According to leaker @QQ_Timmy, the company is working with specialist cooling partners to adopt a “direct-to-chip” design using microchannel cold plates. This marks a sharp shift from its usual liquid-cooling systems and is intended to squeeze out every possible bit of performance per watt from the Rubin Ultra lineup.

For those not in the know, microchannel cover plates (MCCP) work a lot like the direct-die cooling used by hardcore PC enthusiasts. A copper cold plate filled with microscopic channels allows coolant to flow directly around the GPU die, improving convection and reducing thermal resistance. The result is far more efficient heat transfer straight from the chip to the fluid, offering tighter control over thermals and performance.

The move shows how far Nvidia is pushing the limits of its AI architecture. With Rubin expected to draw significantly more power than the already demanding Blackwell series, the company has been forced to think creatively to avoid performance throttling in large rack-scale systems. The days when slapping on a bigger fan could solve the problem are long gone.

Industry sources suggest Nvidia has approached Taiwan’s Asia Vital Components to design the new cooling solution. The microchannel system was originally planned for the base Rubin series, but Nvidia’s aggressive production timeline appears to have forced it to accelerate adoption for Rubin Ultra instead.

The company’s scramble reflects the broader shift in the tech world toward more extreme cooling methods. Microsoft recently revealed its own “microfluidic cooling” approach, where coolant runs either inside or on the backside of the silicon. While technically complex, such designs are quickly becoming necessary as chips consume more power than ever before.

If Nvidia can pull this off, Rubin Ultra could become its most efficient AI platform yet, but without that next-generation cooling, it risks becoming the world’s most expensive space heater.