Finland has offered South African researchers access to Lumi, one of Europe's most powerful supercomputers, as part of the two countries' efforts to turn years of goodwill into an official artificial intelligence partnership.
This offer was made during a dialogue between South Africa and Finland on AI issues, which took place at the Finnish Embassy at the National Research Foundation (NRF) in Pretoria on Thursday. Despite discussions about friendship, a formal agreement on AI research does not yet exist. According to Daan du Toit speaking to TechCentral, universities in both countries are already negotiating, but 'there is no formal cooperation agreement yet. This dialogue is intended to lay the foundation for it.'
Daan du Toit, Deputy CEO for International Cooperation and Resources in the Department of Science, Technology and Innovation, stated that the government is prepared to allocate financial resources. He expressed a desire to involve Finland in the Presidential PhD Program, initiated by President Cyril Ramaphosa in 2023 with initial funding of 1 billion rand to support doctoral studies both domestically and abroad. He also noted Finland's contribution in the 1990s and 2000s to the creation of the former Meraka Institute at CSIR, which, he said, became a platform for CSIR's AI opportunities.
Damien Le Carpentier from CSC – IT Centre for Science, a company owned by the Finnish state and universities that manages Lumi, detailed the opportunities Finland can provide. CSC will host South African personnel and experts for short periods, organize training and exchange for early-career students and researchers, and assist scientists and companies in accessing Lumi or Finnish AI expertise. Furthermore, joint applications for EU funding are open through the EuroHPC Joint Undertaking and the main EU research funding program, Horizon Europe.
According to Lumi's own guidelines, South African researchers cannot apply for Lumi usage independently; the principal investigator of any project must be based in the EU or a country associated with its research programs. The main route for participation is through partnership with European groups.
Lumi is located in a former paper mill in Kaarina, Eastern Finland. It is an HPE Cray EX system, much of whose power is provided by Graphics Processing Units (GPUs). Its GPU section contains 2,978 nodes, each combining an AMD 64-core processor with four AMD Instinct MI250X GPUs, totaling 11,912 GPUs. A separate section provides over 262,000 cores of AMD Epyc processors for traditional tasks. The total storage capacity is about 120 PB, distributed across flash, disk, and object systems, all connected by a 200 Gb/s HPE Slingshot network, according to published Lumi specifications.
When running the Linpack benchmark, used to rank the world's fastest computers, Lumi demonstrates a performance of 379.7 petaflops against a theoretical peak of 531.5 petaflops, consuming about 7.1 MW of energy. In 2023, this made it the third fastest device in the world; as of June 2026, new systems have lowered this ranking to 11th. Nevertheless, it is approximately 95 times more powerful than the 4 petaflop machine installed by CHPC to replace Lengau. Lumi is co-funded by the EuroHPC Joint Undertaking and a consortium of 11 European countries led by Finland.
Le Carpentier noted that the data center operates on 100% renewable energy, and waste heat from the computers is used to heat the city. He stated that Finland could only afford a machine of this scale by pooling funds with other European states. He emphasized: 'Everyone gets a share in a machine they could not afford alone.'
The next step is the creation of the Lumi AI Factory, part of a European scheme aimed at developing skills, data, and services around the large supercomputer. According to information from the consortium member Cyfronet in Poland, the new system will cost over 600 million euros, with funds split equally between the European Commission and participating countries.
Le Carpentier warned that acquiring the hardware is only the easy part. Supercomputers are used for about five years, so funding must be continuously renewed, and each new generation costs more. He added: 'An ecosystem around it is necessary. Without it, your machine is practically useless.'
The computational gap at home
South Africa's own capabilities are significantly smaller. Jabu Mtsweni, executive manager of the cluster for NextGen Enterprises and Institutions at CSIR, reported that the High-Performance Computing Centre (CHPC) in Rosebank, Cape Town, possesses GPU clusters—chips used for training and running AI models—though 'limitedly.'
The CHPC's Lengau supercomputer was listed among the 500 fastest machines in the world at position 121 in 2016 but has since dropped out. A replacement with 4 petaflops capacity, capable of performing about four quadrillion calculations per second, is expected to be installed by March, and Mtsweni expressed hope that this will allow South Africa to return to the list.
He stated: 'The computational gap still exists. Sovereignty is not an event, but a journey.' Mtsweni, who is on the panel reviewing national AI policy, noted that South Africa is seen as a leading AI user in Africa but lags in policy development. He reported that the policy project 'hit a roadblock' after Minister of Communications Solly Malatsi withdrew it upon discovering it cited fictitious sources. Mtsweni concluded: 'AI adoption is outpacing policy.'