WAN Modernization with Segment Routing
In this white paper
- Executive summary
- 1. Why modernize now: The transport inflection point
- 2. Fast reroute with TI-LFA: Sub-50ms as table stakes
- 3. Traffic engineering for everyone: Delay-based routing
- 4. Delivering services over the transport: L3VPN and L2VPN
- 5. Network slicing and deterministic services
- 6. Operational simplification: Collapsing the control plane
- 7. Choosing your path: SR-MPLS or SRv6
- 8. The realistic AI angle: Readiness, not requirement
- 9. The migration case: Why Segment Routing, either way
- References
- Download
Executive summary
For most of the last two decades, we built transport networks around one scarce resource: bandwidth. Capacity planning, circuit design and service tiers all came down to how many bits a link could carry. That era is over. 100G is now common across enterprise, service provider, utility and government backbones, 400G is well into deployment, and 800G and beyond are on the horizon. Raw capacity is no longer the differentiator it once was. What matters now is latency, resilience, deterministic behavior and the ability to deliver services over a shared transport. Traffic has also changed shape, pushing far more upstream than the old download-heavy model expected.
Segment Routing (SR) is the practical way to modernize the WAN for this new reality, in either of its data-plane forms, SR-MPLS or SRv6. SR delivers sub-50ms protection, delay-based traffic engineering, Layer 3 and Layer 2 VPN services, network slicing and a much simpler control plane, all on a standards-based, multi-vendor architecture. It also lets us be honest about AI. Segment Routing is not required for AI, and it does not carry AI training traffic. What a modernized SR transport does is prepare the network for the traffic AI brings, so that when an organization makes that move, the WAN is already built for it.
1. Why modernize now: The transport inflection point
The economics of the WAN have flipped. When 100G is standard, 400G is widely deployed, and 800G and 1.6T are on the horizon, adding capacity is rarely the hard part of meeting a service-level objective. What determines the quality of a voice call, a video stream, a real-time application or an interactive AI session is not the size of the pipe. It is how the network chooses its paths, how fast it recovers from a failure and how well it separates one class of traffic from another. Bandwidth is table stakes now; behavior is the differentiator.
Quality of Service deserves specific mention, if only because it is easy to under-rate. Mixing real-time traffic with bulk on the same link is not new, and for years plenty of networks ran with little or no QoS and got acceptable results. If the pipe was large enough and average utilization stayed low, the occasional collision between a voice packet and a big transfer was rare enough to live with. But that was always a bet on headroom, not a guarantee. Average bandwidth does nothing for a microburst, the sub-second spike when several flows contend at once, and a single large flow filling a queue can still degrade a latency-sensitive application on a link with capacity to spare on paper. Overprovisioning made the problem rare, not absent; strictly, the traffic always warranted QoS, and many operators simply chose to live without it.
What is shifting now is the competing load. Bursty uploads, backup and replication still generate those microbursts, while AI adds long-lived flows that hold capacity rather than spiking and clearing, so there is less quiet headroom to absorb the collisions. The right answer has not changed; the margin for skipping it has narrowed. Classification, marking and scheduling that protect priority traffic and isolate it from bulk are how you do it, and Segment Routing reinforces the point: the same slicing and Flex-Algo constructs that steer traffic by latency also give QoS policy its natural boundaries.
This is not just a design preference; it shows up in measured traffic. Analysis of live service-provider flows in Cisco's AI Impact on Wide Area Networks report[1] found that AI inference flows behave differently from ordinary web traffic in ways that bear directly on queuing. They last about twice as long (a median near 1.3 seconds for inference flows against roughly 0.6 seconds for conventional web) and they move at a far lower sustained rate, roughly a tenth of the median web-flow rate. Rather than arriving in bursts, they deliver smooth, steady throughput, because token-by-token generation effectively paces the flow. The report's own conclusion is that these different rate and duration profiles may warrant different QoS treatment for AI traffic than for conventional web traffic, so the case for QoS differentiation is now backed by measurement rather than intuition.
