- Telcos are chasing the AI factory trend, especially European players
- But they may not be the best candidates to run this infrastructure
- Huge capital investments in AI infrastructure could backfire if telcos aren't careful
What started out as a trickle of AI factory announcements has risen into a full blown wave. But lurking behind the slew of projects is one key question: Do telcos really have any business running AI infrastructure or are they setting themselves up to crash and burn?
Europe is the current epicenter of AI factory activity, though there have also been rumblings in the Middle East and the Asia Pacific regions. In the past two weeks alone, Nvidia has announced partnership deals with Deutsche Telekom, Orange, Fastweb, Swisscom, Telefónica and Telenor to build the infrastructure required for sovereign AI offerings.
Meanwhile, in the U.K., BT appears to be angling for some of the $1 billion-plus the government plans to pour in to AI infrastructure given it jumped aboard as a founding member of the newly launched U.K. Sovereign AI Industry Forum.
"I think there is some geopolitical motivation to ensure that each region controls the destiny of their data, rather than the data owned by U.S.-based cloud providers," Dell'Oro Group Senior Research Director Baron Fung told Fierce. "The supply of AI infrastructure services is also limited. To ensure availability, the regional sovereign entities deploy the capacity that they need, rather than relying on the U.S. cloud providers to allocate limited capacity to them."
But there's a bit more to it. If you clicked any of those links, you may have noticed that many of them were for Nvidia-issued releases. That’s because Nvidia CEO Jensen Huang has been aggressively pushing telcos worldwide to build AI factories, using sovereignty as a motivating force, AvidThink founder Roy Chua told Fierce. And with governments also pushing their domestic telcos to do the same, it’s no wonder a wave has formed.
What’s in it for telcos?
Pressure aside, Chua said telcos have an incentive to play along in part because they believe if they “own the AI infrastructure and work with Nvidia to build these platforms that some of the learning may transfer into these organizations and make them more ready for AI.”
As Fierce Network Research found, telcos are also looking for a way to escape the “dumb pipe” trap in the AI era after being clobbered during the cloud revolution.
But Chua added it’s still an open question whether the move to build AI factories is a smart one or more of a risky bet.
“It’s still unclear from a telco perspective whether they are the appropriate organization in-country to be acquiring and running these AI factories as opposed to a cloud or data center operator,” he said.
Why? Well, telcos have already been here before, having previously pursued a few different infrastructure businesses (data center and cloud, for instance) that ultimately failed despite big investments.
Take for example, Verizon’s public cloud play in the 2010s. The operator acquired Terremark in 2011 for $1.4 billion to strengthen its bid, only to turn around and shut down its public cloud in 2016, sell off its data centers to Equinix the same year and divest Terremark to IBM a year later.
Sovereignty concerns in Europe may be driving operators there to act, but the question of whether telcos are best placed to run this infrastructure stands.
What could go wrong?
It’s true that a shift from training to inferencing could bring AI to the edge. But while telcos have taken that as a sign that their distributed infrastructure (like central offices) could end up ideally placed for edge compute deployments, it’s still entirely possible that AI could jump straight from the data center to on-device computing.
It all depends how close to users the compute needs to be, how much AI oomph can be packed into smartphones, and how picky governments and citizens are about privacy. (Apple has notably proceeded with restraint despite a flood of hyperscaler data center capex. Maybe they know something we don't?)
Such a development would be horrible news for Telefonica, for instance, which is apparently focusing on the rollout of Nvidia GPU-based edge infrastructure across Spain.
Then there’s the whole GPU problem. AI factories are built on GPUs. But given Nvidia and the like are pumping out newer, more powerful models every 12 months, GPU investments are likely to depreciate fast, Chua said. And if the whole AI factory idea flops, that means telcos aren’t going to make their money back, even if they sell whatever they’ve built.
All of this isn’t to say telcos shouldn’t be exploring their options, Chua said. For instance, if they’re building AI infrastructure for themselves so they can build and provide value-added services, great. But they should proceed with caution until they can prove that the whole promise of AI factories isn’t just smoke and mirrors.
“Hopefully they’re spending in a measured way and measuring outcomes before committing too much capital,” Chua concluded. “I think the next nine months we’ll see the outcomes of these early investments and see if they’re actually worthwhile or not.”