NVIDIA's next-gen Rubin Ultra AI GPUs will reportedly consume quite an insane 2300W of power or so, with rumors swirling that NVIDIA is switching to a totally new cooling solution for its bleeding-edge AI chips.
In a new post from @QQ_Timmy on X, we're hearing that NVIDIA is contacting cooling solution partners to integrate "direct-to-chip" cooling through microchannel cold plates for its beefed-up Rubin Ultra AI GPUs. This would be quite the move from conventional liquid cooling solutions, and keep the up to 2300W of thermals in check to hit those performance numbers that NVIDIA wants.
The market previously rumored that cooling for Rubin AI GPUs with a 2300W TDP would use a microchannel cover plate (MCL) but @QQ_Timmy's post suggests that there has been high mass production difficulty for these cover plates. Additionally, the leaker has learned that NVIDIA has reportedly requested Asia Vital Components to design a microchannel cold plate (MCCP) for Rubin Ultra (launching in 2027).
It looks like there are multiple cooling solutions in development right now, with the liquid cooling plate market size expected to continue to grow over the coming years, especially when there are 2300W+ AI GPUs on the market.
Recently, we reported that AMD's next-gen Instinct MI450X AI accelerator "forced" NVIDIA to increase the TDP of its VR200 Rubin AI GPU. Rubin had an 1800W TGP just a few months ago, and then it was upgraded to 2300W. Instinct MI450X was a 2300W TGP design, and to increase competitiveness, AMD boosted it to 2500W. The world of super-power-hungry AI GPUs is here, and it's only going to get even more power-hungry.