Traffic patterns have shifted in ways older designs never planned for. Consumer access networks were provisioned for asymmetric consumption, because we assumed users consume content, not produce it. Network-wide, downstream-to-upstream ratios ran on the order of 10:1 and, driven by video streaming, climbed even higher over the last decade [2]. That model shaped how we built transport. In aggregate, streaming keeps overall internet traffic downstream-heavy to this day, but that balance is breaking down on the networks carrying business and real-time traffic rather than consumer video. Upstream volume is growing faster than downstream, and the workloads driving modern demand (content creation, cloud collaboration, backup and disaster-recovery replication, and AI interaction) are far more symmetric or outright upstream-heavy. Anyone who works with AI tooling sees this firsthand: upload a set of images, a video or a large batch of documents, and a link that used to be download-dominated is suddenly pushing hard in the other direction. A transport network built solely for the old ratio is quietly mismatched to how it is used today.
Measured traffic bears this out, though at the level of individual flows rather than the network-wide aggregate above. Within a single flow (how much comes down for each byte sent up), the Cisco report[1] found ordinary web flows heavily one-directional, a median near 145 to 1, while AI inference flows are far less lopsided at about 21 to 1. Both still lean downstream, but an AI flow carries proportionally far more upstream traffic than a web flow does. The same pattern shows up a second way: roughly 9% of AI inference flows send more data up than down, against only about 0.5% of typical web flows, which makes an AI flow some eighteen times more likely to be upstream-heavy. These per-flow figures run higher than the aggregate because streaming still dominates the total bytes, but the direction is clear, and it should grow more pronounced as usage shifts toward agentic workflows, which push far more up to the model: longer prompts, retrieved documents and tool outputs chained across many steps.
This is not a service-provider-only problem. Enterprises running unified communications and hybrid work, public utilities carrying operational and increasingly latency-sensitive grid traffic, and federal and state agencies with mission and real-time requirements are all running into the same limits.
The limitation is in the traditional control plane. Conventional IGP forwarding chooses the lowest-cost path on a static metric, not the lowest-latency one. Traffic engineering was possible with RSVP-TE, but it carried a real operational cost. The protocol maintains per-LSP soft state that has to be refreshed continuously, which makes it chatty and hard to scale, and its fast-reroute implementation (the backup and bypass tunnels that protect each hop) is intricate to set up and keep working. For most network engineers the complexity outweighed the benefit, so RSVP-TE stayed confined to the paths that genuinely needed it. Even getting traffic onto those paths was a chore, leaning on static routes, autoroute announce, forwarding adjacencies or policy-based routing to push flows into the tunnel.
Segment Routing removes that overhead. Because the path travels as a segment list in the packet, there is no signaled state in the core and nothing to refresh. Source routing does what RSVP-TE needed a signaling protocol for. You keep the capability without the complexity: delay-based routing through Flex-Algo and Performance Measurement, explicit engineered paths through SR-TE policies and native steering that maps a flow onto the right policy by color or intent instead of static routes or PBR. One point to be clear about: this is path selection, not bandwidth reservation. SR does not reproduce the per-class bandwidth guarantees of RSVP-TE's DiffServ-aware TE, and QoS still handles per-hop scheduling. What it changes is how easily we can match a class of traffic to a path that suits it, on one converged network instead of parallel overlays or hand-built tunnels.
2. Fast reroute with TI-LFA: Sub-50ms as table stakes
The first and most universally applicable reason to move to Segment Routing is Topology-Independent Loop-Free Alternate (TI-LFA) fast reroute. It is also, in practice, the single biggest driver of adoption: more operators move to Segment Routing for fast reroute than for traffic engineering, even though traffic engineering tends to get more attention. TI-LFA provides guaranteed sub-50-millisecond protection with 100% topology coverage, precomputing a backup path for every destination prefix and steering traffic onto that path the instant a link or node fails, before the IGP has reconverged. Critically, it does this along the post-convergence path, so traffic lands on the route it would have taken anyway, avoiding the transient micro-loops and suboptimal detours that plagued earlier schemes. There is almost nothing for the operator to do, either. Unlike RSVP-TE, there are no backup or bypass tunnels to define and no per-path configuration. TI-LFA is enabled with a line or two under the IGP, the exact syntax depending on the platform or vendor, and from there the router computes and installs the backup paths in the forwarding table. The primary SPF takes priority; the repair paths are calculated afterward, on spare cycles, yielding to more urgent tasks such as IGP or BGP convergence. On modern routers (where forwarding and much of the feature processing sit on the line cards), the route-processor CPU is idle most of the time, so those backups are calculated without competing with anything critical. Turn it on once and the whole topology is protected.
