As I continue to iterate on, and flesh out, the Internet Innovation Stack which I will use as a way of organizing my technologies of interest, networking is a natural area for me to dive in, as it is the area of technology I know the best.
A typical way of articulating the future of networking, or anything else, is to pick a place in time and describe the way the world will look at that time. As there are NO facts about the future, this is an inherently speculative activity, that often expresses linear extrapolation, even when we are are not consciously aware of it. You know, the whole black swans thing.
Based on nothing but experience, intuition, a sense of where standards are at, and guesstimates about how long it will take network managers to get comfortable with change, this is my current view of where networking is going over the next five years, and where we will be in 2025, from the perspective of new builds. So net-net, I don’t present this as something with any scientific methodology backing it. I will iterate my mental model over time as new information comes to light.
Some of this does not take a great deal of courage or insight, being settled matters. It is hard to imagine the industry launching into the development of a new link-state protocol, for example, or even a new interior gateway protocol (IGP), for that matter. So after decades of development, standards work, and implementation from multiple network managers and routing vendors, some things seem like settled matters (hoping not to tempt the networking gods here). The above table focuses on SP networking, mostly telco, with the obvious observation that hyperscalers have a significant influence on all networking directions, in today’s market. SD-WAN/Enterprise-WAN is an area that straddles Enterprise/SP networking, and that is as far into Enterprise networking as I go in this article.
The next five years
The overall context of IT evolution is important.
All of IT has been disrupted by hyperscalers and SaaS. Combined, they have both radically changed customer experience (compared to on-prem software models) and operations excellence (compared to traditional IT models). They have also changed business models. It is in the area of operations excellence that the future of networking must first be examined.
The hyperscaler model is in a crude sense, one where you simplify basic IT resources (compute, storage, networking) so they are repeatable for scale-out, automating with operations software, within an environment where operations is a software engineering-driven model. In reality, there is much more to DevOps & Site Reliability Engineering, but big picture…So we are talking about a model where network equipment is programmable for orchestration and automation. There is more going on than that, but the story of the next five years begins and ends with this critical point. It may still be nice to have a command-line interface (CLI) as a “security blanket” or for some lab work, but bottom-line, if you are still running your network in 2025 using a CLI, you are going to be way behind the curve in terms of operations excellence. Arguably, you already are. Netconf/Yang usage will become mainstream over the next five years.
Driven by the Site Reliability Engineering ethos to measure everything, router vendors have put a significant investment in high-performance, high-volume, telemetry. There may be more work to do here, but this is a major inflection in networking. If I think back to network operators writing their own SNMP collectors 20+ years ago, and what the industry has tried to achieve over the last 20 years, it is time for a change. SNMP is widely deployed and embedded in operations environments for statistics collection, alerts, and sometimes more, but the future is gRPC/gNMI or some yet to emerge approach. Telemetry is no longer just to support a 6-month capacity planning cycle, it is about having enough information to support much shorter capacity planning cycles, and real-time traffic engineering, orchestration, and automation decisions. Ths industry is stretching for a new operations reality, that requires new engineering in routers, and new ways of collecting/analyzing information.
EVPN has been a hot topic for a few years already, so it takes no great insight to suggest this is the future of layer 2 “services”. That EVPN is a BGP-based approach to layer 2 services has a number of interesting aspects to it. In addition, a form of L3VPN can also be offered by EVPN, sometimes called L3 EVPN. Moving to EVPN may make some of the forwarding/data plane migrations over the next few years easier. See: BGP for Service Migration.
Network operating systems (NOS) are rapidly moving to both Linux and container-based (internally). My intuition is this is the beginning of a new innovation cycle in NOS, and it is difficult to see all the implications of this today. Nokia emphasized the NOS as a development platform in its recent announcements and Cisco emphasized the NOS as something that could be packaged for individual customer requirements when it highlighted IOS XR7 late last year (it was shipping prior). I could speculate about the option space here, but it is ultimately up to vendors and network managers to drive innovation in this area.
