will have 129 million euros to achieve it

It is the fourteenth supercomputer in the world according to the TOP500 listis in Spain and is called MareNostrum 5. But it is also a “living” system because it continues to evolve and adapt to new times. In fact, it has just received a notable injection of capital for a critical update: the one that will allow it to work with AI model training and inference.

129 million for MareNostrum 5. As explained at the BSC (Barcelona Supercomputing Center), the European High Performance Computing Joint Undertaking (EuroHPC JU) has signed the contract for the acquisition of new hardware elements for MareNostrum 5. It does so with the collaboration of Fsas Technologies (Fujitsu) and Telefónica. The budget is close to 129 million euros and will be co-financed 50% by EuroHPC and the other 50% by Spain, Portugal and Turkey. The update will be installed during the first half of 2026. The project, defined in the documentation of the European Commission, started in July 2025.

MareNostrum 5 adapts. This update will not be a “more of the same”, but is focused entirely on artificial intelligence and machine learning workloads. The project also has another objective: to strengthen European digital sovereignty, which will reduce technological dependencies and prevent unwanted technology transfers outside the EU.

HPC used to rule, now AI will also rule. The expanded architecture of MareNostrum 5 will make use of specialized compute partitions. Thus, there will be at least one for LLM training and another for inference. Today this supercomputer has five different partitions and according to the TOP500 list it has an Rmax power of 175.30 PFLOPS (FP64, Linkpack) or 215.40 PFLOPS according to EuroHPC. Its total power capacity is 20 MW.

30 Exaflops FP4 for AI. In the white paper the goal is to achieve 30 FP4 exaflops, which seems like a huge jump, but it must be clarified that the current 0.17 FP64 exaflops are dedicated to classic High Performance Computing (HPC). These FP4 exaflops are dedicated to training and massive inference, and cannot be easily compared to those FP64 exaflops.

The technical description of the GH72 partition, even without explicitly mentioning NVIDIA, clearly fits your DGX GB200 NVL72.

NVIDIA in the pools. Although the technical documentation does not specify what type of hardware will be used, it does talk about a GH72 partition whose specifications are very similar to the DGX GB200 NVL72 supercomputer. It is not specified how many GPUs, what type of GPUs or how much memory they will have, but there are minimum requirements, such as “180 GB of HBM3e memory or higher.” That rules out the H100 and H200, but models like the B200/GB200 would enter (MareNostrum 5 currently uses those) or the B300/GB300, in addition to others like the AMD Instinct MI350/MI355X, which according to market data could arrive with 256 GB of HBM3e memory.

And soon, own chip. Yesterday, the BSC simultaneously announced good news about its efforts to develop its own chip. The so-called Cinco Ranch TC1 developed by the Barcelona Zettascale Lab has been validated and its experimental implementation it has been a success. We are looking at a chip based on RISC-V architecture and that uses Intel 3 3 nm manufacturing technology. It is a relatively modest chip that can operate at 1.25 GHz, but it is still a notable first step for the BSC to also continue advancing its role in semiconductor design.

European supercomputers. This project is part of EuroHPC’s program to keep European supercomputing moving forward. Currently it has 12 supercomputers distributed throughout Europe, among which Jupiter (Germany, 4th in the world on the TOP500 list) and the future Alice Recoque (France), which will be consolidated as the first two exascale systems on the old continent.

In Xataka | The EU wants to close the gap in the race for AI with 750 million euros. And it is good news for Barcelona