The difference is easiest to see by contrasting the two operating states. Without fast reroute, a link or node failure triggers a full IGP reconvergence: every router has to detect the failure, flood updated link-state information, recompute its shortest paths and reprogram forwarding. On modern routers that reconvergence is quicker than it once was: link-state updates propagate fast and SPF runs in milliseconds, so many networks reconverge in well under a second, and only a large, sprawling topology still takes a second or two. But fast is relative. Even half a second is a lifetime for real-time traffic, and for the whole interval, however brief, traffic to the affected destinations is simply dropped. For bulk transfers that is a recoverable nuisance, but for a voice call it is a dropped call and for a video session a frozen stream. With TI-LFA, the backup path is precomputed and installed in advance, so failover happens locally in under 50 milliseconds, before reconvergence has even begun, and the disruption is imperceptible to the application.
That difference is why fast reroute has become a baseline requirement rather than a premium feature reserved for a handful of critical circuits. Voice and video have always demanded fast failover, and real-time and interactive applications are no more forgiving. TI-LFA delivers that protection across the entire topology, not just on the paths an operator chooses to engineer.
AI, for the most part, sits on the tolerant side of this line. A single inference request runs over TCP and recovers from a brief drop, behaving more like a bulk transfer than a voice call. But that is starting to change. Voice-driven assistants are real-time speech, and a lost packet garbles the audio exactly as it would on a phone call. Agentic workflows chain hundreds of requests to complete a single task, and the path between the agent and the model becomes critical to the whole task. In both cases, fast reroute keeps a brief network failure from disrupting the application.
3. Traffic engineering for everyone: Delay-based routing
For years, traffic engineering was a niche discipline. RSVP-TE gave you explicit path control, but its per-tunnel state, signaling overhead and operational complexity kept it confined to large service-provider cores and a few specialized installations. Most enterprises never touched it. Segment Routing changes that. It makes traffic engineering stateless: the path is an ordered list of segments in the packet header, so the core holds no per-flow state and the barrier to entry disappears. Traffic engineering becomes something any organization can operate, not a capability reserved for the few.
What makes this relevant to everyone is delay-based routing, and it comes from pairing two SR features.
Flexible algorithm (Flex-Algo)
Flex-Algo lets us define custom forwarding topologies optimized for a chosen objective (minimum IGP metric, minimum TE metric or minimum measured delay), with optional constraints such as excluding specific links or Shared Risk Link Groups (SRLGs). A latency-optimized topology is expressed as a single algorithm and reached with a single prefix-SID, so steering traffic onto the lowest-delay path means forwarding to the right SID rather than building and maintaining explicit tunnels. It is traffic engineering without the operational burden.
Performance measurement (PM)
Flex-Algo is only as good as the metric it optimizes, and that is where SR Performance Measurement comes in. Using standards-based probes (STAMP/TWAMP), PM continuously measures real one-way and two-way link delay and loss and advertises a dynamic delay metric into the IGP. Flex-Algo then computes its paths against that live measurement instead of a static, hand-configured cost. The result is a network that routes on the lowest actual latency to a destination, not on bandwidth or an administrator's best guess at a metric.
This is the heart of the modernization argument. When every link is fast, the question is no longer which path has the most bandwidth. It is which path has the lowest, most predictable latency right now. For voice, video, real-time data and interactive AI, that distinction is what counts, and Segment Routing lets an organization act on it without a dedicated traffic-engineering practice, with Flex-Algo and Performance Measurement doing the work.
