We are investing in professional hardware suitable for AI workloads with DGX Spark servers and then probably DGX Station from NVIDIA. These systems will offer us:

• The DGX Spark, equipped with the NVIDIA GB10 Grace Blackwell superchip, delivers up to 1 petaflop of AI performance in a compact form factor with 128 GB of unified memory.
• We are closely monitoring the DGX Station and await the release of the GB300 chips, details of how they work, and their dependence on an external GPU.

This infrastructure should enable us to develop and test AI models with up to 200 billion parameters locally, with the possibility of connecting multiple systems to process even larger models.

We have chosen to invest in architectures based on Grace Blackwell superchips, such as the DGX Spark (1 petaflop, 128 GB of unified memory), in order to benefit from unique CPU/GPU convergence and optimise the massive processing of AI models. We favour unified NVIDIA hardware because of the ease of integration offered by CUDA, anticipating a real technological breakthrough that we believe in.

We are not investing time in AMD architectures at this stage because, in order to leverage the NPUs of Ryzen AI Max+, the FastFlowLM backend must be used instead of Ollama, although the most interesting performance still comes from the integrated iGPU for large models.