Inference takes the wheel
The AI infrastructure conversation this week has been around the fact that compute needs are starting to center on inference, rather than training. Synopsys is deploying agentic AI to design the very chips that will run inference workloads, claiming 25-40% faster debug cycles. Cisco is laying out a platform architecture for agentic network operations, arguing service providers can’t monetize AI unless they first automate how they run their networks. And a Nokia-led discussion framed the emerging reality as “scale-beyond” — a world where AI compute is distributed across interconnected domains making placement decisions based on latency, cost, and available power. The common thread: the hard problem is no longer raw capacity. It’s coordination.
That theme shows up just as clearly on the connectivity side. Amazon’s announcement of a 5,105-satellite direct-to-device constellation could be read as another salvo in the satellite broadband wars, but the more interesting signal is how carefully Amazon is framing it as a partnership play with mobile operators rather than a disruption play. The industry’s vocabulary is shifting from D2D to NTN, and from NTN toward “3-D networks” — the idea that fiber, 5G, Wi-Fi, and LEO satellites become different routes in a single system. It’s a smart framework, but Opensignal’s reality check on consumer expectations is a useful corrective: users don’t reset their expectations just because the signal came from space, and most aren’t willing to pay a premium for intermittent backup coverage.
On the agentic front, the gap between technical readiness and organizational readiness is becoming the central tension. ABI Research argues network agents are ready for commercial deployment, but warns that operators who bolt AI onto legacy architectures risk being reduced to bit-pipe providers while hyperscalers capture the service layer. Cisco’s Kazamel made a similar point more diplomatically at DTW Ignite, noting that the technology exists, but change management is “the bigger hurdle.” Treating agents as extensions of the network operations team sounds obvious, but it represents a cultural shift that most telcos haven’t begun to make.
Christian de Looper
Editor
RCRTech
AI Infrastructure Top Stories
Synopsys brings agentic AI workflows to chip design: The EDA giant’s new multi-agent workflows claim to cut debug cycle times by 25-40%, with AMD evaluating the tools on AI infrastructure silicon through Microsoft Discovery.
The next shift in AI infrastructure is scale-beyond: Nokia, Red Hat, and CoreWeave argue that distributed inference workloads will demand networks that coordinate placement across domains by latency, sovereignty, and cost, not just deliver raw bandwidth.
Cisco charts platform path to agentic network operations: Cisco’s Rana El Desouky Kazamel argues telcos need a North Star architecture to make agents trustworthy in high-consequence network environments, warning that change management is the bigger hurdle.
AI Today: What You Need to Know
Amazon’s satellite play for telcos: Amazon plans 5,105 LEO satellites for D2D from 2028, but its careful framing as an operator partnership rather than a disruption signals a maturing approach to satellite-terrestrial integration.
Network agents are ready — are telcos? ABI Research says agentic AI is implementation-ready but warns operators face a choice between AI overlays and AI-native redesigns — with the risk of being reduced to bit-pipe providers.
Satellite services face consumer reality check: Opensignal warns the D2D industry risks undermining its commercial potential by letting 150 Mbps future targets define consumer expectations today, when most users won’t pay extra for intermittent backup service.
Private 5G’s hype and hard maths: Nokia’s retreat from a key private 5G value proposition highlights a market growing fast but not explosively, raising questions about how quickly enterprises will adopt the connectivity needed for physical AI.
AT&T, Ericsson turn 5G towers into sensors: A successful trial shows how existing 5G towers can double as sensor grids, pointing to a future where mobile networks generate revenue well beyond connectivity