When a data centre developer approaches a government, the number that gets the most attention is usually the requested power supply capacity. A proposed facility may request 10 MW, 20 MW or 50 MW of electricity capacity. That number matters. It tells a system planner about the scale of the requested connection and, potentially, the generation and network infrastructure needed to supply it. However, it does not tell the whole story.
Two data centres asking for the same 50 MW connection may place quite different demands on the power system. What matters is not only how much electricity the facility requests, but what equipment sits behind that demand, how densely that equipment is packed, how efficiently the facility operates, how quickly its demand changes and what infrastructure is required to keep it running. For decision-makers assessing AI data centre infrastructure proposals, the question therefore needs to move from “how many megawatts?” to “what is behind the megawatts?”.
This article, the second in the Powering AI series, looks inside the data centre from the power system's perspective. Three terms provide a useful starting point: load, rack density and Power Usage Effectiveness (PUE). Together, they help explain what a data centre is asking the grid to provide. Understanding them will help decision-makers to ask sharper questions beyond the number on the cover page.
Load: how much electricity, and when?
In its simplest sense, load is the amount of electrical power a facility draws from the energy system at a particular time. Not every kilowatt entering a data centre is being used for computing. Some electricity powers the servers and processors doing the actual computational work. The rest supports the environment that allows those machines to operate reliably.
If a developer requests a 50 MW connection, however, that does not necessarily mean that the facility will consume 50 MW continuously from the day it opens. Data centres are often developed in phases, with servers installed as customer demand grows. A developer may therefore seek sufficient connection capacity for the facility's eventual size while initially using much less. For a power-system planner, this distinction matters. Developers often reserve more grid capacity at the outset than they need immediately, to avoid returning for a second connection later (IEA, 2026). For the grid, then, 50 MW is not simply 50 MW. The timing, utilisation and behaviour of that load matter too.
A grid designed to accommodate a 50 MW connection must have enough capacity at the relevant substation, transmission or distribution lines and upstream generation to supply that demand when it materialises. If the facility reaches full utilisation only a few years later, the system may have reserved capacity that is not used immediately. The first questions should therefore be straightforward: What is the requested connection capacity? What is the expected initial load? How quickly will demand increase? What is the expected maximum demand? And what does the facility's demand look like over a day, a week and a year?
AI introduces another dimension. AI training and inference workloads can cause large, rapid changes in power demand. For grid operators, this raises a practical question distinct from total capacity: can the local network absorb rapid swings without protection systems tripping or voltage becoming unstable, and does the facility have batteries or other on-site flexibility to smooth its own demand? Asking about load, rather than only peak capacity, reveals both the timing of a facility's real demand and how well it can integrate into a fluctuating grid.
Rack density: how much computing is being concentrated in one place?
Servers in a data centre are normally installed in racks. Rack density describes how much electrical power the computing equipment within each rack requires, usually expressed in kilowatts per rack. This matters because AI is pushing far more computing power into the same physical space. Traditional computing facilities can spread their servers across many relatively low-density racks. AI facilities increasingly use clusters of specialised processors that concentrate much larger electrical loads into individual racks. The IEA reports that the power density of AI servers increased eleven-fold between 2020 and 2025 and could increase another four-fold by 2027 (IEA, 2026) .
A typical enterprise server rack today draws around 9 to 12 kW (Data Center Knowledge, 2026a). A rack built around Nvidia's current-generation AI hardware can draw well over 100 kW; Nvidia's own reference design for its GB300 platform specifies up to around 142 kW per rack, and successor platforms now entering production are expected to push well beyond that (Data Center Knowledge, 2026b). This is not a marginal difference. A modern AI rack can draw more than ten times what a conventional rack draws in the same footprint.
Recent US guidance on energy-efficient data-centre design notes that high-performance computing facilities have already moved beyond 125 kW per compute rack in some cases (Brown, 2026). That concentration has consequences. A rack drawing 100 kW doesn't just need 100 kW delivered to it. The facility's internal electrical systems must carry that power safely and reliably. Almost all the electricity processors consume eventually becomes heat that must be removed.
Higher rack density therefore affects the design of the facility's power distribution and cooling systems. At sufficiently high densities, conventional air cooling becomes increasingly difficult, encouraging operators to use liquid cooling systems that transfer heat more directly from processors. Lawrence Berkeley National Laboratory notes that direct-to-chip liquid cooling can reduce cooling-related energy use in heat-dense systems.
