What’s behind the push for AI-RAN? Not a single enterprise has, over the last year, even mentioned any interest in or benefits of it to them. Of the 88 operators who I’ve had contact with in that period, only 22 said they believed it might, stress the qualifier, have value, and 49 said they saw no real benefit. So why do we keep hearing about it? Perhaps this Light Reading piece has an answer to that.
There are a lot of theories about AI RAN. Some focus on the benefits it could bring in terms of managing cells and spectrum, as a replacement for ASICs. Some focus on the potential for edge-hosting AI services. Some on both. Operations and security value are both often cited. There is theoretical benefit for all of this, but provable value? Not according to virtually every enterprise (as buyers) and the majority of telcos (sellers).
The problem with the use of GPUs as an ASIC replacement is that you’re looking at a real cost penalty with a speculative benefit. Yes, there are indications that the full exploitation of MIMO could be more likely with GPUs than ASICs, spectral efficiency might be higher, and there might be operational benefits. However, most operators think these benefits are not compelling to the point where they could justify deployment of new RAN equipment, and they’re doubtful whether the benefits could be realized if they were phased in with orderly modernization initiatives. Pockets of AI, in short, don’t seem to operators to offer a value.
You could address the problem of pockets of AI by linking AI RAN to a massive new infrastructure wave, which brings us to the next-generation wireless point. Many say that AI RAN is essential for 6G, but operators note that their own priority for 6G is not to have it be a big fork-lift. Half of operators say that were that to be a requirement, they’d seriously look at not advancing to 6G at all, and almost half think that their pressure on the 3GPP would induce the body to reject any such requirement.
Leveraging something that doesn’t deploy in the first place is surely a challenge, but it may be even tougher than that. The pressure on operator applications for AI RAN inevitably create interest in lowering the cost of achieving some of AI RAN’s promised benefits. About a third of operators believe that promoting GPUs as an ASIC replacement begs the question of ASIC improvement. Why not simply do a better ASIC? The majority believe that even if it were determined that AI could be a benefit, it would not be the generalized AI we see today, driving those massive data centers, but rather a form of machine learning, or perhaps a simple GPU and a foundation model. This “small-model” approach isn’t spontaneously validated, but it is expressed as a counterpoint to the perception that AI RAN means generic-like AI.
This directly impacts the notion of using the AI RAN resources to offer edge computing. The more specialized the RAN hosting mission is, the less likely it is that it could support edge services, since any limitations in what the edge host could offer could stick enterprises who used the service with limited application migration support. What the AI RAN edge might work with today’s application, but a new requirement? That would be a tough sell.
Tougher given that enterprises are more likely to see a risk in using shared AI RAN hosting resources for edge computing than a benefit. A bit over a third of enterprises say, spontaneously, that they’d be concerned about the security of a mobile network whose hosting was shared with edge service users. Some pointed to the issue of GPU hacking, just recently noted. Most just think that if security gains are a justification for AI RAN, then sharing resources with edge services are a more-than-compensating risk. They have challenges securing their own resource pools, after all. Why would operators not have them too?
Interestingly, though, operator comments suggest that a lot of (if not all of) these concerns could be overlooked if their current primary RAN vendor were to adopt AI RAN. Remember that operators have moved to the position that the biggest problem with “openness” is that it presumes a willingness to exercise a broader range of vendor choices, when this to operators just means more integration worries and costs and more finger-pointing if a problem occurs.
So, who wants AI RAN? Two groups, say operators. The first is the “outliers” in the mobile infrastructure space, who want the incumbents to share the wealth. Their hope would be that the innovation that AI RAN might (again, note the qualifier) create could (same) produce something so beneficial it would promote replacement of infrastructure at a faster rate. Second, the AI players, notably Nvidia.
Operators do not want to do a lot of integration. That means that multi-vendor RAN is a heavy lift in itself, and also that pockets of new RAN technology would be avoided even if the theoretical benefits of AI RAN could be achieved in a pocket-deployment environment. In any event, all the giants in mobile infrastructure have learned to embrace open initiatives with the realization that as long as they do that, it’s likely they won’t really be admitting others into their tents.
For the AI players, AI RAN is almost essential, for three reasons. First, they have to keep shaking the earth to sustain the hype wave. A ton of capital has been sunk into AI data centers, which for the most part are running AI that nobody is paying for. Applications needed; line up here. Second, just a little extension in hype-wave life might be enough for something real to come along and sustain AI spending. Third, even just attempts to realize opportunities that don’t actually pan out, or even exist, might create a useful application or technique that could build that “something real”.
I think it’s this group of AI types, more than any collection of buyers or of mobile vendors, that are really pushing the notion of AI RAN. That doesn’t mean that the move is totally self-serving and cynical, but I think that both are there in ample measure. Many, including me, have predicted that the AI wave would, as hype waves all do, crest and crash eventually. Will AI RAN crash with it? Unless it deals with real value propositions, I suspect it will.
