Mitch Wagner of Fierce Network recently published an article that claims that AI is blowing up 30 years of traffic network assumptions. He claims that AI traffic is smoothing the daily peaks and dips in network traffic that all ISPs are familiar with. While every ISP is a little different, any ISP that serves a lot of end-user customers expects traffic peaks in the evening, smaller peaks during the daytime, and very low levels of network traffic at night.
Network engineers have always paid close attention to the peaks, which were the main factor in determining the size of network connections. Nobody wants to have a network that restricts bandwidth when customers want to use it the most. The cable companies learned this lesson the hard way during the pandemic when customers suddenly needed to work and school from home and found the broadband connections unable to meet their needs, particularly in homes where more than one person wanted to use the network at the same time. Every network engineer I know can cite the busy hour, busy day, and busy week on the networks they manage.
Wager says the peaks in traffic are evening out and that networks are seeing a more consistent demand throughout the day. Wagner cites Ed Fox, the CTO of MetTel, from an interview given for a Fierce Network Research report.
I have no reason not to believe Mr. Fox. However, MetTel is an ISP that operates in major urban centers and likely serves the kinds of businesses that have become heavy AI users. It’s not hard to imagine that an urban ISP serving businesses might be seeing a drastic change in traditional network traffic patterns. But I have to think that MetTel and other urban business-centric ISPs didn’t have the same traffic patterns as other ISPs before the advent of AI.
I work with a number of ISPs and I have not heard anybody talking about a big change in traffic patterns. My clients work in a variety of markets, from rural to urban, but none concentrate on the urban market business that MetTel is experiencing in places like New York City.
However, I appreciate the article, because if it’s true, then most ISPs, except for fully rural ones, might eventually see some shift in traffic patterns due to AI. I am curious to get feedback from this blog from IPS to hear if anybody is seeing anything like the big changes being experienced by MetTel.
The Wagner article also claims that upstream bandwidth is growing faster than downstream, reversing a 30-year design assumption built around heavy downloads. This is something that has been well documented by OpenVault. They’ve been shown that upstream usage has been growing faster than download usage starting in 2024. In the first quarter of 2024, national average upload usage increased by 13.2% compared to 7.7% for download. In the first quarter of this year, average upload growth was at 19.8% while download growth was at 7.8%. OpenVault credits most of the growth in upload usage to computer software synching with data centers.
Wagner claims the upsurge in upstream usage comes from AI inference traffic moving towards the edge, particularly for video processing. What he means by that is big growth coming from uses of video cameras for functions like AI-driven video surveillance at retail locations, camera-equipped wearables, and cloud-based operational technology in applications like oil and gas asset management. I hope that the big national companies that monitor traffic begin tracking this issue. I’d love to hear more about the trends in specific traffic, like video surveillance and wearables.
Network engineers all understand that upload traffic is usually a tiny fraction of download usage, with average download usage a dozen times more than upload. With the possible exception of cable companies, upload usage is largely an afterthought for network engineers, who barely consider it when sizing and designing networks.
I have no doubts that there are localized situations where AI traffic is making a big difference in network traffic. But for ISPs that mostly serve residential and small business customers, I still have to think it’s a tiny, possibly unnoticeable blip.