing networks. AI may fundamentally alter the way network traffic growth behaves. If so, the assumptions underpinning today’s wireless forecasts – including future spectrum requirements – need to be revisited.

As with network traffic that pre-dates the arrival of AI, most AI activity occurs indoors. Homes, schools, hospitals, offices, factories, warehouses, airports, and research campuses are rapidly becoming AI environments.  Indeed, Deloitte recently published a list of 130 “most compelling” AI use cases, of which the overwhelming majority are ones that would occur in an indoor location. These are places where Wi-Fi – not high-power macro cellular networks – serves as the primary wireless platform. Furthermore, Cisco reports that AI is not limited to smartphones and will be embodied in a host of devices – with most of those devices communicating on a Wi-Fi platform.

AI’s center of gravity is local. The AI software on a laptop. The AI camera in a factory. The AI diagnostic system in a hospital. The AI-powered robot moving through a warehouse. The AI assistant helping a student complete an assignment. All of those interactions begin on local networks… and all of them rely heavily on Wi-Fi to move AI traffic.

Given that use cases skew heavily in favor of indoor applications, understanding how that indoor traffic might behave differently is particularly important for Wi-Fi and the unlicensed spectrum that powers it. Cisco’s analysis, which reiterates the findings of a 2025 Brattle report that is perhaps even conservative in its findings, lays out several important distinctions:

1. AI systems operate continuously. AI systems generate requests, retrieve information, communicate with models, coordinate with applications, and exchange data with other systems – often with no direct human involvement. Instead of a person opening an app, an AI agent (running on a local computer overwhelmingly likely to be connected via Wi-Fi) may complete an entire task autonomously, interacting with multiple systems along the way. Cisco’s research suggests those interactions can be far more network-intensive than traditional user activity. In fact, agentic AI tasks were found to generate approximately 450% more traffic than performing the same task manually – a trend that highlights how the traffic burdens on Wi-Fi networks are likely to grow significantly in the coming years.
2. AI traffic lasts longer. The company found AI inference flows persist approximately twice as long as traditional web transactions. That may sound like a technical detail, but for wireless networks, duration can be just as important as volume. Longer flows mean greater airtime occupancy, more contention among devices, and increased pressure on shared spectrum resources. Wi-Fi 7 is tailor-made to help manage this kind of demand, using wider channels, lower latency, and more efficient coordination to move more data with less waste – but even the most efficient protocol can become congested if traffic keeps growing and Wi-Fi networks do not have enough spectrum bandwidth to work with.
3. AI traffic is increasingly bidirectional. For decades, broadband networks have been optimized primarily around downloads—streaming movies, loading webpages, and consuming content. AI changes that equation. Prompts, images, videos, telemetry, context windows, sensor feeds, and machine-generated data all travel upstream before AI systems can generate responses. Cisco found that roughly 9% of AI inference flows contain more upstream traffic than downstream traffic. By comparison, only 0.5% of traditional web traffic behaves that way. That shift matters because Wi-Fi is especially well suited to handle more symmetrical traffic: unlike cellular networks, which are typically engineered around download-heavy consumer use, Wi-Fi can use wide, local, shared channels to move large amounts of data efficiently in both directions.
4. AI traffic is increasingly autonomous. A future powered by AI assistants, AI agents, machine vision systems, digital twins, robotics, and intelligent infrastructure will create traffic that is generated by software, not people. Machines do not wait for business hours. They do not sleep. They do not pause to watch television. They continuously exchange information, update models, analyze environments, and coordinate actions. And because the overwhelming share of these machines are in indoor or fixed facilities – offices, labs, factories, hospitals, warehouses, etc – Wi-Fi will carry the lion’s share of their AI-driven traffic.

Cisco projects that AI inference could account for approximately 25% of total network traffic by 2035. But are today’s spectrum planning assumptions capturing the full impact of AI? That question is particularly important for unlicensed spectrum.

The United States took a historic step when it opened the full 1,200 megahertz of the 6 GHz band for unlicensed use. That decision provided the foundation for Wi-Fi 6E and Wi-Fi 7 while helping to ensure sufficient capacity for the next generation of applications. 

But the emergence of AI will accelerate Wi-Fi demand faster than policymakers anticipated when that decision was made.

One important lesson from Cisco’s research is that AI may fundamentally alter the wireless demand curve itself, pushing America toward a Wi-Fi capacity crunch.

If future traffic is more persistent, more autonomous, more uplink-intensive, more device-dense, and increasingly driven by software agents rather than people, then traditional forecasting models may no longer provide a complete picture of future capacity requirements.

Policymakers should be asking a new question: If AI changes the way wireless networks are used, what does that mean for the future of unlicensed spectrum and its ability to carry this substantial tonnage?

The answer will determine whether America remains prepared for the next decade of AI innovation—or finds itself playing catch-up once again.