Last week I took part in some fascinating discussions about the use of artificial intelligence and machine learning for spectrum management. The venue was the International Symposium on Advanced Radio Technologies (ISART), a U.S. government–sponsored conference hosted by the Institute for Telecommunication Sciences (ITS). ITS is the science and engineering laboratory of the Commerce Department’s NTIA.

ISART mainly a place where discussions of leading-edge work on wireless networks leads to forecasting the capabilities of future systems and laying out blueprints for securing resources and removing obstacles. Past conferences dealt with propagation models for cell sites, while this one, in part, examined ways to create propagation models from real measurements.

The networking industry has been using AI and ML in production systems for 30 years, initially using backtracking algorithms, written in the Prolog programming language, to predict and resolve equipment failures in enterprise networks too complicated to be managed through conventional means. While common discourse about AI ranges from the fabulous to the bizarre, AI in networking is based on real experience.

Where We’re Going

Where we are is in a state of massive investment – often fueled by exuberant speculation – as well as public anxiety about the effects that the AI buildout will bring to labor markets and the environment. Hyperscale data centers in particular have become a bogeyman for fears about the next chapter in technology evolution.

The public is concerned and confused about the race to AI dominance between the US and China, the impact of AI on jobs, data scraping to feed AI models, and the often laughable hallucinations we get from AI chatbots. In lieu of studying these and the other AI issues, public anxiety is zeroing in on data centers.

The public doesn’t want them in our neighborhoods, but we also don’t want Kevin O’Leary wrecking fragile ecosystems in remote parts of Utah. Without powerful new data centers AI will stall, so we need to solve the infrastructure problem one way or another. The industry is in the mode of “we have real work to do” on data center messaging, step one of finding a workable resolution.

AI is Genuinely Useful

While every shiny new thing coming out of Silicon Valley is over-hyped, most of them turn out to be genuinely useful and desirable. Programmers, writers, analysts, and students certainly find that AI LLMs increase their productivity, even if their results need to be carefully examined.

AI helps wireless carriers site their towers and employ advanced RF mechanisms such as beam-forming more efficiently. It also works for consumer versions of these problems such as WiFi access point location and home theater design.

I’ve been genuinely blown away by how helpful AI can be on this latter exercise. It knows how to blend equipment into seamless designs that overcome manufacturer focus on narrow parts of the problem, and it knows all the tools that professional installers use on home theaters.

Comic Relief

While talented and serious people are developing AI systems and learning to use the ones created by others, others struggle to catch up with the rocket ship. The cable industry draws smirks by setting itself up as the champion of WiFi, a technology that developed no thanks to them – and often despite their early attempts to neuter it.

Now they’re telling us that AI is all about WiFi. Huh? The argument goes like this:

Every AI request follows the same journey. It begins in a data center, travels over high-capacity broadband networks, and reaches users over Wi-Fi. That last hundred feet is where nearly every AI interaction actually happens.

Excuse me, but the AI transaction actually takes place in the data center. The very short first & last segment of the data’s journey can take place over WiFi, Ethernet, 5G, USB, or even Bluetooth without changing the nature of the AI chat; you’ll get the same answer in every case.

AI is Compute Intensive

Companies are building all of these data centers because AI is all about computation, and that means power. Lots and lots and lots of power. We can’t do very much of that on a battery-powered device like a smartphone, a tablet, or a laptop. Moreover, the battery-powered device that sends the initial query is most useful when it’s either fully mobile or fully stationary, plugged into Ethernet and wall power.

Cable also makes the mistake of assuming that AI – like IoT before it – sends lots and lots of data all the time. It turns out that AI queries, like the web, are highly asymmetric, consuming much more data than they create. AI data center networks certainly can be busy, but most of their traffic is in-building and over a wire. AI may well mean more network traffic, but we’re probably looking at 50 – 100 percent increases rather orders of magnitude.

And even if WiFi were to be the most important part of the AI ecosystem, it wouldn’t be an advantage for any nation because we all have WiFi that’s good enough to use. Real WiFi advocates tell us that the 6 GHz band works best when using the same 80MHz channels we had in the early days of Wi-Fi 5.

So no, cable guys, America’s potential advantage in AI will come down to our models, our training databases, our engineers, our data centers, our fixed and mobile networks, and our appetite for investment. Taking credit for other people’s work is just plain rude.