Who makes direct-to-chip cold plates for AI data centers

The cold plate is the first thermal interface in the loop: coolant enters through a quick disconnect, flows across channels machined or etched into the plate above the silicon, and exits carrying the heat toward the rack manifold. Every accelerator, CPU, and switch die on a liquid-cooled tray gets its own plate, so the part is qualified against a specific chip package and reference design rather than sold as a generic component. That qualification tie is the main structural feature investors should track: a plate design that wins a reference socket tends to stay for the life of that platform, and switching vendors mid-cycle means re-validating thermal performance against the die. AI GPUs run hotter and denser than prior server chips, which is why liquid cold plates displaced air heatsinks as the default rather than an option, and why two-phase and microfluidic variants are being developed for the next step up in power density. Some rack integrators build cold plates in-house for their own designs, which narrows the market available to independent cold-plate makers even as unit volume grows. Watch for supplier concentration at the top GPU platforms and how much of the volume in-house integrators keep for themselves versus outsource.