For as long as most ISPs have been building networks, capacity planning has followed one basic rule: traffic rises in the evening when people get home and stream video, and falls off overnight. You size your upstream links, your oversubscription ratios and your maintenance windows around that daily rhythm. Two pieces of industry data published this year suggest that rule is starting to break -- and the cause is AI.
Long-haul demand just doubled, and it isn't shaped like a spike
Zayo's 2026 Bandwidth Report, released in July and built from purchasing data across nearly 6,000 customers, found that long-haul fibre demand doubled in 2025 while metro-level demand rose by as much as 20x in some markets. The report's core argument is less about the size of the growth than its shape: AI training, inference and distributed compute generate large, continuous data flows between data centres, colocation facilities, cloud regions and edge nodes, rather than the short, predictable peaks that video streaming or office hours used to produce. Demand for 400-gigabit wavelengths and dark fibre is rising alongside it, driven largely by hyperscalers, cloud providers and the new wave of "neocloud" AI infrastructure buyers.
Operators are already seeing it on real links
This isn't just a forecasting exercise. At an industry webinar covered by Capacity Media in August, Telstra Digital Infrastructure and advisory firm Teneo described what they called an "architectural effect" showing up on specific data-centre-to-data-centre corridors: sharper traffic peaks and far less tolerance for latency or congestion than traditional patterns allowed for. Telstra gave a concrete example -- a sustained AI traffic flow of hundreds of gigabits, running for hours over several days, that reversed the network's usual direction of flow and forced an unplanned resize of an interconnect that hadn't been built for that pattern. Both speakers pointed to agentic AI as the next step up in difficulty, since agent-to-agent traffic is persistent and needs guaranteed low latency -- delays between cooperating agents compound rather than just annoying a single user.
Why this matters even if you're not selling to hyperscalers
Most ISPs reading this aren't negotiating wavelength contracts with cloud providers. But the same underlying shift is reaching residential and small-business connections from the other direction. Cloud backup tools, AI assistants and agents, video calling, and cloud-synced storage all generate background traffic that doesn't wait for "peak hours" -- it runs whenever the device is on. The old assumption that daytime hours are quiet enough for maintenance, or that an oversubscription ratio tuned to evening peaks will cover the rest of the day, gets less reliable as more of a subscriber base runs always-on AI-adjacent traffic in the background.
What ISPs can actually do about it
You don't need hyperscaler-scale infrastructure to respond sensibly to this trend:
- Watch utilization by time of day, not just peak hour. If the gap between your quietest and busiest hours is shrinking, your maintenance windows and upstream contracts need to reflect that.
- Keep visibility at the subscriber level. Monthly aggregate usage hides which packages or clients are driving sustained background load -- real-time session data shows it.
- Revisit package shaping with real data, not guesswork. Dynamic, RADIUS-level control over individual sessions lets you respond to changing usage patterns without renegotiating your entire upstream contract.
That last point is exactly where platforms like XpressRADIUS fit in. Real-time RADIUS control over MikroTik routers means an ISP can see live session and bandwidth data per subscriber, not just a monthly total -- and adjust packages, caps or disconnect rules in response to how usage is actually shifting, rather than reacting months after the fact.
Want visibility into how your subscribers are actually using bandwidth, in real time?
Start a free trial