4. Delivering services over the transport: L3VPN and L2VPN
A modern transport network is more than a set of fast, well-engineered paths. It is a platform for delivering services. Segment Routing natively supports both Layer 3 VPNs and Layer 2 VPNs, including EVPN, over the same converged fabric. Instead of relying on raw high-speed links or standing up parallel overlays and point-to-point circuits, we can carve the transport into segmented, multi-tenant services (VRFs for routed isolation, E-LINE and E-LAN for Layer 2 connectivity), all riding one SR underlay.
The value is universal, though it shows up differently by sector. Service providers deliver customer VPN services and monetize the transport directly. Enterprises segment departments, sites and security zones without building separate networks. Utilities and government agencies keep mission, operational-technology and corporate traffic apart with hard boundaries on shared infrastructure. The point is the same in every case: the network delivers services, not just connectivity. Consolidating those services onto one SR transport cuts the protocol sprawl and operational cost of running several purpose-built networks side by side.
5. Network slicing and deterministic services
Building on Flex-Algo and VPN services, Segment Routing enables network slicing: partitioning one physical transport into multiple logical networks, each with its own topology, path selection and performance profile. One slice, optimized for latency, carries voice, video and other real-time traffic. A second carries bulk or best-effort data. A third isolates critical operational traffic such as utility teleprotection and NVIDIA SCADA™ or government mission systems. Flex-Algo gives each slice its own topology, so one slice's traffic can take entirely different links from another's. Where slices share links (and often they will), slicing by itself does not keep them apart. That is QoS's job: it schedules each class so one cannot starve another, and hard partitioning such as FlexE (Flexible Ethernet) can go further and reserve dedicated bandwidth on the link. Slicing defines the classes and their paths; QoS or hard partitioning enforces the separation where those paths overlap.
One point on the utility case, because it is easy to overclaim: slicing gives teleprotection its own low-latency lane, but the failover inside that lane is not TI-LFA. Line current differential and distance protection need symmetric, co-routed paths and deterministic latency, which TI-LFA's post-convergence repair does not guarantee. Those flows rely on dedicated mechanisms instead: purpose-built teleprotection equipment, co-routed bidirectional LSPs that keep both directions on the same path or, within SR, Circuit-Style SR policies that hold a co-routed path with 1:1 backup. Slicing isolates and protects the class; the protection inside it is matched to teleprotection's budget, not the general-purpose fast reroute the rest of the network uses.
Slicing is what turns one big fast network into a set of differentiated, SLA-backed services on shared infrastructure. It is also the answer to the mix of workloads modernization brings. As real-time, bulk and increasingly AI-related traffic share the same backbone, slicing and QoS together keep them from contending without the cost of physically separate networks.
6. Operational simplification: Collapsing the control plane
One of Segment Routing's most underrated benefits is what it removes. Because the path is encoded as a segment list in the packet, the core no longer distributes and maintains per-path state. That eliminates LDP and RSVP-TE outright and collapses the control plane to an IGP (IS-IS or OSPF) plus BGP. Fewer protocols means fewer failure scenarios, less state to troubleshoot and a network that is simpler to operate and scale. When we need centralized path computation, an SR-PCE or controller can compute and program paths on demand, but the day-to-day control plane stays lean.
This simplification is standards-based and multi-vendor, which matters for procurement and long-term flexibility. It also compounds with everything above: the same architecture that gives us TI-LFA, delay-based routing, VPN services and slicing is reducing the number of moving parts. Modernization here is not only adding capability; it is retiring complexity.
7. Choosing your path: SR-MPLS or SRv6
Segment Routing comes in two data-plane forms, and every benefit described so far applies to both. The right choice depends on where an organization is starting from.
For organizations already running MPLS, SR-MPLS is the natural, low-risk step. It reuses the existing MPLS data plane and hardware while shedding LDP and RSVP-TE, and it delivers TI-LFA and SR traffic engineering with minimal disruption. For greenfield builds, IPv6-native environments or operators optimizing for long-term scale and simplicity, SRv6 is the stronger fit.