I will likely dive into Segment Routing and SD-WAN in another article, because they are both “from the edge” technologies and the whole end-to-end thing is so much a part of the Internet culture ethos. Performance management and quality of service have been significant conversations in networking for decades, with many different approaches to both having been deployed. It is a big subject, but it does not take much insight to note that both Segment Routing and SD-WAN are on the ascendency today, and both will be part of a larger end-to-end conversation in the industry, that will include, for at least one of them, server-to-server / server-to-handest. Given there are already some SRv6 deployments, suggesting that SRv6 won’t be mainstream for new builds until the 3-5 year period probably deserves some explanation. I will discuss that in another article.
Integration is not something you can force, it happens when the technology is ready for it to happen. With 400ZR/400ZR+, it seems the industry, packet, and optical, feel this is the right time for it to happen in data center interconnect use cases (up to 80KM, maybe 120KM), in the form of coherent optical pluggables. High-performance, spectrally efficient, long-reach transmission appears today, to still be in the realm of specialized optical systems. This will be interesting to watch over the next decade, especially given Cisco’s Acacia acquisition and the remaining sellers of optical pluggables. Also interesting to watch will be whether symmetrical 10G PON takes off us a pluggable, and what transitions may occur in access, in a post-pandemic world.
5G radio deployments have started. How ubiquitous the highest-bandwidth services will be, based on millimeter tech, is an unknown at this stage. Presumably like all tech it is just a matter of time, but there are naturally some significant changes to drive to go from a macro-cell topology to a small-cell topology, everywhere. Also, the journey from 4G service cores (mobile “packet” core) to 5G services cores appears to be in the future, and that transition may gate some of the overall promises of 5G. Nonetheless, 5G is an opportunity to rethink total network architecture from segment routing to virtualization, orchestration, automation, and unified access, aggregation, and core. Definitely watch this space, should be interesting.
By definition, you can’t see Black Swans coming. So this next section is highly speculative, by nature.
Quantum networking would be about the most disruptive thing I could think about over the next five years. It is also the most unlikely. To say quantum anything is a science project at this stage, is to err on the side of generosity. In addition, the quantum computing community has come to the conclusion that it will be a co-processor to classical computing and not a replacement. The quantum networking world appears to have come to a similar conclusion. That said, I do get a real sense that the innovation cycle in quantum is accelerating, error rates are reducing, and the next 5-10 years is going to be an exciting period. Will quantum be another example of something we all get excited about but never makes a significant dent in our day to day lives? There obviously is that risk. Nonetheless, the potential is cool, and I definitely intend to keep an ongoing eye on it.
5G is not ubiquitously deployed yet, and some are already talking about 6G. Given the existing challenge with 5G investment, it seems highly unlikely that 6G would be a disruptor in the next 5 years. The one big outstanding issue in this area that gets my attention is whether WiFi should be the basis of telecom mobile networks going forward. There were some that believed that WiFi would dominate the 5G era, which as far as I know, is not a matter or radio technology per se, but other areas of standardization. Will there be another stronger push by the WiFi ecosystem during 6G? In my mind, that could be a real game-changer if it ever happened. We shall see. It will also be interesting to see how fast Open RAN approaches are adopted over the next five years, which will provide some insight into the overall trajectory of the mobile infrastructure.
Serverless is bleeding-edge technology, from the perspective of mainstream adoption. Like containers before it, the requisite management infrastructure is not where it needs to be. Putting aside the discussion around serverless being stateless (pros and cons), there are some exciting possibilities with a function-as-a-service model. As we are just at the beginning of our NOS journey to containers, I don’t see serverless becoming mainstream as a NOS paradigm in the next five years. However, there is one other place where serverless could make an appearance in networking.
Cloud and SaaS have disrupted everything. There are significant risks and opportunities in not applying/applying this to networking. We have already seen significant cloud-based services in the area of security. Security is a natural fit for a cloud subscription model because, by nature, there are ongoing updates. Parallels in networking are not as obvious to most. There is of course some cloud-based solutions in the area of configuration. There is cloud-based management in some segments of networking, for example, Enterprise branch wireless/security. Will something more disruptive happen in networking? It is not something I hear much discussion about. The potential is there but will require imagination, vision, investment, and commitment. That combination is hard to forecast. Either it happens or it does not.
