We’re breaking down how AI would possibly impression enterprise networks in data-driven deep dives. Utilizing TeleGeography’s WAN Price Benchmark—a custom-made platform that makes it simple to mannequin and observe your WAN community prices—we’re modeling potential real-life community setups and configurations for AI.
In these posts, we’re evaluating the whole price of possession (TCO) of a legacy MPLS community—one that also represents many multinationals’ WANs in 2025—to varied hybrid or internet-first alternate options designed to accommodate adjustments introduced by AI and cloud. Whereas our argument for shifting previous MPLS-centered WANs was traditionally largely about price, it’s now far more about getting the community proper sized for the approaching AI-driven bandwidth crunch, resiliency, app-specific coverage and efficiency, and full utilization of cloud.
That is the final submit of this sequence. Meet up with the primary three posts:
Hub and Spoke WAN
Earlier analyses on this sequence explored tiered hybrid fashions and internet-first methods, each of which leverage the flexibleness and cost-effectiveness of web connectivity, managed and secured by an SD-WAN overlay. However on this submit, we study essentially the most radical revision but: the Hub and Spoke WAN.
The Hub and Spoke WAN mannequin represents a major problem to a elementary assumption of recent networking: the necessity for ubiquitous any-to-any connectivity. As an alternative of a meshed non-public community, reminiscent of MPLS, or a versatile overlay, like SD-WAN, this structure includes creating a non-public, “carrier-like” community the place department places of work (“spokes”) are instantly tethered to regional information facilities (“hubs”).
Adopting such a “carrier-like” community represents greater than a technological shift; it indicators a elementary transformation within the enterprise’s IT operational mannequin. The group transitions from being a passive shopper of a managed WAN service to an energetic supervisor of its personal non-public spine. This new position includes sourcing and managing a various set of parts for the spine and doubtlessly co-location services for the hubs themselves. This shifts the duty for end-to-end community efficiency, routing integrity throughout the core, and long-term capability planning from the service to the in-house IT staff. The choice, subsequently, can hinge as a lot on a corporation’s inner capabilities because it does on the technological deserves of the structure.
Modeling a Hub and Spoke WAN
These are the adjustments we applied at every WAN web site tier to create such a community:
- DCs— We added a spine of huge bandwidth wavelengths between the info facilities.
- All Tiers—One DIA port of an analogous pace to the unique MPLS, together with an EoMPLS non-public line of an analogous pace again to the closest regional information middle. We included native entry to all DIA websites, excluding these in the USA, Canada, and Western Europe.
- We saved MPLS for the Chinese language websites for personal connectivity at these websites, relatively than the DIA (we nonetheless added non-public traces for these websites).
- As a result of the places of work have a direct non-public connection, there isn’t any must undertake SD-WAN.
So why would an enterprise forgo SD-WAN, the market’s main expertise pattern? The reply lies in what this structure prioritizes as a substitute of final flexibility.
As a result of all non-public visitors from a spoke has a single, deterministic vacation spot—its designated hub—the complicated, dynamic path choice and coverage enforcement of SD-WAN turns into pointless. This architectural alternative has profound implications for each price and operational complexity. With necessary company visitors to the info middle going by non-public traces, safety can be a lot easier than in a hybrid or internet-first WAN structure.
For a selected class of enterprise, the advantages of absolute visitors isolation on non-public traces, radically simplified routing logic, and enhanced management over the community core outweigh the benefits of a totally meshed, internet-centric design. This isn’t a legacy structure, however a contemporary, strategic alternative for organizations whose enterprise necessities demand safety and predictability above all else.
Common Complete Bandwidth per Web site in Every Subregion—Twin MPLS WAN vs. Hub and Spoke WAN
Notice: Every bar represents the common whole web site bandwidth, together with a number of ports or underlay companies, throughout all websites within the listed subregion.
- The Hub and Spoke community is actually the identical bandwidth as the unique MPLS WAN. We did this to make the comparability as one-for-one as doable with an MPLS community and since bandwidth is much less of a problem in a P2P scenario the place visitors is on a non-public community destined for the info middle.
- In some areas, the common bandwidth is barely down, whereas it was extra prone to go up a bit in locations the place we swapped a decrease capability MPLS port with a barely bigger non-public line.
- All P2P connections, besides these between information facilities, acquired an entry line.
Distribution of Complete Web site Speeds—Hub and Spoke WAN
Notice: Every part represents the share of whole web site bandwidths that fall inside every bandwidth vary.
- Within the authentic, we have been including collectively two identically sized MPLS ports. Right here we’re including collectively the DIA port and the EoMPLS line, that are sometimes barely completely different speeds however very shut.
- Over half of the websites on this community have a complete web site bandwidth of 200 Mbps or under.
- About 10% of web sites have a complete web site bandwidth above 1,000 Mbps.
Unique Twin MPLS and Hub and Spoke WAN State of affairs TCOs
Notice: Every column represents the whole annual price of possession for that WAN state of affairs, damaged out by product in every shade part.
- Whereas the bandwidth is actually the identical, the Hub and Spoke community is extra reasonably priced, with the TCO lowering about 12% from $8.1m to $7.1m.
- This community is inherently safe (delicate company visitors goes over non-public traces not on shared infrastructure) and resilient (EPL over a meshed community) even when it lacks the power to share information between any two random places of work on the community.
- Entry is almost all of the price of the community–57%–with the subsequent largest share going to the non-public traces at 23%.
- There aren’t any extra SD-WAN prices, which retains the worth down in comparison with different eventualities.
The Hub and Spoke mannequin analyzed right here is outlined by its deliberate omission of an SD-WAN overlay. As an alternative, the enterprise successfully builds its personal non-public spine connecting key regional hubs, with all different websites connecting to this core. It is a trendy, hybrid structure that blends non-public and public connectivity with a transparent objective for every, and is customized for AI, particularly as including incremental bandwidth to EoMPLS hyperlinks shall be easy and cost-effective.
The Hub and Spoke mannequin’s readiness for AI is rooted in its centralized structure, which instantly addresses the distinctive visitors patterns and information necessities of synthetic intelligence workloads. By centralizing compute and information in well-connected hubs, the Hub and Spoke mannequin creates an environment friendly and safe atmosphere that’s purpose-built for the calls for of large-scale AI mannequin coaching and inference.
Mannequin Your Personal WAN
You do not have to make costly commitments to learn the way a lot your WAN will price. TeleGeography’s WAN Price Benchmark platform permits you to mannequin your WAN community, strive completely different configurations, and get actual insights into how your price range shall be impacted.
Some enterprises prioritize chopping prices, however we extra generally see enterprise clients use TeleGeography’s WAN Price Benchmark platform to arrange a extra resilient and higher performing, AI-ready community with out having to ask the c-suite for extra money. You’ll be able to get extra particulars and a video tour right here.
WAN Managers: We Need to Hear From You
Are you a WAN architect or IT infrastructure skilled? Do you’re employed with networks at a non-telecom service enterprise?
Take our WAN Supervisor Survey to know how enterprises like yours are designing their underlays to stability AI-driven bandwidth calls for, prices, and efficiency.
Your response to our temporary 10-20 query survey will earn you free entry to the complete report in addition to a TeleGeography infrastructure wall map! Click on right here for the survey.