SRv6, specifically the uSID flavor, encodes segments as IPv6 addresses. That brings address summarization, a simpler transport model and a path toward richer service programming. The trade-offs are the IPv6 learning curve for organizations that have not adopted it yet, and a feature set that is still catching up to SR-MPLS in a few areas.
We cover this decision in depth in a companion WWT article, "Cisco SR-MPLS or Cisco SRv6 uSID: Choosing the Right Path for Your Network"[3]. It compares the two across architecture, scale, feature gaps, operational complexity and business drivers. For WAN modernization, the essential point is that both paths lead to the same place: a simpler, latency-aware, service-capable transport. The choice is about starting position and strategy, not about whether to modernize.
8. The realistic AI angle: Readiness, not requirement
AI is the reason many organizations are revisiting their transport strategy, so it deserves an honest treatment rather than a marketing one. The clearest statement is the accurate one: Segment Routing is not required for AI, and the WAN does not carry AI training traffic. Training a large model splits the work across many GPUs, and they have to exchange data constantly to stay in sync. Some of that traffic keeps every GPU's copy of the model aligned after each training step (all-reduce); some moves data between GPUs that each hold part of a model too large to fit on one (tensor and pipeline parallelism). Either way it is heavy, continuous and very sensitive to delay, so it lives in the back-end GPU fabric. That fabric runs on InfiniBand or lossless RoCEv2 Ethernet, inside a single data hall, at single-digit-microsecond latencies.
Physics sets the limit. At roughly five microseconds per kilometer in conventional single-mode fiber, a synchronous training collective that waits on its slowest member cannot span WAN distances. That figure is not fixed, and it is worth a note for the optical side of the house as RON and next-generation optics enter the picture: emerging hollow-core fiber guides light through air rather than solid silica, cutting propagation delay to about 3.3 microseconds per kilometer, close to the speed of light in vacuum, and roughly thirty percent faster than conventional fiber. That is a real gain for latency-sensitive long-haul and DCI routes, but it does not change the conclusion. Shaving thirty percent off a millisecond-scale WAN path still leaves it orders of magnitude above a single-digit-microsecond fabric budget. No amount of traffic engineering, or faster fiber, makes a synchronous training collective span the WAN.
Segment Routing has, in fact, earned a place inside that back-end fabric, and that is worth calling out as a mark of how far SRv6 has come. A recent transport called MRC (Multipath Reliable Connection), developed by OpenAI together with Microsoft, AMD, Broadcom, NVIDIA and Intel and published through the Open Compute Project (OCP), uses SRv6 source routing to spread a single connection's packets across many fabric paths at once, and it runs in production on some of the largest AI training clusters built to date. This shows real support of the SRv6 data plane: it holds up at the most demanding scale there is. It is a different job than WAN traffic engineering, though: MRC runs inside a single site with dynamic routing turned off, so it is not evidence that training traffic rides a Segment Routing WAN.
Where AI does touch the WAN is narrower, and more realistic:
- Inference distribution: request/response inference is latency-tolerant relative to training and can be sensibly geo-distributed across regions and sites.
- Dataset and storage movement, checkpoint replication, disaster recovery and cloud bursting: bulk and asynchronous flows between data centers.
Data-center interconnect deserves a plain word too. Most DCI today is built on Routed Optical Networking (RON): coherent pluggable optics such as 400G ZR and ZR+ placed directly in the router, running over dark fiber, usually over short or metro distances. Even where organizations build dedicated longer-haul interconnects, RON over leased or owned fiber is the common approach. Segment Routing does not replace this. It is the natural IP-layer forwarding and traffic-engineering layer to run across a RON-based transport. Carrying DCI over a modernized SR backbone, rather than as a set of isolated point-to-point circuits, is itself a reason to modernize: it brings the same TI-LFA protection, delay-aware routing, slicing and service delivery to the interconnect that the rest of the WAN already has.
How much does the WAN actually matter to inference?
