An article by Vikas Kumar AI Strategy Lead, Telcos APJ at Dell Technologies & Harbinder Lally, P.Eng, Professional Engineer .

The Network Already Knows. Your AI Just Can’t See It Yet.

Most telco AI initiatives don’t stall at the model — they stall at the data. Dell Technologies and Ocient are changing that by giving network teams governed access to full-fidelity data across all network data sources, in real time, without replacing existing OSS or BSS systems.

Why Telco AI Stalls at the Data Layer

The short answer: Network data is siloed, pre-aggregated, or discarded before it can be used. When an AI model can’t access complete, timely data, it can’t produce reliable answers.

Consider a common scenario: At 2am, a degradation ripples across a metro cluster. Telemetry catches it in seconds: latency climbing, sessions dropping on a handful of cells. Then come the questions every assurance team knows by heart. What changed? Was it the configuration pushed last night? Which control-plane events correlate with the symptom? Did the user plane carry the impact and, the question that decides escalation, how many customers were hit, and were any of them roamers, enterprise accounts, first responders, VIPs?

This is the quiet constraint underneath almost every telco AI ambition – autonomous networks (AN), AI-driven assurance, smarter customer experience (CX), stronger fraud defence. These programmes rarely stall at the model; they stall at the data.

What Are the Five Network Data Silos Telcos Need to Unify?

Answering any of these questions means stitching five very different sources into one picture across the exact window an event occurred. The user plane records what subscribers did. The control plane records how the network decided to do it.  Telemetry shows how the infrastructure behaved. OSS holds configuration and topology. BSS holds the subscriber, product and service context that says who was affected and whether it mattered. The first three tell you what happened. The last two make it interpretable.

 

Figure 1: The five network data sources that matter for telco AI 

Operators don’t have to accept data that ages out before you can act on it.  It isn’t that a platform can’t express the query.  It’s that the data is siloed, sampled at ingest to avoid overwhelming storage costs, held in the wrong format, pre-aggregated, or aged out of a rolling window before anyone thought to ask. The data was deleted before the right question could be asked, because retaining every record at full granularity seemed too costly. Networks sample their data and keep it only briefly. Both choices put uncertainty into the dataset before any AI touches it. Remove that one constraint and almost everything downstream changes.  When AI models train on complete, high-resolution data rather than sampled approximations, the results are measurably different: faster fault detection, more accurate root cause analysis, and insights that reflect what happened on the network, not a statistical estimate of it.

How Do Dell AI Data Platform and OcientAIQ Turn Full-Fidelity Data Into AI Answers?

Keep those five sources whole, synchronized and join them on demand, and the outcomes operators have chased stop being aspirational.

The architecture pairs two complementary strengths.

• Within the Dell AI Factory framework, theDell AI Data Platform turns that governed data into AI.  It runs analytics, processing, search and orchestration engines, and building the semantic layer and knowledge graph that give agents, OSS and BSS a governed surface.
• OcientAIQ ingests, retains and analyses high-volume network data at full-fidelity and cost-effective scale keeping the raw records and producing governed datasets ready for analysis

One keeps every signal; the other makes every signal answerable without replacing the OSS, BSS or catalog of metadata that is being stored already in place. Build it once and reuse it, and the tenth use case is far cheaper than the first.

What Does this Mean for Telcos?

• Network Assurance & Autonomous Operations – Full-fidelity root cause and closed-loop assurance
• OSS / BSS Integration – One governed foundation across network and business systems
• Security & Fraud Detection – Anomaly and threat detection on full resolution data
• Lower TCO with On-Prem AI – CPU-first analytics and right sized models cut cost

Those four outcomes are the destination. What makes them credible is how quickly they surface in day-to-day operations not as a future roadmap, but as wins the assurance, security and network teams feel every shift. The same governed foundation that answers the 2am question is the one that runs the shift after it: alarm storms suppressed at source, silent 5G degradations caught early, siloed data unified without movement, AI grounded in the operator’s own procedures and history, and faults traced through to the customers and VIPs they affect (Figure 2).

Figure 2: Common telco challenges, the capability that addresses each, and the operational value it delivers.

One Governed Platform for Every Data Type

Dell AI Data Platform handles structured, semi-structured, and unstructured data together on one open foundation, so agents and teams always work from governed, contextually rich information.

Much of the context needed to interpret an event is not structured records at all. It lives in MOP/SOP procedures, ITSM incident and change history, raw logs, configuration exports, images and even LiDAR from field surveys. A carrier-grade foundation must handle all types of data types.

                                    Figure 3: One platform for structured, semi-structured and unstructured telco data

The Dell AI Data Platform is purpose-built engines on one open foundation:

• Analytics – federated SQL across lakehouses, databases, and object stores, so data can be queried where it lives without unnecessary copying.
• Processing – Apache Spark–based batch, streaming, transformation, and AI/ML preparation pipelines.
• Search – Elasticsearch-powered hybrid retrieval that combines keyword and vector search for RAG, semantic search, observability, and unstructured content discovery
• Orchestration – governed multimodal workflows that ingest, chunk, index, enrich, and coordinate AI pipelines across the lifecycle.

OcientAIQ adding full-fidelity landing and analytics for the highest-volume data.  Three capabilities make this practical:

1. Ingest and orchestrate every data type The orchestration engine automates ingestion, chunking, indexing and metadata enrichment, turning PDFs, MOP/SOP procedures, ITSM tickets and logs into governed, retrieval-ready data products; grounding RAG and agent workflows in the operator’s own history, not generic knowledge.
2. Federate instead of copying Federated SQL queries data in place across lakehouses, databases and object stores, avoiding an ETL copy, staleness and egress per use case. It interoperates with existing inventory and OSS/BSS rather than replacing them, and supports ODF, convertible to TM Forum SID so network, inventory and service data share one semantic model.
3. Connect meaning and serve it as context The semantic layer and knowledge graph are built inside the platform on the same governed store, aligned to a telco ontology (TM Forum SID/ODF). That becomes the context layer, exposed through an MCP server and TM Forum Open APIs – TMF621, TMF628, TMF639, TMF641 and TMF642; enabling agents, OSS and BSS call one governed surface for contextualised, permissioned answers.

How the Infrastructure Works Under the Hood

Dell AI Data Platform provides the governed path from source to AI consumption – land, structure, consume with PowerScale and ObjectScale underneath and NVIDIA RAPIDS, cuDF and cuVS for training and vector search. OcientAIQ sits in front as the full-fidelity landing and preparation layer, publishing in through open formats and connectors (Parquet/CSV extract, Spark push-down, JDBC, Python and federated query).

The CPU/GPU split is deliberate: OcientAIQ’s analytics run CPU-first because the work is bandwidth-bound, not FLOPS-bound, with GPUs reserved for the dense floating-point work model training and vector search, fed reduced, governed results rather than raw records.

The Tests for Telco Grade Data Platform Architecture

Figure 4: The five tests a carrier-grade data platform is a must for telecom workloads