One training server, up to ten inference servers, one liquid loop. Delivered racked, plumbed and benchmarked, with one partner to call.
Teams typically run 10 to 20 inference servers for every training server. The rack is sized for that ratio, so the model you fine-tune at night is the model your users talk to in the morning.
Nova 8 Training adapts open models to your data on site.
Push the new model to the inference servers in the same rack, on the same network.
Ten Nova 8 Inference servers answer users at up to 550,000 tokens a second.
Start in a standard air-cooled rack. Connect the liquid loop and the heat leaves as hot water instead of hot air, ready for a simple dry cooler or for heating buildings.
Heat is mixed into room air, chilled back out, then evaporated away: three stages of plant after the server.
60 °C water is hotter than outside air on the hottest day, so a dry cooler rejects it all year. No compressors, no chilled water.
In a sealed rack the room no longer needs air conditioning for the servers.
Closed loop with dry coolers. Nothing lost to cooling towers.
We size the training-to-inference ratio, power and cooling with you.
Servers, loop and quick-connects installed and pressure-tested before shipping.
Every server is benchmarked and the results are handed over with the rack.
One support line for the servers and the cooling, through the Renda network.
Book a two-hour remote session in a fresh, isolated environment. See your own tokens per second before you commit.