Want high-performance enterprise AI without sending your data to the public cloud or locking yourself into a proprietary superpod? Open-weight models (DeepSeek, Qwen, Kimi, Mistral, Apertus…) now rival the best. The real question is how to run them on your own premises, fast and predictably.
On the agenda:
🧠Private AI on standard hardware: off-the-shelf GPUs, DGX Spark and an open RoCE Ethernet fabric, from 4 to 32 GPUs
âš¡ Why the network sets the speed of your model: when every microsecond counts, on every token
📊 Measured figures, not promises: up to 3.3× single-node throughput with prefill/decode disaggregation over Ethernet
🇨🇠Data sovereignty: inference, cache and logs under your control, built for Swiss and European requirements