Meluxina never sleeps at its data centre in Bissen. The supercomputer regularly reaches its maximum capacity. However, despite being open to businesses, the bulk of its computing time is still taken up by researchers.
The Ministers for the Economy and Digitalisation, 
In practical terms, a supercomputer enables tasks that are too demanding for a conventional computer to be distributed across numerous processors. Meluxina is used for simulation, modelling, the analysis of large volumes of data and artificial intelligence. The projects presented by LuxProvide include, in particular, weather forecasting, mapping breast cancer susceptibility, research into new materials, molecular simulation, industrial particle separation and satellite position estimation.
Searching: a compulsive habit
Meluxina is owned jointly by Luxembourg and the European joint undertaking EuroHPC. The latter holds a 35% stake and reserves the same proportion of the capacity for European projects selected through calls for proposals. The remaining 65% constitutes the national capacity, currently estimated at around 365,000 ‘node-hours’ per month, comprising both conventional and graphics processors. A ‘node-hour’ corresponds to the use of a computing node for one hour. However, this unit does not provide a uniform measure of either the complexity of the computation carried out or its economic value.
Over the last three years, public research organisations have accounted for around 90% of the national capacity utilised, amounting to approximately 6 million ‘node-hours’. Innovation projects run by private companies account for the remaining 10%. More than 64 start-ups and SMEs have used Meluxina during this period. Of the user companies, 80% are based in Luxembourg and 20% are from other European countries.
However, the government does not distinguish between usage by start-ups, SMEs and large enterprises. Nor does it specify whether government departments have used the system, or for what purposes. The breakdown therefore confirms the central role of public research, but does not yet allow for a precise assessment of the uptake of Meluxina by the Luxembourg economy.
Meluxina AI to be available from 2027
This clarification comes at a time when the country is preparing Meluxina AI, a second supercomputer specifically designed to train, test and deploy artificial intelligence models. Its capabilities are set to become available from 2027 as part of the Luxembourg AI Factory. The new machine will not simply replace Meluxina: it will feature an architecture tailored to the computations required for AI.
The project presented in December 2024 envisaged more than 2,100 accelerators. The contract that was ultimately signed covers 1,008 graphics processing units. Ministers insist that this difference does not constitute a scaling back of the project. The initial figure was based on the performance, technologies and prices available at the time of its preparation. The chips ultimately selected are fewer in number, but more powerful and more expensive.
The company selected following the tender process has committed to supplying a machine 10 times more powerful than Meluxina for training AI models and 40 times more powerful for inference operations, that is, the use of a pre-trained model. The system will be based on 252 nodes and 1,008 Blackwell-generation Nvidia graphics processing units. According to EuroHPC, its peak performance will exceed 20 exaflops for artificial intelligence calculations.
The total announced cost of Meluxina-AI amounts to €112m: €80m for the acquisition and €32m for hosting and operation over five years. Half of this is to be funded by EuroHPC. For hosting and operating costs alone, the Luxembourg government will contribute €16m (excluding VAT), with the European Union covering the remaining €16m.
Pricing still to be determined
Half of Meluxina-AI’s capacity will be reserved for Luxembourg. However, the access rules and pricing have not yet been finalised. In principle, businesses will be required to pay a market price, which will vary depending on the type of processor, the volume and duration of the computations, as well as the level of service and security required.
Start-ups will be able to access resources free of charge through the Initiate programme. Eligible SMEs will be able to have up to 80% of certain costs covered. Expenditure may also be eligible for research and development support schemes where it relates to a qualifying project. Access by public research organisations is governed by their multi-year agreements.