For policymakers, rack density provides a useful clue about the kind of facility being proposed. A developer describing a project as an “AI data centre” should be able to explain what computing infrastructure will actually be installed, the expected power density of its racks and how those densities are expected to change as the facility develops. Two data centres requesting the same megawatt figure can be doing very different things once you account for rack density.
Rack density therefore helps distinguish a genuine high-density AI deployment (with the cooling, water and grid-stability implications that follow) from a facility that uses the language of "AI data centre" to describe more conventional hosting or storage activity at ordinary density.
PUE: how much electricity reaches the computing equipment?
Power Usage Effectiveness (PUE) compares the total electricity a facility draws to the electricity that reaches its computing equipment. It is expressed as a ratio of total facility power to IT equipment power. A PUE of 1.0 would mean that every unit of electricity entering the facility went directly to computing, with no electricity required for cooling, power conversion or other supporting infrastructure. In practice, this is a theoretical limit that no real facility reaches. A PUE of 2.0 means only half of the incoming electricity reaches the computers.
Suppose two facilities each require their IT equipment to consume 50 MW. If one operates at a PUE of 1.2, the total facility demand would be about 60 MW. At a PUE of 1.5, the same 50 MW of computing demand would require about 75 MW from the power system. The computing requirement is identical. The demand seen by the grid is not. This is why PUE matters to energy planners. The industry's average annual PUE was about 1.52 in Uptime Institute's 2026 survey, although newer large facilities routinely report figures of 1.3 or lower (Davis, 2026).
In practice, PUE varies widely. Older or poorly designed facilities can run above 2.0, while the most efficient modern hyperscale data centres, built with purpose-designed cooling, report PUEs close to 1.1 (Data Center Knowledge, 2026a) (Huawei Digital Power, 2025). Global averages sit well above that best-in-class figure, and the industry's central expectation is that average PUE will keep falling over the rest of this decade as newer, more efficient facilities are built and older ones are upgraded (IEA, 2025) .
However, policymakers must interpret PUE carefully. A lower number generally indicates more efficient supporting infrastructure, but it does not tell policymakers whether the computing itself is efficient. Nor does it capture every environmental trade-off involved in achieving that efficiency. Uptime Institute cautions, for example, that factors such as water consumption sit outside the metric. PUE therefore answers one useful question: How much additional energy does the facility require to support the computing taking place inside it?
Reviewing a proposal with these terms in mind
None of these three terms turns a regulator into an engineer, and that is not the point. Together, they give policymakers a way to move past the headline megawatt figure and ask what is actually behind it. A 100 MW proposal is not enough information. Imagine a developer announces plans for a 100 MW AI data centre. Before deciding what that project means for the power system, policymakers need considerably more information.
Is 100 MW the facility's maximum connection capacity, its expected operating load or the power required by its IT equipment alone? Will the facility begin at 20 MW and expand to 100 MW over five years, or does it expect most of that demand immediately? What proportion of the load will actually serve computing equipment? What PUE is expected under local climatic conditions? What rack densities are planned? Will the facility use air or liquid cooling? How much water will it require? Does its load remain relatively steady, or will AI workloads create significant power fluctuations?
And one further question matters enormously for a grid planner: how firm does that 100 MW need to be? A facility that can temporarily reduce its grid demand during periods of system stress creates a different planning problem than one requiring uninterrupted supply at maximum load at all times. The first article in this series showed why the terms on which a data centre connects to the grid matter. Understanding what sits behind its requested megawatts is the necessary step before negotiating those terms.
The water behind the cooling
Computing equipment generates heat continuously, and that heat must ultimately go somewhere. Estimates of cooling's share of total data-centre electricity use vary by facility and climate, but industry assessments commonly put it between 30% and 45% of the total (IEA, 2026) (IEA, 2025). Powerful GPUs generate far more heat per unit of floor space than the servers that ran most conventional digital services a decade ago. As AI hardware spreads, the cooling load grows too.
Data centres solve this problem in different ways. Some rely primarily on air-based systems. Others use chilled-water systems, evaporative cooling, direct-to-chip liquid cooling or combinations of these technologies. Some systems recirculate water internally, while others consume water through evaporation as they reject heat from the facility.
