I’ve spent the past few months asking executives and analysts across the telecom and tech sector the same question: is the edge opportunity for telcos actually real? I waffled back forth at first, but the more I’ve asked the question, the more skeptical I’ve become. There isn't an easy answer.
The impetus for this quest was the release of Nvidia’s AI Grid and subsequent implementations involving Cisco, AT&T and Comcast in April. For a moment, it seemed like just the breakthrough operators needed to bring the edge to life. AI, I thought, could be the use case operators have been searching for to make the edge make sense. After all, analysts have been saying for years now that inferencing won’t require the centralized compute resources that training does.
But, lo and behold, it’s more complicated than that. And the truth lies somewhere in between claims made by the edge believers and the critics.
The position of the critic camp is best embodied by comments AvidThink founder and principal Roy Chua made to me following the AI Grid news in April, when he pointed out that the whole “edge AI” pitch is reminiscent of the failed mobile edge computing initiatives operators chased the late 2010s and early 2020s. The problem – both then and now – is that there isn’t a solid use case to justify edge investments.
Sure, there’s AI. But mission critical AI, for instance for physical AI applications like cars, drones and humanoid robots, will likely be deployed on-device to avoid catastrophic failures. Everything else has enough of a latency budget that it can be processed in a data center.
As Red Hat VP and telco chief Fran Heeran, who led core networks are Nokia for many years, put it in a recent interview, “The notion of putting any sort of intelligence in a condition-critical situation away from the device itself, I think, is madness. Why would you put the driving logic for a car anywhere other than a car? Because you’re just inviting trouble.”
“I would challenge anybody to find the definitive use case just yet as to what needs to be that far out there,” Heeran added.
Suman Kanuganti, founder and CEO of Personal AI, was similarly unconvinced that running GPUs at the edge would end up being a major business for telcos. He suggested instead they should focus on strengths they can offer around identity, authorization, regulations and privacy.
The case for telco edge AI
Then there are believers like Ericsson Americas CTSO Joe Constantine, who argued that humanoid robots will need edge compute for inferencing. Indeed, Taara CEO Mahesh Krishnaswamy argued that the ability to put compute power on-device will be limited by space, batter power and cooling requirements.
Device pricing will also factor in – the cost of powerful compute on a more expensive device (like a car) is likely easier to hide than, say, trying to squeeze a $40,000 GPU into a $500 drone or a $20,000 humanoid. That means inferencing will have to be done at the edge.
The first green shoots of edge interest already seem to be appearing. AT&T’s GM of Strategic Managed Services Andy Foerstner this week told Fierce’s Mitch Wagner that customers are starting to request local instances on which to run their AI models. But even he conceded that the use cases still aren’t entirely clear and the edge isn’t something AT&T has “productized” yet.
The voice of reason
It’s easy to see why the telcos are so eager for the edge opportunity to work out this time. As I’ve pointed out before, telcos are sitting on prime edge infrastructure. And as Heeran noted, “it's potentially a piece of real estate and a function that the hyperscalers can't match.”
Perhaps the best take on the edge opportunity – and challenge – came from AWS Global Director for Telco Solutions Amir Rao. The problem, he said, isn’t just finding use cases that demand inferencing at the edge but finding ways to productize those.
“The use cases will come, but we have to have an offering, a product as an industry which makes sense for those use cases,” Rao said.
“If you are just going to put inference infrastructure at the edge and not offer an experience – which include the last mile connectivity to the device – then you are letting the developer of the application have to sort out the latency separately,” he explained. “That’s not ideal because if they have to do this then they may as well use the cheapest kind of resources available” which would be hyperscale compute.
In a nutshell, telcos not only have to identify edge AI use cases, but those where they can actually add value by offering things like managed SLAs, contextual awareness and so on.
Notably, while Roa thinks robotics and humanoids will be a huge use case, he specified that he doesn’t mean industrial robotics. That specific genre of robotics, he said, is better served by a fiber or private wireless connection.
Whatever use cases telcos decide to chase, Rao said the matching product offering has to be a must have to make that experience work.
“The offering would have to be built on specific use cases and it would actually only flourish, in my humble opinion, if the use case absolutely requires it or cannot come to life,” he concluded. “If it is a nice to have, then of course we’ll run into some challenges.”
That, though, leaves telcos with a chicken and egg sort of problem. They can’t really afford to wait to build out their edge infrastructure, lest they get beat out by startups with deep, private equity-filled pockets. But if the ideal use cases will require a telco’s product to come to life, then to some degree at least they’re stuck trying to dream up the use cases they’re hoping to enable.
Maybe they can use AI to dream up the killer edge application.
Opinion pieces from industry experts, analysts or our editorial staff do not represent the opinions of Fierce Network.