SambaNova targets air-cooled data centers

This is a distribution advantage, not just a hardware spec. Dense GPU racks increasingly assume data centers can deliver very high rack power and liquid cooling, which rules out many regional facilities that were built for ordinary air cooled servers. SambaNova is selling into that gap with inference systems designed to fit existing power envelopes and cooling setups, so telecoms, regional clouds, and sovereign operators can launch AI capacity without first rebuilding the building around the rack.1 2 3 4

• NVIDIA’s newest rack scale GPU systems are built around liquid cooled designs. NVIDIA documents describe GB200 NVL72 as liquid cooled, and DGX GB200 racks at roughly 120kW each, with larger SuperPOD units reaching 1.2MW. That is a very different facility profile from a normal regional colocation hall.1 2
• SambaNova positions SambaRack for the opposite environment. It says the systems are purpose built for inference in existing air cooled data centers, without a liquid cooling retrofit, a new build, or a power upgrade. That makes the product easier to place in second tier cloud regions where speed to deployment matters more than absolute training density.3 4
• The channel pattern fits this constraint. OVHcloud is adding SambaNova RDUs to AI Endpoints, and the company page ties similar deployments to SCX, Infercom, Argyll, SoftBank, and TEPCO Systems. These partners already have enterprise customers and data center footprints, but not necessarily hyperscaler grade GPU facilities.4 5 6

The next step is a split market. The biggest model training clusters will keep concentrating in newly built, liquid cooled GPU campuses, while more inference capacity spreads through existing regional data centers using lower friction hardware. That gives SambaNova a clear lane wherever local operators need domestic AI supply faster than they can upgrade physical infrastructure.1 2 3 4 6