When you have hundreds of thousands of AI accelerators in your datacenters, like Microsoft, problems easily build up, especially with cooling. Today, Microsoft showcased its custom cooling plates, which utilize microfluidics to carry liquid coolant into microscopic channels etched directly into the back of silicon chips, promising to pull heat away far more effectively than today's standard cold-plate systems. In internal tests, the approach removed heat up to three times more efficiently than conventional cold plates and cut the peak temperature rise inside a GPU by roughly two-thirds. The company sees these results as a path to allow data centers to squeeze more performance from the same hardware while using less energy for cooling.

Instead of relying on a metal cold plate that sits above a packaged chip, Microsoft's design carves tiny microchannels into the silicon itself so coolant can flow much closer to the actual heat sources, the actual transistor network. The channels are hair-thin and follow a unique pattern inspired by nature, resembling leaf veins more than straight pipes. Microsoft used AI to direct fluid toward hotspots and map each chip's thermal fingerprint, improving efficiency over standardized channel layouts.