BGP as IGP? I will have more to say about BGP in another article. For now, keep an eye on link-state vector routing, and let’s see if it goes anywhere beyond its current use case.
RIFT: Do we make network equipment simple, or do we make networks simple. RIFT says the later. RIFT also asserts that getting to really good zero-touch provisioning requires changing routing protocols as we currently know them, even if they use the same name and are based on the same technology. RIFT is currently focused on data center scale-out networking architectures with regular topologies and does not obviously look like a fit for WAN networks today. RIFT is asking some interesting questions and providing some interesting capabilities. Definitely watch this space.
Virtualization. What can be said here that has not already been said, I for sure have had plenty to say. As the networking industry continues to explore the best way to leverage virtualization, the most interesting questions for me are the following. Do we want to strip network equipment of a distributed control plane? Segment routing says no, and I would probably say no as well. There are advantages to the distributed exchange of information. Do we want to have the full-stack of VPN-related functionality, for every VPN, embedded in routers? That question is less clear to me, and there are currently startups exploring this question, and it will be interesting to see the results of that exploration over the next five years. The answer is likely not binary and probably depends on the quality of competing approaches.
Controllers. Centralized vs distributed. I have a propensity to feel the answer is to let distributed do what it does well and augment it with scale-out, data-based, global optimizations, and insights. As the industry invests in segment routing flex-algo, service-aware segment identifiers that perform network resource reservations and VPN constructs, the distributed vs centralized questions in networking continue to be an area where there remain different perspectives and competing approaches. Controllers have the potential to herald a dramatic change in how we manage and/or optimize networks. What happens over the next five years will be fascinating to watch. I personally believe the industry should take the view that controllers augment a distributed control plane, but we will have to watch and see what happens.
While there are implementations of “unequal” cost multipath (a variant of ECMP – equal-cost multipath), developments like RIFT suggest there may be more focus on non-equal cost multipath to come. ECMP is an approach to network management, specifically traffic load balancing, that is often positioned as an alternative to fine-grained traffic engineering, though it has historically had some deficiencies in networks with asymmetrical bandwidth on different paths. To the extent that more research and investment is put into this area of networking, it may expand the number of network managers that forego traffic engineering (TE). The tradeoffs require a separate article and just suggesting ECMP and TE are comparable in any way is risking significant negative feedback, because they don’t focus on the same problems, per se, or have the same capabilities. Today, I believe there will continue to be significant networks with an interest in TE.
AI/ML. What is AI and what is ML is a discussion on its own. The aspiration has over the last few years been articulated by enough people for it to be on everyone’s radar. Intent-based/policy-based self-driving networks. Networks that are truly autonomous in a way that preserves business rules, priorities, and interests. Easy to say. Hard to realize. If the industry can pull this off, that would be a radical transformation, bringing networking fully into the cloud-era. My guess is, by 2025 the industry will have some capabilities in networks worthy of being designated “self-driving”. I doubt the industry will be fully there. However, of all the things in this section of the article that has the potential to surprise, this would be high on the list, if not the highest.
In broad strokes, the direction of the industry is known. SaaS-like customer experiences. Hyperscaler-like operations excellence. Programmable, automated, and perhaps even autonomous. Increasing packet/optical integration. A shift from IP/MPLS to segment routing is likely, and with it, whatever leverage comes from a source-routed paradigm. Telemetry feeding enough information for centralized controllers to augment distributed control planes. Continued leverage of BGP, and perhaps even a compelling role for virtualization. SD-WAN reminds us there is something interesting going on in the overlay that is different from what is going on in the underlay, with a different, multicloud, multiaccess, cloud-managed, mission. There are no facts about the future. The precise timelines and exact technologies are not known. Some directional statements can be made.