Predicting the future is hard.
As an analyst, that is exactly what people hope I can do. Of course, no one can predict the future, but we can model various outcomes given the best data we have on hand.
So using the data, let's make a prediction about the future impact of AI on enterprise networks.
There are many ways to approach this, but the key area I will address in this blog series is the impact of AI on enterprise network configurations. The two key areas where our data are useful for this are predicting the budgetary impact of:
- AI-driven bandwidth demand increases.
- The ongoing move away from office-to-office connectivity to connecting users to assets in the cloud.
I’m not going to address how automation will change network management itself, but rather how those above changes impact the total cost of ownership (TCOs) of resulting WAN configurations.
Deep learning and LLMs require large transfers of data for training and inference. Palo Alto recently released a report that they saw an 890% increase in bandwidth demand last year. That is massive! Of course that is largely between data centers, but the implications for any branch office or site are still clear. Nokia predicts that AI-specific traffic will grow at a CAGR of 24% a year through 2033. Much of the work of AI is happening in the cloud, making traditional MPLS architectures with centralized breakouts inefficient and slow. Not to mention expensive
While we can’t predict much about the impact of AI on specific lines of business, we can certainly look at how large increases in bandwidth demand and cloud-centric hybrid architectures impact network costs. This blog series will address modeling several possible AI-age network configurations and how they will impact budgets. I do this all the time for customers of our WAN Cost Benchmark platform, which is the tool we are using to do this modeling.
Modeling Network Configurations
We have been running various hypothetical scenarios since 2019, but in 2025, we needed a major revision to keep pace with these new AI-driven configurations. This analysis will look at a standard and high-capacity MPLS network and compare that to:
- a Conservative Tiered Sites WAN,
- a standard Tiered Sites WAN,
- a pared-down Remote Worker Hybrid WAN,
- an Internet-First WAN,
- and the most radical revision, a Hub and Spoke WAN.
Other than the Remote Hybrid WAN, all of the hypothetical network scenarios have the same physical geography across 150 global sites. Our aim when we created this original hypothetical network in 2019 was to capture a smaller enterprise/multinational network (and to keep the comparison work easier for our analysts!) and to make sure that it captures trends in key markets all over the world. Our network includes seven data centers among the 150 sites in key regional data center hub cities:
- Tokyo
- Hong Kong
- Frankfurt
- London
- Ashburn
- San Jose
- São Paulo
We don’t specify a vertical for this hypothetical enterprise; however, it is likely closest to a business services company (think consulting, accounting, law firm) as most sites are in a key metro’s downtown or suburbs. This does not represent an industry with rural or remote sites that have very long access lines; most of our access lines are in the 0-5 km and 6-15 km ranges. While our sites fall mostly in the U.S. & Canada and Western Europe, the network spreads across the globe.
Percentage of Sites in Each Subregion–All WAN Scenarios
Note: Each section represents the total contribution of the listed region to the global WAN size.
Hypothetical Network Map
Dual MPLS WAN
Our baseline or control Dual MPLS network is meant to represent what enterprises are leaving behind: an all-MPLS network with active-active dual ports and fully redundant access lines. The geographic distribution of the average bandwidth available at each site featured below reflects the total bandwidth availability of the network and how sites are distributed within each region. We will use this chart to compare changes to the network configurations and how they impact available bandwidth. This chart represents the total bandwidth of all underlay products at each individual site, averaged across the entire subregion. For the starting network that is simply twice the MPLS port we selected as it is an active-active MPLS network.
Average Total Bandwidth per Site in Each Subregion—Dual MPLS WAN
Note: Each bar represents the average total site bandwidth, including multiple ports or underlay services, across all sites in the listed subregion.
- East Asia has the highest bandwidth, with an average of 2,868 Mbps, because it has two data centers (DCs) and relatively fewer sites than Western Europe or the U.S., which also has two DCs.
- Similarly, the average in Latin America is brought up to an average bandwidth of 907 Mbps by the Brazilian DC.
- Other regions show much smaller average bandwidths, especially those that are expensive. In the network design, we kept cost in mind to reflect the reality that bandwidth may be limited in high-cost locations.
The average bandwidth per site globally is 1,298 Mbps; however, this is brought up significantly by the large capacities at the data centers and a few large headquarters or campuses. Typical sites are much lower, in the 10-200 Mbps range.
Distribution of Total Site Speeds—Dual MPLS WAN
Note: Each section represents the percentage of total site bandwidths that fall within each bandwidth range.
- Since the WAN has active-active MPLS ports, the total site speed is actually twice that of the single port, which is the number we focus on to determine the available bandwidth at any given site or across the network.
- Only about one-quarter of sites have a total bandwidth of 100 Mbps or less.
- About 15 sites, or 11%, are above GigE. These represent the seven data centers and a few large regional hubs.
High Capacity Dual MPLS WAN
Our starting control network would prove strained under AI-drive bandwidth demands, so we added an additional control network. In the MPLS environment, many enterprises would renew on a three- or even five-year basis and perhaps only then address growing bandwidth requirements. This additional control is in place to demonstrate the scenario where perhaps a WAN manager in 2025 was renewing a network they had purchased in 2022 and now needed to boost bandwidth by at least 50% across the board. If this enterprise was not yet able to adopt SD-WAN/SASE and still needed MPLS at every site, we wanted to see the cost impact of a larger MPLS network.
Average Total Bandwidth per Site in Each Subregion—Dual MPLS WAN vs. High Capacity Dual MPLS WAN
Note: Each bar represents the average total site bandwidth, including multiple ports or underlay services, across all sites in the listed subregion.
- The average bandwidth per site comes out to 2,133 Mbps, which is a 64% increase over the original MPLS network.
Distribution of Total Site Speeds—High Capacity Dual MPLS WAN
Note: Each section represents the percentage of total site bandwidths that fall within each bandwidth range.
- We no longer have any bandwidths in the 11-20 Mbps range.
- The smallest bandwidth range, 21-50 Mbps, has dropped from about 13% to 3% of the global WAN.
- The percentage of sites falling in the bandwidth range 501-1,000 Mbps has decreased by about 1%, now accounting for 13% of sites.
- Around 25% of the sites now have a bandwidth larger than GigE.
So, how do these larger bandwidths impact cost? Well, the cost per bit goes down with higher capacities, so while we expect costs to go up significantly, we did not expect costs to go up by as much as the bandwidth increase.
Original Dual MPLS and High-Capacity Dual MPLS Scenario TCOs
Note: Each column represents the total annual cost of ownership for that WAN scenario, broken out by product in each color section.
- While the average bandwidth available per office went up by 64%, the TCO only rose by 43%, going from $8.1m to $11.6m.
- While the unit cost advantage is clear, this would still represent a drastic change in budget for the network to accommodate the increasing bandwidth demands of the modern enterprise.
While no one would want to increase their network budget by 43%, that isn’t the end of the story. An all-MPLS architecture is inefficient for internet-driven cloud traffic, and old class of service technology cannot take advantage of application specific policy and performance management when enterprises adopt SD-WAN.
In the next posts we are going to look at various hybrid scenarios that take advantage of SD-WAN, local breakouts for better cloud connectivity, and increased bandwidth.