The choice creates a trade-off that matters especially in water-stressed locations. Some cooling designs can reduce electricity consumption but increase water use. Others can reduce water consumption while requiring more electricity. PUE alone will not reveal this trade-off. The US Department of Energy notes, for example, that measures intended to reduce cooling-water consumption can themselves increase energy use and worsen PUE. How much water a facility needs depends heavily on its cooling technology and climate: one recent analysis found evaporative water use across hyperscale facilities in different US climate zones ranging from about 0.1 to 1.8 litres per kilowatt-hour of IT electricity used, against 2 to 4 litres per kilowatt-hour for traditional evaporative cooling systems, a roughly tenfold difference depending on the technology chosen (Akinade et al., 2026).
Two African examples show why this matters to policymakers, not just operators. In Cape Town, opponents of a planned 174 MW data-centre campus calculated, using an industry rule of thumb of about 25.5 million litres of water per megawatt per year for conventional cooling, that the facility could consume roughly 4.4 billion litres of water annually; a figure that had not featured in the original land-use application, which covered electricity and space but not water (Kretzmann, 2026). The gap was less an oversight than a structural one: the officials who approve land use and the utilities that manage water supply are not always the same people who approve an electricity connection, so a proposal can clear an energy review without ever being tested against a city's water balance.
In Kenya, a large data centre proposed for Olkaria, near Lake Naivasha, has drawn scrutiny not only over its power requirements but over what a new large industrial water user would mean in an area where residents already report shortages, even though the project is designed to run on geothermal energy and includes water-conservation measures (Al Jazeera, 2026). In both cases, the institutions responsible for water were distinct from those responsible for power, illustrating why data-centre water use is not simply a subset of the electricity conversation. It is a separate infrastructure and political question, and one that a purely energy-focused review can miss.
Current grid realities, and why they change what these terms mean
None of the three terms above means the same thing in every African power system, because the systems they apply to are not in the same condition. Africa has around 313 GW of installed generation capacity for close to a fifth of the world's projected population by 2030, and produces under 6% of the world's electricity (African Energy Week, 2026).
Demand is forecast to grow 4-6% a year this decade, and the continent needs to add roughly 182 GW of new capacity by 2030 just to keep pace; unreliable supply already costs sub-Saharan economies an estimated 2-5% of GDP a year (African Energy Week, 2026). Regional power pools illustrate the same strain at a smaller scale: the Southern African Power Pool serves some 400 million people across 12 countries from about 82 GW of installed capacity, and still faces deficits, congested transmission and ageing infrastructure (Capital Ethiopia, 2026). Set against this backdrop, a data-centre proposal is not landing on a system with comfortable headroom; it is landing on one already working to catch up with its own population and economy.
However, "grid strain" does not mean the same constraint everywhere, and this is where load, rack density and PUE start to matter differently depending on which part of the system is under pressure.
In South Africa, the binding constraint is shifting from generation to location. National supply has stabilised. Eskom's General Recovery Plan delivered over a year of uninterrupted supply, and the newly approved Integrated Resource Plan envisions 105 GW of additional generation by 2039 (Engineering News, 2026). Yet Eskom's own medium-term outlook shows that adequacy is now as much about where power is generated as how much exists nationally: specific load centres such as Gauteng and North West face localised deficits from network congestion, even as the national balance looks sufficient on paper, and roughly 9.5 GW of existing capacity is due to retire by 2030 before major transmission upgrades are complete (Daily Maverick, 2025). For a data centre here, the question is less "does the country have enough megawatts" and more "does this substation, on this part of the network, have room", which is precisely why rack density and load profile at a specific site matter more than a national capacity statistic.
In Nigeria, the constraint is reliability itself. The national grid collapsed multiple times in the first weeks of 2026 alone, after two major collapses in 2025 and a reported twelve in 2024, with frequency performance oscillating well outside the safe operating band regulators require. Industry warnings suggest the country could see far more collapses in 2026 without structural intervention (Davidson,2026). In a system this unstable, PUE and rack density are almost secondary; the immediate question a data-centre proposal raises is how much of its own generation, batteries or backup it needs simply to keep computing continuous, echoing the diesel dependence already documented among Nigeria's existing facilities.
Kenya sits between these two pictures: a smaller system with real capacity in reserve on paper, but concentrated in a handful of generation sources; mainly geothermal and hydropower, whose seasonal and site-specific limits mean a single large, high-density facility can still strain a specific transmission corridor or a specific water catchment, as the Naivasha case above illustrates, even where national totals look adequate.
