Opinion: AI-RAN could give telecom a second chance to own the AI economy

- AI-RAN could turn the RAN from a permanently hungry cost center into a distributed AI revenue platform - Nokia/NVIDIA and Ericsson are reviving a 40-year telecom argument: specialized infrastructure versus general-purpose compute - The really enormous prize lies beyond the traditional network edge, where AI meets factories, robots, transport, energy and the physical economy When I turned up last Friday to host a discussion of AI-RAN with three top telecom executives, I was expecting a conversation about radio performance, capacity, automation and cost. What I got was a lot more exciting — and way more disruptive. My guests were Nokia Chief Technology and AI Officer Pallavi Mahajan; NVIDIA Vice President for AI and Telecoms Soma Velayutham; and SoftBank Corp. SVP and CTO Ryuji Wakikawa. We were ostensibly there to discuss AI-RAN — using AI and accelerated computing inside the radio access network. But the conversation quickly moved beyond merely “making the RAN smarter.” The much bigger idea is that the vast distributed infrastructure built by telecom operators could become part of the global AI computing platform itself. That would fundamentally change the economics of telecom and perhaps finally answer the question of what carriers are actually supposed to become. The answer to that question is something they have spent most of this century blundering around in the dark, bumping into sharp objects — such as debt, angry shareholders and confusing 5G business models — trying to figure out. Carriers build networks big enough to cope with their busiest periods. The inevitable result is expensive computing capacity sitting around twiddling its digital thumbs at quieter times. NVIDIA’s Velayutham called this “fallow capacity.” His argument is simple: once accelerated compute enters the RAN, why use it only to run the network? Use spare capacity for inference and other AI workloads as well. That matters because AI itself is moving outward. As I argued last year in “The AI opportunity heads to the edge,” the first AI boom was about gigantic centralized GPU clusters; the next one increasingly takes intelligence closer to the people, enterprises and machines actually using it. “Nobody can do this better than delivering AI to consumers and enterprises than the telecom networks,” Velayutham said. The old argument, with much bigger stakes And here we hit one of telecom’s oldest arguments. Nokia is making a major bet on the convergence of RAN and accelerated AI computing. Its alliance with NVIDIA — which included NVIDIA taking a $1 billion stake in Nokia — explicitly envisages Nokia’s 5G and 6G RAN software running on NVIDIA architecture and distributed edge inference becoming part of the operator proposition. Ericsson is taking a materially different route. It is developing AI-RAN but emphasizing software portability and compute choice: Ericsson Cloud RAN can run on COTS hardware, including NVIDIA infrastructure, or on purpose-built Ericsson Silicon. We have, of course, been here before. For at least 40 years — or four years longer than I’ve been covering it — telecom has oscillated between highly optimized purpose-built systems and general-purpose computing promising flexibility, programmability and better economics. Proprietary switches versus software-controlled systems. Appliances versus NFV. Custom baseband versus Cloud RAN. AI-RAN is the latest rematch. What makes this round different is that the general-purpose platform is no longer merely a cheaper server. It could become the computing substrate of the AI economy. There is a second divide, too. The hyperscalers are pouring almost unimaginable amounts of capital into huge centralized AI data centers. Reuters recently calculated that Microsoft, Meta, Oracle, Amazon and Alphabet alone have around $1.09 trillion in future lease commitments, much of it related to data-center expansion. I have been somewhat less restrained about this. In “Meta’s $100B house of GPU cards”, I argued that the extraordinary concentration of capital, power and infrastructure around hyperscale AI factories deserves considerably more scrutiny than it is getting. I am not predicting Lehman Brothers with GPUs. Yet. But anyone who thinks this level of concentrated capital expenditure is risk-free has apparently misplaced their copies of “2008 — That Was the Year That Sucked” and “The Dummy’s Guide to Why Banks That Over-Leverage Complex Debt Securities to Build Stuff Are Really Stupid — Second Edition, Updated to Cover Data Centers as Well as Subprime Mortgages.” What fascinates me about NVIDIA is that, while prospering magnificently from the AI-factory boom, it has also identified the much bigger geography beyond it: distributed inference and physical AI. This is particularly gratifying for someone who enjoys a solid “told you so” moment, because I have been banging on about exactly this for several years. My recent Fierce Network pieces “When the network becomes the operation” and “Introducing: The Unified Infrastructure Stack” make essentially the same argument: AI acquires its greatest economic importance when networks, compute and intelligence become embedded in factories, mines, ports, utilities, logistics and other physical systems. Our own market analysis shows why. FNTV forecasts a $650 billion annual equipment and software market by 2035, of which Industrial OT & Physical Systems represent roughly $435 billion — 67% of the total. That is the hinterland. And it is territory that U.S. hyperscale cloud operators largely ignore. From cost center to AI platform Wakikawa explained why operators desperately need it. SoftBank’s mobile revenue, he said, is declining while infrastructure investment remains unavoidable. “So we keep investing in the infrastructure, but the revenue is declining.” AI-RAN, he said, could shift infrastructure “from a cost center to the revenue generator.” And that, really, is the shining city on a hill for the entire carrier ecosystem: a world where operators stop being data lackeys feeding hyperscalers’ money-printing AI machines and become active participants in the AI economy themselves. SoftBank’s AI Grid already points in that direction, dynamically allocating suitable AI workloads across available GPU capacity. Network compute becomes fungible. Mahajan added that distributed inference is not simply about latency. Power, cooling, sovereignty and the cost of moving data all push intelligence away from gigawatt-scale centralized facilities. Velayutham took the argument to its logical conclusion: “Every base station is now an AI inferencing data center.” Connect millions of them and suddenly telecom possesses a geographically distributed computing fabric extending remarkably close to consumers, enterprises and machines. And this is where things become properly exciting — assuming, like me, you are the sort of person who becomes unreasonably animated by massive transformative shifts in global communications markets. The ultimate AI economy is not just ChatGPT. It is factories. Robots. Vehicles. Energy. Logistics. Healthcare. Industrial automation. Mahajan called the network the “central nervous system” of that emerging grid. That is bang on. Telecom spent the first internet revolution watching enormous amounts of economic value migrate upward to cloud platforms and applications riding across its infrastructure. AI-RAN may give it a second chance. Read more from Stephen Saunders: The AI opportunity heads to the edge Telecom’s great AI dilemma: Everybody wants the future, nobody knows how to monetize it My cunning plan to solve America’s AI computing shortage When the network becomes the operation Introducing: The Unified Infrastructure Stack Meta’s $100B house of GPU cards Join Fierce for our AI-RAN Summit on September 9, featuring speakers from AT&T, T-Mobile, Orange, Nokia, Nvidia, SoftBank and Samsung. Click here to register! Stephen M. Saunders MBE is a communications analyst and USPTO-registered inventor examining how digital infrastructure — 5G, cloud, and AI — is reshaping industry, power and society, as well as underpinning the emerging, ubiquitous global digital economy. As anchor of FNTV and a longtime industry insider, he focuses less on growth narratives and more on execution, risk and how hyperscale technology is distorting markets, governance and society at scale. Opinion pieces from industry experts, analysts or our editorial staff do not represent the opinions of Fierce Network.