Hitachi’s New GPU Cold Plate Aims to Cut AI Data Centre Cooling Power by 30%

Hitachi has developed a cold plate for cooling GPUs that it says could cut the electricity needed to cool next-generation AI data centres by 30%, the company announced on October 21. The prototype is built around what Hitachi calls a U-shaped tunnel flow path, which sends coolant straight toward the hottest part of the chip instead of along its surface.

In conventional cold plates, coolant runs across the face of the GPU and warms as it goes, so the far end of the chip receives warmer liquid. Hitachi’s design arranges thousands of fine channels in parallel through the thickness of the plate. Coolant travels to the area near the heat source, absorbs heat, then turns back toward the outlet, so cold and warmed liquid do not mix. Because each channel is short, friction stays low, which keeps pressure loss, and therefore pump power, in check.

In measurements at a coolant flow of 4 L/min, the prototype reached a thermal resistance of 0.011 kelvin per watt with a pressure loss of 27.6 kPa, giving better cooling than the conventional structure for the same pumping power, the company said. That matters because the industry is increasingly interested in warm-water cooling at around 45°C. If GPUs can be held below their target temperature with warmer coolant, more of their heat can be released straight to outside air instead of through power-hungry chillers.

What it means in practice

The 30% figure is an estimate rather than a measured result. Hitachi calculated it from changes in annual chiller operating rates using outdoor temperatures for representative cities such as Tokyo, and notes that results vary with conditions. The company plans to combine the plate with pumps, chillers and piping under integrated control through its HMAX Data Center digital service, and will present part of the work at InterPACK2026, the ASME conference in San Diego from October 26 to 29.

Power is fast becoming the constraint on Japan’s AI build-out. Projects such as NTT’s proposed Anthropic core data centre in Osaka and Sharp’s domestic orders for Foxconn-built AI servers point to ever denser GPU racks, while Hitachi itself is also moving into physical AI for factory floors with Fanuc.

What to watch: whether the plate moves from prototype to product, how its cost compares with existing cold plates, and whether operators can confirm the 30% savings in real facilities outside the climate assumptions Hitachi used.