What does this mean for policymakers?
The lesson for policymakers is that these three terms (load, rack density and PUE) are diagnostic tools, not fixed thresholds. Together, they turn a headline announcement into a more useful conversation. For African countries seeking data-centre and AI investment, this conversation matters. Governments need digital infrastructure, but they also manage power systems with finite capacity, and millions of homes and businesses still need more reliable, affordable supply. Before policymakers can decide what an AI data centre means for the grid, they first need to understand what is actually behind the megawatts.
In addition, a data-centre proposal can create two infrastructure questions at once. Can the electricity system supply the facility? And can the local water system support the way it requires to remove heat? Those questions may fall to entirely different institutions. An energy regulator may examine the electricity connection. A water utility, environmental regulator or local government may assess water abstraction, discharge or supply. An investment agency may be negotiating incentives, while a ministry responsible for digital policy focuses on connectivity and data infrastructure.
The data centre, however, experiences these systems as one operating environment. For governments, that is an argument for assessing large AI infrastructure proposals across institutional boundaries rather than treating power, water, land and digital connectivity as separate approvals.
References
African Energy Week (2026), "Africa's Demographic Surge is Forcing a Power Reckoning," 16 February 2026. https://aecweek.com/news/africas-demographic-surge-forcing-power-reckoning
Akinade B. A., Amanambu, A. C., Frame, J. M., Ren S. (2026). AI data centers and the water use feedback loop [Preprint]. https://arxiv.org/pdf/2606.21760
Brown, J.W.J. (2026) ‘Rethinking data centre design’, REMI Network, 17 September. Available at: https://www.reminetwork.com/articles/rethinking-data-centre-management/
Al Jazeera (2026), "Can Kenya's AI ambitions coexist with Naivasha's water needs?", 24 August 2026. https://www.aljazeera.com/economy/2026/8/24/can-kenyas-ai-ambitions-coexist-with-naivashas-water-needs
Capital Ethiopia (2026), reporting on the AFREC–IEA webinar on Africa's power grids, 6 September 2026. https://capitalethiopia.com/2026/09/06/africas-power-grids-face-rising-demand-high-losses-and-digitalisation-pressure/
Daily Maverick (2025), reporting on Eskom's Medium-Term System Adequacy Outlook 2026-2030, https://www.dailymaverick.co.za/article/2025-11-03-risks-flagged-in-eskoms-2026-30-system-adequacy-outlook-require-critical-balancing-act/
Data Center Knowledge (2026a), "AI Rack Density's Real Limits: Power, Cooling, Failure Risk," citing Uptime Institute's 16th Annual Global Data Center Survey and Nvidia's GB300 NVL72 reference architecture.https://www.datacenterknowledge.com/ai-data-centers/ai-rack-density-s-real-limits-power-cooling-failure-risk
Data Center Knowledge (2026b), "Data Center Rack Density Has Doubled. And It's Still Not Enough." https://www.datacenterknowledge.com/hardware/data-center-rack-density-has-doubled-and-its-still-not-enough
Davidson Iriekpen (2026), "Nigeria: Endless Collapse of Nigeria's National Grid," 1 February 2026. https://allafrica.com/stories/202602020399.html
Davis, J. (2026) ‘The growing PUE advantage of larger data centers’, Uptime Intelligence, 6 August. https://intelligence.uptimeinstitute.com/resource/growing-pue-advantage-larger-data-centers?
Huawei Digital Power (2025), "What Is a PUE Data Center? A Complete Guide to Power Usage Effectiveness," citing global average and hyperscale-leader PUE figures. https://digitalpower.huawei.com/en/blogs/pue-data-center
International Energy Agency (2025), Energy and AI, Executive Summary. https://www.iea.org/reports/energy-and-ai/executive-summary
International Energy Agency (2026), Key Questions on Energy and AI, Executive Summary. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
Kretzmann, S. (2026). Giant data centres get the first green light from Cape Town tribunal. Bizcommunity. https://www.bizcommunity.com/article/giant-data-centres-get-the-first-green-light-from-cape-town-tribunal-397138a
Mining Weekly / Engineering News (2026), "African Power Utilities Join AEW 2026 as Grid Expansion and Electrification Take Center Stage," 5 June 2026. https://www.miningweekly.com/article/african-power-utilities-join-aew-2026-as-grid-expansion-and-electrification-take-center-stage-2026-06-05
