Elon and the 200 GW Telco AI Factory

Author: Peter Jarich, Head of GSMA Intelligence 

Part of the big questions being answered at Nova Future Summit. 


If the SpaceX IPO taught us anything (beyond the importance of getting your IPO out ahead of rivals), it was the value of going big when thinking about the promise of technology and our ability to execute on that promise.  

Deployment of 100 gigawatts of space-based compute per year? Sure. Establishment of a permanent human colony on Mars with at least one million inhabitants? Cool. Building out a substitute for all current terrestrial cellular services? Great idea.  

Of course, when the world’s richest man says he plans to compete head-on with all your connectivity services while chasing a 16X bigger AI opportunity, you can forgive telcos for doing their own “big thinking.” Particularly about AI.  

But are they thinking big enough?  

To be sure, they’re all paying attention to potential traffic impacts. And, because monetizing traffic has become increasingly tough, some are trying to monetize AI via selling token packages or pushing deeper into AI and cloud services. Against this backdrop, rolling out compute capacity might seem boring. Right? Maybe. But when frontier model makers are chasing down GPU shipments and locking down datacenter capacity before it’s built, it’s also probably a good business model.  

Telcos get this.  

Most see DC capacity as the biggest thing they can do to support AI usage. Some are rolling out GPU as a service offers and building out AI compute capacity. But if they want to truly scale that support, rolling out sub-scale datacenters will only get you so far. Thinking like Elon and his myriad tiny AI space-based data centers (a marginally more compelling idea vs. using spare Tesla capacity) might get the industry farther. And that’s where the AI-grid concept comes in.  
 

Whether it’s satellites, or EVs or edge nodes or base stations, the concept is similar. Build one giant set of compute from a grid of smaller distributed compute. If this smacks of Telco’s fascination with Edge Compute which hasn’t panned out in a meaningful way, that’s totally fair. But if you think it’s the same old story, you’re missing out on a few things 

• Demand. There was never any shortage of potential telco edge compute use cases. On paper. Matching that up with real world demand in a scalable way, however, has proven difficult. An AI grid removes the need to sell the edge value prop (you’re selling the capacity, not necessarily the location) while targeting demand that is very real in the here and now.  
• Scale. As hyperscalers, neo-clouds and former footwear brands all rush to build out massive AI factories, we need to ask what sort of scale telcos could achieve. In a blog from earlier this year, Nvidia argued that telcos maintain, “distributed network data centers worldwide, spanning regional hubs, mobile switching offices and central offices, with enough spare power to offer more than 100 gigawatts of new AI capacity over time.” It’s a big number but leaves out the RAN. That’s roughly 10M base stations. A small share isn’t connected to the electric grid. The majority are at a smaller number of sites (towers), many of which have added power to tap. Add those up and you get a conservative 100 gigawatts of added capacity. If you’re doing the math, that’s a total of 200 gigawatts – about six to seven times more than total global AI data center capacity in 2025.  
• Edge AI Value Prop. While telcos haven’t always been successful in selling the edge compute promise, AI should change that. Low latency has long been the edge tag line whether or not a critical mass of applications need it. In an AI world, latency will still be important to some use cases, but bandwidth efficiency, resilience and data security should seal the deal.  
• Internal Demands. As much as telcos want to monetize AI, they’re also major AI users – and their token consumption will only grow as they put chatbots and AI-driven automation into scaled use. That could quickly get expensive unless they build out their own AI compute. 

200 GW split across the telco universe isn’t nearly as sexy as half that capacity promised by a single company. And that’s a risk. If telcos don’t see the broader opportunity – and seize that opportunity – it’ll be a major loss of telco AI monetization plans, their operational efficiency, and anyone looking to telcos to deliver sovereign AI services.  
 

Be part of these big conversations at Nova Future Summit  

Join other leaders from across global connectivity and technology this September—for closed-door discussions on the next chapter of intelligence infrastructure and value creation. Request your invitation.