Aropet Joel
← Back to blogAropet Joel

Karuma's Failure Is Why Africa's First AI Factory Is Being Built in Uganda

Uganda spent $2.2 billion on the Karuma hydropower dam. Since it opened, it has run at roughly 30% of installed capacity, held back by a national grid that cannot move all the power and an export line to South Sudan that is now six years late. In 2026 the pieces of a different plan for that stranded electricity clicked into place. Africa’s first sovereign AI training facility, running on NVIDIA Blackwell GPUs, is scheduled to go live at Karuma in the second half of this year. It is being built there specifically because the dam has hundreds of megawatts nobody else can buy.

That connection is the story the coverage is missing. The Aeonian Project is being framed as a digital-sovereignty statement, and it is one. But the reason it can be built at all is that Uganda broke its own power-export plan first.

The 30% dam

Karuma is a 600 MW plant financed with a $1.4 billion loan from EXIM Bank of China, part of a total project cost of roughly $2.2 billion once transmission and associated works are counted. In its most recent audited year, it produced 808 GWh against a declared available capacity of 2,652 GWh. That is 30%. The Uganda Electricity Generation Company blames three things: not enough domestic demand, a merit order that dispatches cheaper thermal plants and private producers first (the latter earn capacity charges whether they generate or not), and periodic waterweed clogging the intake.

The transmission story is worse. The 400 kV Karuma-to-Olwiyo line was supposed to feed a further 400 kV line running to Juba in South Sudan, opening a real export market. The Juba line has been delayed six years and construction is now pushed to 2028. Meanwhile Uganda’s generation capacity crossed 2,000 MW, well past current grid draw. Karuma sits paying loan interest and capacity charges while its dispatched output is measured in single-digit percentages of its design.

For years this has been reported as a policy failure. In 2026 it starts looking like a resource.

What Synectics and Schneider are actually building

The Aeonian Project, led by Synectics Technologies with Schneider Electric, is a $1.2 billion buildout of 100 MW of AI computing inside the Karuma facility. Phase one is a 15 MW AI training module plus a 10 MW sovereign supercomputer named USIO, both scheduled for the second half of 2026. Five more modules follow, reaching the full 100 MW by 2028. USIO is being built on NVIDIA’s Blackwell platform, with AI infrastructure firm MDCS.AI and Belgium’s Automation NV on the delivery side. European development money from GIZ, HAUS, and the EU Development Fund is in the mix.

Two technical details matter. First, the site is powered by what its promoters call Karuma’s “excess pre-transmission electricity.” In plain terms, the compute takes power directly at the plant, before it has to reach a substation and travel down a grid that cannot carry it anyway. That sidesteps the very bottleneck that stranded the dam. Second, it is cooled by Nile water, a thermal sink big enough, and effectively free enough, to run high-density GPU racks that most existing data centres cannot host. The site also plugs into a 2,500 km fibre optic network connecting to submarine cables in Kenya and Tanzania, giving the compute a route to global users.

Why here and not Kenya, Nigeria, or South Africa

Look at what modern AI training infrastructure actually needs. A single NVIDIA GB200 NVL72 rack pulls about 120 kilowatts under nominal load, with production deployments observed drawing 130 to 132 kW at full load. Facilities being planned in 2026 want 100 to 750 MW per site, delivered as cheap, always-on, low-carbon power that is not already spoken for by an urban grid. According to industry surveys, fewer than 5% of the world’s existing data centres can even handle 50 kW per rack.

Kenya’s hydro is largely dispatched to a demand-constrained grid that still imports peaking power. South Africa has been publicly warning that its AI data-centre plans are hitting water and power limits. Nigeria’s grid instability disqualifies it for training loads that cannot tolerate an outage. Uganda has the opposite problem: several hundred megawatts of dispatchable renewable power with no buyer, sitting at a plant already built, next to a river cold enough to cool the racks.

That is the arbitrage. The Aeonian Project is not primarily a story about African values winning over Silicon Valley. It is a story about a stranded hydro asset matching, almost exactly, the shape of what Blackwell-generation training wants. Roughly 98% of Africa’s data is currently processed outside the continent; a 100 MW facility does not close that gap by itself, but it is the first physical bet in the direction of doing so. The window exists because the Juba export line failed. If the Juba line hits its 2028 date, the arbitrage narrows. If domestic demand catches up first, the arbitrage narrows. That timing is why the phasing starts this year.

The three other layers landing at the same time

The compute story only reaches Ugandans if the rest of the stack catches up. Three other things happened this year that read differently once you put them next to the AI factory.

On 15 July, the Ministry of ICT launched the Updated National Information Security Framework 2026. It is not a headline document. It is an assessment tool with minimum-security controls for government institutions, meant to raise the floor of what the state can be trusted to protect. That floor matters more once national-scale compute is on Ugandan soil.

The National Identification and Registration Authority finished its mass-enrollment programme on 8 February 2026 with over 35 million Ugandans registered under a system that now includes iris recognition. That is a functional national identity layer.

The Bank of Uganda’s Digital Drive Project is rolling out one merchant code that works across both MTN and Airtel, plus a national QR standard any wallet can scan. That is the first serious attempt to give small merchants a rail that does not force customers to choose a telecom.

Compute, identity, cybersecurity policy, payments. Four sovereign infrastructure layers, all pushed in the same window.

The gap

The tell in this pattern is the National Payment Switch. It has been reported as “stuck in limbo amid lobbying claims and competing business interests.” The switch is the piece of infrastructure that would make interoperability real at the settlement layer rather than the merchant-code layer. Its stall matters more now, because a sovereign AI factory without sovereign payment rails is just a hosting arrangement with better marketing. Data still trains foreign models. Compute still serves foreign customers. The value leaks out through the missing middle.

At Kirafiki Pay we spend most of our engineering time on that middle. The lesson from Karuma is that the top and bottom of the Ugandan digital stack are being built by people with billions of dollars and national mandates. The middle (where a boda rider gets paid and a hospital’s revenue lands) is where private builders still have to do the work themselves. Getting that layer right in the next 24 months is what decides whether the AI factory is a Ugandan asset or a Ugandan address.

What to watch

Three specific things over the next 18 months.

First, whether the Aeonian Project’s 15 MW AI module actually energises in H2 2026 or slips into 2027. That decides whether the arbitrage window is real or aspirational.

Second, whether the Karuma-Juba transmission line hits its new 2028 date. If it slips again, the case for domestic AI absorption of Karuma’s output strengthens sharply, and expect more phases to accelerate.

Third, whether the Bank of Uganda unblocks the National Payment Switch before the end of 2026. If it doesn’t, everything above it in the stack is worth less than the press releases suggest.

The Karuma story used to be about wasted money. It is quietly becoming a story about how a country turned a transmission failure into a compute strategy. That is a good story. It only stays a good story if the rest of the stack ships.

Sources

aiugandainfrastructuredata-centersaeoniankarumasovereignty