Last week, we received our two NVIDIA GB300 NVL72 racks. Now live in Stockholm, this is one of the first GB300 NVL72 deployments in Europe, adding 144 Blackwell Ultra GPUs to our research infrastructure.

Each liquid-cooled rack combines 72 GPUs and 36 NVIDIA Grace CPUs. Together, the two racks provide 74 TB of combined CPU and GPU memory and a theoretical peak of 2.16e18 FLOP/s of dense FP4 compute. With this cluster, we will be able to train models in the 30B-parameter range from scratch.

The new racks join our six DGX B200 nodes, bringing our owned GPU count from 48 to 192. For a small, independent research group, this is a substantial expansion of the experiments we can take on.

Our focus remains on understanding and improving reasoning in language models, including reinforcement learning with verifiable rewards. The additional compute gives us more room to run experiments at scale and develop our infrastructure for pre-training and post-training. We look forward to putting it to work.

Thank you to DELTA for installing the racks, and to T.Loop for hosting them in Stockholm. We appreciate the work both teams put into getting the racks in place and ready for research.