Swiss AI firm runs sovereign inference tier on four Intel Arc Pro B60 GPUs

Swiss company SOKKAN Inference launched a new 'Swiss' tier this week, serving AI inference from a self-owned machine located in Meyrin, Geneva, with no NVIDIA hardware involved. The setup uses four Intel Arc Pro B60 GPUs purchased for around CHF 2,450 total, chosen primarily for their availability and VRAM-per-franc value during a GPU shortage. Three open-source MoE models run simultaneously across the four cards, enabling data to remain entirely within Switzerland for clients with strict data-residency requirements. Performance benchmarks show strong throughput for the gpt-oss-20b model under vLLM, but the larger 80B model suffers degraded aggregate throughput under concurrency due to llama.cpp lacking continuous batching support. The entire machine, including co-hosted production workloads, costs approximately CHF 750 per year in electricity at Swiss rates.