Last Updated 1 hour ago by Kenya Engineer
The global AI boom is usually counted in three units: money, chips and gigawatts. None of them, on its own, shows how much computing is actually available.
That problem was brought into focus by a Guardian investigation published on 17 August. Citing internal documents, the newspaper reported that Microsoft had about 2.2 million installed AI chips. It compared that figure with estimates derived from Microsoft’s reported data-centre expansion and concluded that the company’s operational chip estate appeared smaller than some analysts expected.
Microsoft rejected the calculations as inaccurate and said the analysis relied on incorrect assumptions. The company does not publicly report volumes of specific AI chips, and the internal documents are not independently available. The reported 2.2 million figure should therefore not be treated as a verified inventory, and the investigation does not establish that Microsoft has failed to obtain chips.
It does establish a more useful question: what does a technology company mean when it says it has added data-centre or AI capacity?
An announced gigawatt is not an operational gigawatt
A large data-centre campus passes through several different capacity states. A developer may control land, reserve utility capacity, sign a connection agreement, begin civil works, complete a weather-tight shell, install mechanical and electrical systems, energise a substation, fit server racks, connect fibre, load software and finally pass commissioning tests.
Each milestone is real, but they are not equivalent. A gigawatt in a development pipeline may refer to a future maximum build-out. A gigawatt secured under an energy agreement may not yet have transmission infrastructure. A completed building may not have the switchgear, cooling or network systems needed to accept computing equipment. An energised hall may still be only partly fitted out or lightly utilised.
This explains why corporate statements, planning applications, electricity-connection queues and actual energy consumption can describe the same market very differently. They are often measuring different points in the delivery chain.
Why converting megawatts into chips is uncertain
The Guardian attempted to compare reported capacity with a broad estimate of how many graphics processing units that power could support. Such calculations can be valuable as an order-of-magnitude check, but they cannot produce an exact inventory.
First, the power figure may represent total facility capacity, critical IT load, contracted supply or ultimate campus build-out. Second, a data centre uses electricity for cooling, pumps, power conversion, lighting and other systems. Power Usage Effectiveness estimates this overhead, but the result varies with climate, design and utilisation.
Third, GPU generations have different power requirements and are installed in servers with CPUs, memory, networking and storage. A cloud fleet also performs conventional workloads and may include Microsoft’s own silicon alongside Nvidia, AMD and Intel hardware. Redundancy and design headroom further separate nameplate capacity from average consumption.
Finally, installation is not the same as productive use. Chips may be present but awaiting networking or commissioning; capacity may be reserved for reliability; and utilisation changes by hour and workload. Any single conversion from gigawatts to GPUs therefore depends on definitions that companies rarely publish.
The physical bottleneck is better supported than the chip estimate
Although the exact inventory remains uncertain, the underlying infrastructure constraint is well documented. Microsoft chief executive Satya Nadella has said that the immediate problem can be finding powered buildings close enough to available electricity, leaving equipment without a suitable place to be installed.
Microsoft’s 2025 annual report said the company operated more than 400 data centres across 70 regions and added more than two gigawatts of capacity during that year. In its first-quarter FY2026 earnings call, it said total AI capacity would increase by more than 80 percent during the financial year. These statements show extraordinary expansion, but they still do not identify how much capacity was fully energised, fitted with accelerators and available for customer workloads at each reporting date.
The International Energy Agency reached a similar physical conclusion at industry scale. It reported that data-centre electricity use rose 17 percent in 2025 and projected that total consumption would roughly double by 2030, with AI-focused data-centre demand tripling. It also identified grid connections, transformers, gas turbines, advanced chips and planning approvals as tightening constraints.
The limiting component can change during a project. A chip shortage may dominate one quarter, followed by transformer lead times, utility connection works, cooling equipment, skilled commissioning teams or delays in local approval. AI capacity is a system output, not a warehouse count.
A better capacity ledger
Electricity planners and the public need a capacity ladder rather than one headline number. At a minimum, major projects should distinguish ultimate planned capacity; utility-approved connection capacity; capacity under construction; capacity commissioned and energised; operational critical IT load; and average or peak consumption over a stated period.
Computing disclosure can remain commercially sensitive while still being useful. Operators could report accelerator capacity in broad classes, the share installed and commissioned, and whether figures refer to physical chips, server systems or equivalent computing performance. They should state the reference date and avoid combining assets at different delivery stages.
Efficiency figures require the same discipline. PUE and Water Usage Effectiveness should be reported alongside utilisation and local climate, because a lightly loaded new facility can produce misleading ratios. Renewable-energy claims should distinguish annual contractual matching from electricity available at the site in every operating hour.
Construction reporting should include the energisation date, first customer workload, initial operational phase and ultimate build-out. These milestones would allow investment markets, grid operators and communities to see whether announced capacity is becoming usable infrastructure.
The lesson for Kenya is immediate
Kenya is attracting larger data-centre projects and framing AI as a national development priority. Kenya Engineer has reported a planned 44-megawatt Nxtra facility at Tatu City, while Konza Technopolis is operationalising a Tier III-certified national data centre. These are important assets, but the sector will need consistent definitions as projects multiply.
A promoter should not be able to describe land acquisition, a reserved power connection and a commissioned computing hall with the same word: capacity. Nor should national AI progress be measured by processors ordered without disclosing when they will be installed, powered and made accessible to researchers, start-ups or public institutions.
Kenya’s draft AI and Emerging Technologies Policy already recognises energy and water use in data centres as part of a full sustainability value chain. The final framework could turn that principle into a reporting standard for large computing projects. NEMA, energy regulators, utilities, county planners and the ICT ministry would then work from the same project-stage definitions.
This matters more in a smaller power system. A large speculative connection request can distort network planning, while a delayed substation or transmission project can strand expensive equipment. Transparent phases and financial security for major connection reservations would help utilities separate credible demand from proposals that may never be built.
Measure the system that delivers the computation
The Microsoft investigation may never yield a definitive public GPU count. That is not a reason to ignore it. It is a reminder that the AI economy is advancing faster than the language used to describe its physical assets.
Investors need to know whether capital expenditure has created productive capacity. Utilities need to know when load will arrive. Communities need to understand electricity, water, noise and backup-generation impacts. Governments need to know whether national computing ambitions exist in presentations or in commissioned equipment.
The most credible AI metric will therefore not be the largest number of chips or gigawatts announced. It will be the amount of reliable, efficient and accessible compute that is actually energised, commissioned and serving useful work.
-
Product on sale
AWEBook Volume 2Original price was: KShs 3,500.00.KShs 3,000.00Current price is: KShs 3,000.00.
