Who makes gpu baseboards & compute trays for AI data centers

The compute tray sits between the packaged silicon (layer 1, layer 2 memory) and the rack-scale system (4.01): it's the board that physically carries the GPUs, any paired CPUs, VRMs (4.08), connectors (4.07) and the traces that bring scale-up fabric (3.01) to each chip. Nvidia's HGX and Bianca designs and the OAM/UBB open standard define the mechanical and electrical envelope; the ODMs build to that spec while Supermicro builds boards in-house rather than sourcing from the ODM tier. AI demand is the entire reason this part category exists in its current form — a baseboard carrying multiple high-power GPUs, coherent CPU links and dense high-speed interconnect is a different design problem from a general-purpose server motherboard, with far tighter power delivery, signal integrity and thermal requirements. For investors, the baseboard is where board complexity, layer count (4.05) and connector density all stack up, so it's a useful proxy for how much of a rack's bill of materials sits in high-spec board manufacturing versus commodity components; watch which board builders get first-wave qualification on each new GPU generation, since that tends to set the pecking order for the whole rack cycle.