Is networking the next bastion of AI hype, or is there actually a lot of value we could harvest there? That’s an important question, one this Light Reading article shows is already being asked. It’s true, as the article suggests, that mobile operators are hopeful, and also true that so are operators overall and even enterprises. If all this hope converges on some common AI missions, that convergence might be an AI proof point. But is is converging, and is there any other validation out there? Do operators, and network vendors, have an opportunity related to AI? It’s complicated.
The article offers three areas AI might target, “performance and efficiency and drive profitability”. The first two seem to me to be dependent on an AI mission in capacity planning and/or traffic management, and the last would have to target revenue, capex, or opex. I would argue that capex savings would have to derive from improvements in performance or efficiency, so that leaves revenue augmentation or opex reduction.
I think AI could likely be a benefit in capacity planning, and since applying it to that mission wouldn’t likely incur a major cost, I think it would be worthwhile step for AI and operators, and even as a tool to help enterprises size data center trunks and branch access connections. However, enterprises agree with my view that the value to them would be limited unless major changes in network traffic were contemplated. That matches operator views that this particular idea is most useful in greenfield situations. One that many mentioned was FWA, where node placement optimization is considered a significant issue.
Traffic management via AI faces two primary questions. First, how much could be accomplished without adopting a central control paradigm like that of SDN? You can’t do local traffic management to achieve any benefits beyond dealing with a local failure or congestion. MPLS offers traffic engineering benefits, but adaptive routing would seem to defeat any AI attempts at network-wide optimization through MPLS, beyond initial route planning. Second, is optical capacity so cheap that you couldn’t justify spending on optimizing how you use it? And the most capex-intensive part of network infrastructure is the access part, where there are rarely many alternate paths for traffic management to choose from. Operators say this AI mission is likely effective only in greenfield applications, too.
Are we out of options here? I don’t think it’s that dire, but I do think that we may be proving that AI isn’t a panacea for operator profit problems or enterprise network challenges. There are two broad AI opportunities related to networking, and one really related to the business of networking. Of course, they all have potential issues.
Security is perhaps the biggest network-specific opportunity for AI, one that exists for service providers and enterprises alike. A dozen enterprises and twice as many operators say that they’ve determined that almost all the significant security issues they’ve ever encountered or even heard of leave a traffic footprint, and a majority could be at least addressed and minimized by network reaction. Operators believe that network security can be sold, and enterprises agree.
The problem is collecting the knowledge. It’s not practical, and may not be possible or even legal, to inspect packets in detail to identify problems, but looking at the patterns in traffic and in interactions is another matter, one where AI could clearly help. One enterprise told me they’d minimized a ransomware attack by detecting a pattern of spread, which generated an unusual relationship between user activity and database service activity. In this case, finding the pattern was a happy accident, but the company is now exploring AI options to solidify the strategy.
You can improve AI’s potential in security if you have the ability to recognize “sessions” or relationships among users and applications. You can improve it further if you use policy to manage what sessions are allowed. Most malware manifests its presence by attempting connection, many of which are prohibited, and just knowing this is happening and detecting a spread is a solid signal of a problem. It also identifies the specific risk actors, allowing them to be cut off.
AI could be especially valuable as an alternative to using explicit session policy violations to detect suspicious behavior. The session patterns of users or applications could be enough in themselves to identify a problem in the making. However, this mission may pose challenges for operators, because enterprises are wary of having anyone inspect their data to the point needed to achieve session awareness, so traffic volume alone might be the only path.
Problem detection and analysis is the second network-specific AI mission with good potential. Operators who see a managed service opportunity view AI as a way to offer a valuable service without incurring a lot of human cost, improving both sales and margins. Enterprises list this as their own primary area of AI interest in networking, both for current and proposed/planned AI use. It does require a lot of visibility into the network to be effective, which so far has tended to make it most successful in enterprise-deployed scenarios.
Managed services are nothing new, and the obvious truth is that were they a transformational revenue opportunity, operators would have already realized it. The question is whether AI could change that picture by lowering the cost so significantly that managed services could look compelling to buyers and still be profitable to sellers. I think it could, but if that were the case, couldn’t enterprises buy the AI tools themselves? Enterprises say that the visibility risks would lead them, by more than a five-to-one margin, to pick their own AI over a managed service. Might AI be useful to SMBs, though? Yes, but operators are hesitant to target the space at this point, citing the price sensitivity known to exist there.
What about that traditional-business opportunity? Well, operators and enterprises are all businesses, all with customers to sell to, accounting, employees, and so forth. All of these activities could be improved through the use of AI, but particularly the area of prospect/customer relations. However, these applications are more difficult to frame as services, and because they aren’t networking-specific they’re beyond the scope of this blog.
How can we sum up the potential of AI in networking? There are compelling value propositions for enterprises, and Juniper has already demonstrated that. HPE might push it even further. There are some interesting opportunities for operators, but probably none that would be transformational, and many of the proposed missions there have minimal credibility. About a third of operators told me they thought AI could help lower opex, but none thought it was the long-term answer to their profit-growth prayers. I don’t think so either.
