Can Operators Extract More Capacity Without Building New Sites?
African telecom operators face a capital allocation challenge long before they encounter a technology decision. Network coverage remains incomplete, capital costs are elevated, and the list of competing investment priorities is extensive: constructing new cell sites, expanding 5G coverage, laying fiber backhaul, building data centers, and now acquiring GPU compute capacity.
Artificial intelligence complicates this capital allocation process because it yields returns across two distinct layers. At the network layer, AI can optimize existing spectrum, cut operational expenses, and extract higher capacity from deployed infrastructure. Above the network layer, AI supports a specialized compute business that sells GPU access and enterprise model integration.
These two opportunities carry fundamentally different risk profiles, depreciation schedules, and operational demands. The strategic challenge for operators lies in correctly sequencing these investments.
This #TechTalkThursday examines what AI-driven network optimization delivers on the ground, where the limits of spectral efficiency sit, and how operators can evaluate near-term operational savings against larger, high-risk compute bets.
Distinguishing Capacity Gains from Physical Network Expansion
AI-driven Radio Access Network (AI RAN) technology optimizes how radio frequencies are scheduled and configured across existing cell sites. While this software layer improves capacity and throughput, it cannot create a physical tower where none exists, nor can it extend coverage into unserved rural regions.
The distinction between software efficiency and physical coverage is critical when analyzing vendor performance data. For example, an operator deploying AI scheduling on an existing 4G network reported an 11% increase in network capacity alongside a 40% improvement in peak-hour user throughput. While these vendor-published results highlight genuine performance gains from existing spectrum, they represent optimization rather than expansion.
The boundaries of this approach become clear when viewed alongside physical buildout metrics. During the same deployment period, the operator completed 603 of its planned 1,228 new cell sites , leaving its 4G population coverage at 82.23%, short of its 85% strategic target. AI software increased the throughput of existing towers, but it could not substitute for missing steel, fiber backhaul, or physical site construction.
AI RAN helps operators extract maximum utility from built infrastructure, but it does not reduce the total number of physical cell sites required to close national coverage gaps.
Resolving Operational Silos to Cut Rural Site Expenses
For operators managing expansive rural footprints, operational expense (opex) reduction is often more valuable than raw spectral capacity gains. Energy consumption, remote maintenance, and field intervention costs dominate the economics of rural cell sites, where average revenue per user (ARPU) remains low.
The primary obstacle to operational efficiency is system fragmentation. Operators generate massive volumes of performance, power, and field telemetry, but this data frequently remains locked in isolated operational silos.
Highlighting this structural friction in a TechAfrica News interview, Jawad Maaloum, CEO and Co-founder of MLnetworks, noted:
“It’s not a lack of data. Telecom operators already have massive amounts of data across network performance, energy, OSS/BSS, and field operations. The real issue is that this data is fragmented across silos that don’t operate in a connected or operationally aware way. Each system sees only a slice of reality, and none of them naturally translate into a unified, actionable view of the network. So even when something is obvious at a system level, like a site that is underperforming and over-consuming energy at the same time, no single team has the full picture, and no system is responsible for turning that into coordinated action.”
— Jawad Maaloum, CEO and Co-founder, MLnetworks
Connecting these fragmented data streams through automated network platforms allows operators to address operational issues proactively, such as identifying a tower that is underperforming while simultaneously over-consuming power.
Field results from automated self-optimizing network platforms demonstrate efficiency gains of up to 75%, alongside drastic reductions in minor fault repair times, cutting resolution windows from four days down to roughly 35 minutes across legacy 2G and 3G sites as well as modern 4G and 5G networks.
For an operator managing marginal rural towers, shortening repair windows and reducing generator fuel consumption directly protects operating margins, helping keep low-ARPU sites financially viable.
Structural Differences Between Telecom Infrastructure and Compute Assets
The operational logic used to justify AI network optimization is increasingly cited to support building dedicated AI compute infrastructure. However, operating a GPU-based AI factory is fundamentally different from running a traditional telecommunications network.
The core divergence lies in asset depreciation. Traditional telecom hardware follows a five-to-seven-year refresh cycle. In contrast, AI compute assets depreciate rapidly, with GPU hardware requiring replacement every 18 to 24 months as newer processing architectures emerge. This rapid turnover forces compute operators to generate returns within a significantly narrower window.
Power availability represents another major structural constraint. Data center demand across Africa is projected to grow from 0.4GW to over 1.5GW by 2030 , requiring up to $20 billion in infrastructure investment according to estimates cited by the World Economic Forum. However, total installed data center capacity across key markets like Nigeria, Kenya, South Africa, Egypt, and Morocco remains under 500MW.
To manage these financial and power constraints, operators are adopting different investment structures:
Capital-Light Joint Ventures:
Structuring minority-stake joint ventures with foreign capital partners, such as launching large-scale AI data center programs in Nigeria and South Africa, allows operators to participate in compute infrastructure without carrying full balance-sheet exposure.
Vertically Integrated Services:
Deploying specialized AI factories alongside enterprise model distribution, such as launching NVIDIA-powered facilities paired with reseller access to models like Anthropic’s Claude, creates an end-to-end enterprise service layer.
Outlining the commercial rationale behind enterprise model delivery, Hardy Pemhiwa, President and Group CEO of Cassava Technologies, emphasized:
“What matters to our customers is being able to use AI in their business, and that starts with access to the best models. Our goal is simple, advanced AI within reach of every business, government, researcher, entrepreneur and university across the emerging markets globally.”
— Hardy Pemhiwa, President and Group CEO, Cassava Technologies
While enterprise compute presents long-term market opportunities, it carries short asset lifespans, heavy power requirements, and direct competition from global hyperscalers. Operators must manage compute investments as an independent business line rather than a simple extension of connectivity.
Four Tests for Evaluating AI Performance Claims and Strategic Allocations
To evaluate competing AI products and infrastructure proposals, decision-makers require clear evaluation criteria to separate verified field performance from vendor projections. Four primary tests provide a structured evaluation framework:
Measured Results Versus Future Projections
Near-term capacity improvements on deployed hardware represent verified operational data. In contrast, vendor roadmaps promising multi-year spectral efficiency gains of over 100% represent forward-looking projections that should not be treated as current capabilities.
Alignment with Operational Bottlenecks
Spectral capacity gains solve urban congestion, whereas automated fault detection and power optimization address rural site viability. Operators must align software adoption with their specific network bottlenecks.
Capital Reinvestment Strategy
Operational savings generated by software optimization only create long-term value if the resulting margin is explicitly reinvested into physical buildouts, fiber backhaul, or coverage expansion.
Independent Governance for Compute Assets
AI data centers and model resale require distinct risk pricing, shorter depreciation modeling, and dedicated operational governance to ensure they do not draw capital away from core network obligations.
Addressing the necessity of clear commercial alignment, Nitesh Singh, Managing Director for Communications, Media, and Technology at Accenture, noted:
“Without a clear value proposition, AI won’t deliver lasting impact.”
— Nitesh Singh, Managing Director for Communications, Media, and Technology, Accenture
Reinvesting Efficiency Savings into Core Connectivity Growth
AI software delivers measurable operational value to African telecom networks today, expanding capacity on congested urban sites, reducing fault resolution times from days to minutes, and lowering operating costs on energy-intensive rural towers.
However, software optimization is inherently finite. It cannot build missing cell towers, lay fiber backhaul, or acquire raw spectrum.
The near-term efficiency dividend buys operators time and operating margin. The ultimate strategic test is how leadership allocates that margin. Reinvesting software-driven savings into physical network expansion converts short-term operational efficiency into long-term market coverage.

