Infrastructure

The substrate everything runs on — from carrier radio and optical transport, through telco-cloud, to the GPU-powered AI data center.

If a workload cannot run without the Linux and Kubernetes layer beneath it, that layer is infrastructure. This domain spans roughly twenty-five years, from the carrier transport and radio plant to today’s GPU fabric. The oldest skills are still part of it, because the same fundamentals carry forward: optical transport became co-packaged optics, and carrier five-nines became data-center resilience. What follows leads with the newest work; the earlier eras are context for how the substrate got here, not claims about where it is now.

Convergence

The substrate is now a fusion of AI compute and networking. Rack-scale accelerated systems built on data-center GPUs (Blackwell) sit alongside Arm host CPUs (Grace) coherently coupled to them, with DPU and SmartNIC offload (BlueField, ConnectX) shearing networking, storage, and security off the host entirely. These arrive packaged as turnkey AI factories — Dell PowerEdge XE and similar — orchestrated through Kubernetes-native ML platforms such as OpenShift AI.

The defining concern of this era is the training fabric. Distributed and LLM training scale over collective-communication backbones (NCCL) riding RDMA, either on InfiniBand or on purpose-built lossless Ethernet such as Spectrum-X. Congestion signalling (ECN) and topology-aware GPU scheduling are what keep an expensive fleet saturated; without them the accelerators idle while the network sorts itself out.

In parallel, the telecom substrate has converged onto the same cloud-native base. Disaggregated virtualized RAN (CloudRAN, O-RAN 7.2x, SMO) runs on carrier-grade Kubernetes distributions such as Sylva and Wind River. AI is being embedded in the network itself through GPU-accelerated RAN (NVIDIA ARC) and core analytics (NWDAF), with 6G on the horizon. General-purpose x86 silicon persists, but its role has changed: it is the control-plane and host layer now, feeding the accelerators rather than doing the heavy numeric work. That CPU/GPU/DPU split is the thing that frames how this era’s data centers are wired.

Transition to AI

The substrate straddled telecom and cloud-native compute, with its centre of gravity in disaggregated mobile access. vRAN and Open RAN build-outs, and private 5G, sat on a carrier-grade telco-cloud foundation (Wind River, OpenShift), with customer-premises equipment marking the edge.

Underneath it all was general-purpose iron and open networking: Dell PowerEdge hosts under iDRAC management, alongside PowerSwitch leaf-spine fabrics running SONiC. These were the physical hosts that the virtualized RAN and the containers actually landed on. VMware’s software-defined data center and the emerging internal-developer-platform pattern sat at the periphery, where operations tooling and developer abstraction were beginning to consolidate over the raw substrate.

Closed-loop assurance

This era is defined by two things: disaggregation of the radio-access network, and the cloud-native turn in the core. O-RAN interfaces (E2, A1, O1, O2) and vRAN split the monolithic basestation into interoperable CU, DU, and RU components. The 5G core recast the mobile core as service-based functions running as CNFs on carrier-grade Kubernetes.

Making general-purpose servers carry telecom traffic demanded a hardware-fluent plumbing layer that ordinary IT clouds never needed: SR-IOV and Multus for line-rate multi-network pods, CPU pinning for deterministic NUMA-aware latency. On that foundation ran 5G NR with massive MIMO and network slicing, with compute pushed outward to the edge cloud and ETSI MEC to cut backhaul and latency.

Cloud-native operations

The substrate pivoted from virtual machines to containers and declarative orchestration. Docker standardized the portable image, while a Kubernetes control plane — API server, scheduler, etcd — became the de-facto pattern for reconciling cluster state. This layered over OpenStack IaaS and SDN overlays such as Contrail.

In a telecom and enterprise context, this cloud-native tooling still landed on operator-owned iron: self-managed on-premises and bare-metal Kubernetes on enterprise data-center platforms, with Ceph providing distributed software-defined storage on commodity nodes. The through-line of the era is hyperscaler-style declarative orchestration married to private, physically-owned substrate.

Distributed systems

This is the hinge where dedicated telecom hardware began dissolving into cloud primitives that still anchor today’s stack. On the radio side, LTE arrived as the all-IP flat 4G core, with VoLTE migrating voice onto the IMS packet core and retiring the circuit-switched 3G world.

Underneath, the enduring pieces took shape: commodity COTS blades displacing bespoke and mainframe iron, HAProxy and SAN block storage as the load-balancing and pooled-storage workhorses, and the private/public/hybrid cloud triad plus CDN edge caching establishing the elastic multi-tenant model. Threading through it all was the first wave of network programmability — SDN’s control/data-plane split, seeded by the now-faded OpenFlow.

Network modeling

The carrier transport foundation everything later was built on. Aggregated voice and data rode synchronous optical pipes (SONET/SDH) and their managed-wrapper successor OTN over DWDM fiber, provisioned through Telcordia’s TL1 command language from operations-support systems.

Above the optics, the packet layer took root: Ethernet framing and best-effort IP routing, with MPLS label-switched paths adding fast forwarding, traffic engineering, and L2/L3 VPNs to the carrier backbone — stitching access, aggregation, and core into one transport network. The network elements themselves were purpose-built embedded systems on real-time operating systems such as VxWorks and QNX; dedicated boxes rather than general-purpose compute. Ethernet and IP endure as living substrate. The SONET/TL1/MPLS transport stack is the largely superseded rootstock that the packet-optical fabrics grew out of.

Control loop

The foundational era was the circuit-switched, standards-defined telecom plant. 2G GSM digital cellular (TDMA, SIM) carried voice, SMS, and early packet data over a SONET/SDH synchronous-TDM transport backbone with sub-50 ms ring protection.

Network elements were embedded systems on real-time kernels such as pSOS, managed through OSI/TMN stacks — CMIP and Q3 interfaces. Even as that was being built, the IETF’s RFC-driven IP, TCP, and routing work was laying the connectionless, packet-oriented groundwork that would eventually subsume all of it.

Infrastructure is the first of four domains on this site. It is the substrate that operations runs, that the software lifecycle delivers onto, and that agentic transformation is now changing the way we build.