- Telcos can’t beat hyperscalers on AI compute
- The edge AI “opportunity” lacks a killer use case to justify investment
- Network modernization and AI optimization can help telcos turn their networks into critical utility infrastructure hyperscalers need
Telecom operators won’t want to hear this, but the winning AI strategy is simple: Go for the utility play.
Utility is a dirty word in the telco industry. Operators fought hard against the label in the battle over net neutrality. But embedding themselves as AI utility infrastructure seems to be their best bet in a market that is dominated by hyperscalers and, to date, lacks a killer use case that would require their unique touchpoints.
Rather than chasing AI baubles, AvidThink Founder and Principal Analyst Roy Chua told Fierce, telcos should invest in modernizing their networks, making them more programmable and consumable, and reclaiming the ground they’ve lost in the long-haul, sub-sea and metro connectivity realm.
Losing battles
It will come as no surprise to anyone paying attention that telcos have already lost the macro compute opportunity. In reality, they never really stood a chance.
Hyperscalers have spent the past decade-plus building global cloud compute footprints and perfecting the art of operating this infrastructure at scale, meaning they already had a leg up when ChatGPT kicked the compute race into high gear.
Last year, Amazon, Alphabet, Meta and Microsoft spent a collective $416 billion on footprint expansion. This year, that number is expected to jump to well north of $600 billion – even more if you include Oracle – meaning they will have collectively spent over $1 trillion on AI compute infrastructure in two years. That’s not a number telcos can match, and even if they could, they just don’t have the expertise to operate data centers at scale.
There’s been some talk about operators seizing on the sovereign cloud opportunity, given they already operate trusted infrastructure in countries across the world. But hyperscalers have moved quickly to snatch this opportunity up as well.
It’s true, Deutsche Telekom has had some success in this vein in Germany, but as Fierce has noted before, its model isn’t one that is easy to replicate – especially for U.S. operators and standalone telcos which lack business units dedicated to this kind of thing. Because again, hyperscalers have already flooded much of the market and managing this kind of infrastructure at scale is difficult.
With recent geopolitical events and attacks on AI infrastructure, it’s also no surprise that folks are suddenly rerunning the cost-benefit analysis on sovereign cloud and AI deployments.
Justifying edge investments
The edge has also emerged as a contender for operators’ attention and investment, thanks in part to Nvidia’s AI-RAN push and efforts like Cisco’s AI Grid. But Chua pointed out there’s one huge hole in this plan: There’s no real, game-changing use case to justify the investment.
At this stage, Chua said, the edge AI pitch is just “MEC all over again.”
For those who don’t remember, mobile edge computing (MEC) was popular at the turn of the decade and focused on the idea that deploying real-time processing power at the network edge could enable new, monetizable use cases.
Verizon leaned into MEC, investing big and partnering with AWS to deploy its Wavelength and Outpost infrastructure. But revenue failed to materialize and while MEC is still mentioned in tucked-away corners of Verizon’s website, the last mention of it on an earnings call seems to have been in Q2 2024.
Here’s why that’s strange: if AI at the edge is a killer use case for telcos, Verizon be shouting about its capabilities from the rooftops. It would, after all, have a multi-year lead over competitors. But it isn’t shouting about it, and it doesn't have a lead over competitors as far as we know.
Mobile edge computing was a flop
MEC was a flop because there weren’t yet any use cases that needed it. As Chua explained, applications that require real-time performance, like self-driving cars, are run on-device. Those that are ultra-business critical run on-prem. And everything else is pretty well-enough served by local cloud zones.
Indeed, VaporIO’s Kinetic Grid and the Open Grid Alliance ran into the same issue.
With AI-RAN and AI Grid, Chua said the industry is hearing the same pitch over again, but the use cases haven’t changed. Folks today, he added, are still talking about computer vision – a use case that has been in demo booths for at least a decade. And a few extra AI bells and whistles do not make an investment case.
Without new use cases – truly disruptive ones – the edge is a moot point.
But that’s not to say telcos shouldn’t invest in their networks for the AI era. They should. The focus, however, should be on modernizing them and making them programmable and consumable.
Network as a (utility) service
Networks are and always have been telcos’ biggest asset and their area of expertise. Owning that in the AI era isn’t a bad thing. On the contrary, embracing their role as critical, utility-style infrastructure is a winning strategy. Call it Network-as-a-Utility-Service, if you will.
“They could come in and show that they do have a unique advantage in running large-scale networks – between buildings, in metro regions and over the wider area – that they should be the partners for that,” Chua said.
Hyperscalers have eroded some of telcos’ dominance in the network space – with their subsea cables and metro fiber infrastructure – but it’s worth trying to take back ownership of the network. The way to do that is to modernize and optimize network infrastructure using AI to give hyperscalers and other enterprise clients the kind of flexible, programmable fabric they’ll need long-term.
In a nutshell, telcos need to architect and modernize their networks such that they can give customers “a path across the internet anywhere…to move data around and run my models anywhere GPUs are available," Chua said.
Lumen Technologies and Zayo in the U.S. and Colt Technology Services in the U.K. are examples of how this can be done successfully, he noted.
“Telcos providing flexible, programmable networks that [they] can provision and command — I think that’s viable because enterprises and all these cloud providers will actually want to be able to program routes with certain amounts of capacity, with certain SLAs,” he said. “That’s really a more programmatic utility ... basically a better class utility.”
Verizon, AT&T or Telefonica could do the same to monetize their backbone infrastructure. The catch is they need to invest in both talent and network modernization to make it a reality.
“They could [do this]. They just need to get their acts together in terms of making their networks programmable,” Chua concluded. “But they haven’t.”