Firstly, can I commend Vish’s piece about the implications of the Hugging Face attack? It has been covered everywhere, of course, and has opened up the debate about AI safety, and the AI apocalypse. But Vish has an angle: that the problem might be that AI learns to evade the tests designed to decide whether it is safe – so that falling violation rates only reflect better concealment, and not better behavior. Once the thing being tested learns from the test, and optimizes against the measurement, then the test itself is part of the agent’s game, and becomes a moving target. It should be randomized and independent, he says. Keep it on its toes, and never let it know – sort of thing.
Besides, Ericsson had a second bite of the AI-RAN cherry at Intelligent RAN Forum a couple of weeks back – per the latest in RCR’s serial coverage of the event. AI-RAN isn’t just about making RAN smarter, it says; it’s about making the network a part of the system. The network is both a workload and a platform – and the uplink is killer when the physical world grows a brain. As we know (as the whole industry tells it), and as we’ve already discussed. The thing, out of the forum, is that the for-networks angle has legs already – as told by Optus (to Ericsson), most powerfully. But this for-AI piece, about integrating the network, still sounds hyped-up, rather than buttoned-down.
And then we have some proper news, as linked among all the press releases at the bottom of this letter – including Deutsche Telekom putting some serious numbers on its AI ambitions: €2.5bn in indirect-cost savings by 2030, €800m in AI-related revenue from businesses outside the US. So AI is primarily an efficiency play for the German telco, rather than a new-revenue play – at least on the numbers it is putting forward. So its AI story, as it is, is less about selling AI (networks for AI) than using it (AI for networks, and service) to make its operations cheaper, faster (maybe better too) – with savings recycled into the network, and notably into fiber (for AI).
So yeah, as it stands, the AI story in the telecoms sector looks more like a proof point for AI, than a proof point for telecoms. Give it time, sure; but the telco industry needs to tell a half-decent story about carrying workloads for enterprises on its networks, or networked GPU-gubbins. And to get paid for it – and fast.
James Blackman
Executive Editor
RCR Wireless News
RCR Top Stories
AI-RAN’s needs: Ericsson says AI RAN requires both AI-optimized networks and networks built for AI applications, with stronger uplinks and distributed intelligence becoming increasingly important.
Artificial cunning: The Hugging Face intrusion by AI agents raises a question: can systems learn to evade the safeguards designed to control them? Vish Nandlall argues that falling violation rates may not mean safer systems. Testing needed.
Multi-hop mmWave: How to extend high-frequency mmWave coverage without just adding more base stations? Kyocera’s multi-hop repeater validation points to a more flexible architecture for NLOS environments.
Starlink binaries: Starlink is not about to replace terrestrial networks – but its the threat is not an idle one. As orbital capacity scales, the contest is over the edges of connectivity, spectrum, infrastructure and, ultimately, who controls the customer.
London DC crunch: London remains EMEA’s largest data center market, but power, planning, and grid constraints are increasingly shaping delivery as strong demand keeps vacancy low and projects move forward despite tighter conditions.
AI-Powered Telecom Infrastructure
Supermicro, in collaboration with NVIDIA, delivers AI-powered infrastructure tailored for telcos, enhancing operational efficiency, network management, and customer experiences. Explore now
Beyond the Headlines
Beyond connectivity: Physical AI could push networks beyond connectivity, with Samsung saying future systems will need to coordinate connectivity, distributed computing and workload placement across devices, edge and cloud.
Agent operators: Verizon is moving from scripted RAN automation to agentic network ops, where AI reasons across domains, shares context and takes action – with the operator retaining control of intent, outcomes and guardrails.
Supercycle fever: AT&T has signed a $3bn fiber supply deal with Corning, adding to a string of major commitments as AI data-center construction tightens optical supply and telcos bulk up their networks.
AI-RAN risks: Don’t repeat your 5G mistakes with AI-RAN, SK Telecom tells its fellow telcos – while Optus considers how they should re-architect their infrastructure as “social networks for agents”.
IoT war chest: Onomondo is getting a new owner and €100m in fresh funding as it pitches integrated global infrastructure, rather than just airtime – as the platform for IoT sensing to support physical AI sense-making.
What We're Reading
DT eyes AI cuts: Deutsche Telekom expects €2.5bn in indirect cost savings from AI by 2030, while AI-related revenue outside the US reaches €800m. AI is also deployed across its network operations and customer service.
BT buys TalkTalk: BT has acquired TalkTalk and PlatformX, out of admin, ostensibly to protect 2.5m customers and critical services. The deal carries an estimated £400m FY27 cash impact, with regulatory review expected.
Home-made P5G: Eight UK firms have launched a £4.3m private 5G project to integrate home-grown 5G tech into a multi-vendor private network, backed by £3.1m from Innovate UK. MOSAIC-5G aims to cut deployment from weeks to hours.
Key cross-licensing: Huawei and Qualcomm have agreed cross-licensing for 5G, computing, AI, and network patents, with Qualcomm also buying certain Huawei US patents. The agreement remains subject to regulatory approval.
B2B comms platform: Airtel Business has a new enterprise platform combining messaging, RCS, and voice with fraud protection and network intelligence. The DPDP-compliant service works as a single control plane for customer comms.
