How the Global Grant Architecture Traps African Machine Learning in a Performative Loop

African artificial intelligence is celebrated more loudly than it is capitalised. In April 2025, fifty-four signatories in Kigali endorsed the Africa Declaration on Artificial Intelligence. They announced a US$60 billion Africa AI Fund, a headline figure that remains largely undefined more than a year later. Over the same period, African AI startups captured roughly 0.02 to 0.03% of global AI venture funding [3][4], the continent held under 1% of global data-centre capacity [5], and sub-Saharan Africa produced under 1% of the world's computer-science AI publications. The gap between declaratory ambition and executional capacity is not incidental. An incentive structure manufactures it, made up of donor key performance indicators, short grant cycles, authorship conventions, and career ladders that systematically reward convening and reporting over engineering. This Viewpoint argues that the binding constraint on African machine learning is not talent, and not solely compute, but a funding architecture that finances demonstrations rather than durable technical capability. It proposes four concrete reforms: fund compute as infrastructure rather than as a project line item; make maintained code a fundable deliverable; count first authorship and software artefacts in evaluation and promotion; and replace three-month "pilotitis" grants with programmes built to run for years, not quarters.