It helps to put numbers to this instead of asserting it. Cisco's white paper A Day in the Life of a Prompt[4] breaks time-to-first-token into four parts: queueing, tokenization, prefill compute and network. In its worked example, the network contributes about 40 milliseconds of a roughly 350-millisecond total; batching, queueing and GPU prefill account for the rest. Since a full response often runs into seconds, the WAN is on the order of ten percent or less of what the user actually experiences. At that scale, whether network latency is 20, 40 or 80 milliseconds barely moves the result.
Two trends could change that balance. Inter-token latency is falling fast as new inference hardware arrives (the same paper cites recent results approaching a fraction of a millisecond per token), and agentic workflows chain tens to hundreds of sequential model calls into a single user request. As the compute per call shrinks and the number of network round-trips grows, the network's share of perceived latency rises; Cisco projects it reaching a quarter to nearly half within a few years. The point to hold onto is that network latency is not growing here. The compute it is measured against is shrinking. This is a projection, not a present-state measurement, but the direction is credible, and it is the argument for modernizing the WAN before the demand lands rather than after.
The practical result is distributed inference. As the network's share of latency grows, inference has to move closer to users, across edge, metro and core rather than concentrating in a few large data centers. This is the one AI-driven reason to modernize the WAN that holds up without overreach: not training, which stays in the fabric, but low-latency inference delivered near the user. And it maps cleanly onto the capabilities already covered: delay-based Flex-Algo and Performance Measurement to steer inference along the lowest measured-latency path, slicing to protect it from bulk traffic, TI-LFA to keep it available and RON-based DCI to reach the metro sites where it will increasingly live.
That is the readiness argument, and it is the honest one. AI changes the shape of traffic: more data pushed upstream, more bulk movement between sites, more sensitivity to latency and jitter on the flows that do cross the WAN. An organization that has modernized its transport with Segment Routing is ready for that shift. SR does not make AI possible, but it makes the network ready for the next step when an organization takes it. That readiness, not any claim that SR enables AI, is the right way to frame it.
9. The migration case: Why Segment Routing, either way
The decision to modernize does not depend on whether an organization already runs MPLS. If it does, Segment Routing is the natural next step. It preserves the existing data-plane investment, retires LDP and RSVP-TE, and unlocks TI-LFA, delay-based traffic engineering, VPN services, slicing and programmability through a straightforward migration. The technology an operator already knows becomes simpler and more capable at the same time.
If an organization does not run MPLS today, Segment Routing still makes sense, maybe more so, because it can be adopted without inheriting any legacy signaling. Fast reroute, latency-aware routing, converged L2 and L3 services, network slicing and a lean control plane stand on their own merits for the latency era, with or without AI in the picture. Enterprises, service providers, public utilities and government agencies share the same underlying need: a transport network whose value is measured in behavior and services, not just bandwidth.
Bandwidth is largely a solved problem. Latency, resilience, service delivery and readiness for what comes next are the mandate now, and Segment Routing delivers all four on a single, standards-based architecture. That is the case for WAN modernization with Segment Routing, and it applies to nearly every organization running a transport network today.
References
[1] J. Gustawsson, J. Antich, and W. Sagheer, "AI Impact on Wide Area Networks: Cisco 2026 Report," Cisco Systems, 2026 (research conducted with Opanga Networks). https://www.cisco.com/c/dam/en/us/solutions/collateral/artificial-intelligence/mass-scale-infrastructure/ai-network-traffic-report.pdf
[2] Broadband downstream-to-upstream traffic-ratio data: NCTA – The Internet & Television Association (U.S. fixed broadband ratio rising from ~3:1 in 2010 to ~14–16:1 by 2019–2020); Cartesian household-usage study (~17:1). [PLACEHOLDER: confirm exact publication title, date, and URL before publication.]
[3] M. DiVincenzo, "Cisco SR-MPLS or Cisco SRv6 uSID: Choosing the Right Path for Your Network," World Wide Technology, June 2026. https://www.wwt.com/blog/cisco-sr-mpls-or-cisco-srv6-usid-choosing-the-right-path-for-your-network
[4] J. Antich, "A Day in the Life of a Prompt," Cisco Systems, White Paper, July 2026.