In the first quarter of 2027, Serbia will receive a supercomputer equipped with 640 Nvidia GraceHopper chips, which will be twenty times more powerful than the one launched in April and worth 50 million euros.
It will be delivered by the French company Bull SAS in cooperation with Mistral AI, while a dedicated facility with water-cooling systems is already being built at the State Data Centre in Kragujevac. A far more important question than the purchase itself is how many domestic companies and researchers actually know how to turn this kind of infrastructure into products, research and competitive business advantages.
What the GraceHopper chip actually does compared to the current infrastructure
The difference between the existing system in Kragujevac and the new infrastructure is not only in speed, but also in the type of tasks the system can support. GraceHopper is Nvidia’s chip designed primarily for training large AI models and running scientific simulations. Unlike a standard computer that processes tasks one by one, this chip can handle millions of operations simultaneously, meaning it can complete in minutes tasks that would take a conventional computer days.
The first National Artificial Intelligence Platform was based on four Nvidia DGX A100 systems with a total of 32 GPU units and performance of five petaflops for AI processing. During 2025, the Republic of Serbia upgraded this system with six additional Nvidia DGX H200 systems featuring 48 GPU units and 32 petaflops of performance, representing an approximately sevenfold increase in computing power.
The third supercomputer planned for 2027 will be a system with 640 Nvidia GraceHopper superchips intended for training large language models, advanced simulations and processing massive datasets in real time.
In practical terms, the current system allows smaller teams and researchers to train specialised AI models and experiment with machine learning. The new system enters the category of infrastructure used by states and the world’s largest technology companies for developing their own AI models, from medicine and autonomous systems to national language models and security simulations.
A reference example comes from Britain. King’s College London and Guy’s and St Thomas’ Hospital use the Nvidia supercomputer Cambridge-1 to train models on tens of thousands of MRI brain scans, with the aim of detecting dementia, stroke and brain cancer earlier. Sebastien Ourselin, head of the School of Biomedical Engineering at King’s College, described the impact in a single sentence: “For the first time, this kind of computing power is becoming available to medical researchers and it will be truly transformative for patient health.” The same type of infrastructure, using the same Nvidia chips, is coming to Kragujevac.
Its application is not limited to medicine. Legal systems trained on court practice analyse contracts and assess litigation risk within seconds. Banks use the same capacities for credit scoring and real-time fraud detection. Energy companies optimise grid loads, while transport systems predict traffic congestion. For Serbian companies, this means infrastructure will be available to sectors that previously had no access to such computing capacity.
Who can make use of the new infrastructure
The National Artificial Intelligence Platform in Kragujevac is available free of charge to Serbian universities, institutes and start-ups based in science and technology parks. The platform is already being used by 80 companies, together with professors from technical faculties. The third supercomputer will add further capacity to the existing system and will also be available to businesses, with 2.5 petabytes of data storage space, thirty times more than at present.
Serbia currently has 13 data centres, while the European Union has more than 3,000. Slovenia and Croatia, as EU member states, have access to the Horizon Europe programme, which allocates 485 million euros annually for research projects in artificial intelligence. Serbia holds associate member status and can participate in such projects as a partner, but not as a coordinator, reducing its chances of securing funding. Bosnia and Herzegovina, North Macedonia and Albania currently do not even have this limited access, meaning researchers from those countries must seek EU partners or pay market prices for foreign cloud systems.
The new supercomputer is also intended to support the development of a Serbian language model (Large Language Model or LLM) based on Mistral’s system, adapted to the Cyrillic alphabet and the specialised vocabulary of medicine, law and education. If the project succeeds, Serbian companies would for the first time have a model that understands the local legal system and medical protocols, rather than only general English-language contexts.
Through this project, Mistral AI gains access to the market for localised AI systems for the public sector and regulated industries, where European competition is still practically non-existent. Serbia, in return, gains a partner developing a European alternative to American AI companies such as OpenAI and Anthropic. For Mistral, this is also a flagship state project in Central and Eastern Europe that could help the company secure future government contracts across Europe.
Over the past two years, countries around the world have been rapidly purchasing AI infrastructure because they believe computing power will become a strategic resource, similar to energy or telecommunications. However, experience shows that hardware without domestic teams, researchers and products often remains underused.
If Serbia succeeds in building domestic AI companies, research teams and products for the regional market around this infrastructure, Kragujevac could become one of the most important data centres in Central and Eastern Europe. If it fails, the greatest benefit of the system may remain limited to political prestige and a few isolated projects.
(Bloomberg Adria, 28.05.2026)
https://rs.bloombergadria.com/tehnologija/digitalizacija/105087/superkompjuter-u-kragujevcu/news
