Four Universities Are Pooling GPU Clusters to Give Researchers Shared Access
A federated scheduler will let researchers at four institutions queue jobs across roughly 900 pooled accelerators, most of which sat idle for two-thirds of the year.
Universities in Accra, Ibadan, Nairobi, and Stellenbosch have agreed to link their research clusters under a shared scheduler, creating a pool of about 900 accelerators that any researcher at a member institution can queue against.
The motivation is embarrassingly simple: utilization. An internal survey circulated among the four institutions found average cluster occupancy between 28% and 41%, with long idle stretches during teaching terms and multi-week backlogs during grant deadlines. Pooling smooths both, since the peaks do not arrive at the same time in Ibadan and Stellenbosch.
"We each bought hardware to survive our own worst week," said Prof. Kofi Adjei, who chairs the consortium's technical committee. "Collectively we bought four worst weeks and used them about a third of the time. This is not a fundraising story. This is arithmetic."
The consortium has budgeted $2.4 million over three years for the scheduler, the interconnect upgrades, and a small operations team, funded jointly by the institutions and a continental research foundation. Two further universities have asked to join. If the model holds, it offers a template for research compute that does not depend on any one government deciding to build a national facility.
Why the offline-first bet is reshaping African AI · The Build Loop
When you purchase through links in our articles, we may earn a small commission. This doesn't affect our editorial independence.
Kwame Asante
Enterprise Correspondent
Kwame Asante writes about enterprise technology, procurement, and the vendors selling to Africa's largest institutions. Before joining Afrikons he covered corporate litigation in Accra.
View BioLoading the next article