Frore says its diamond-based AI coolers cut temperatures by 22°C for Nvidia Rubin chips rated at 2,300W.
Frore Systems is betting that one of the most expensive materials on Earth can solve one of the data center's most expensive problems. The cooling specialist has unveiled a new direct-to-chip coldplate that embeds diamond wafers directly above GPU hotspots, a design it says can tame the extreme heat of next-generation AI accelerators.
The product, called LiquidJet Diamond, targets processors that push far beyond the thermal limits of conventional silicon. Frore claims the technology handles hotspot densities up to 770 W/cm², reduces GPU temperatures by 16 to 22 °C compared with standard multi-inlet skived coldplates, and improves AI token throughput per watt by 30 to 35 percent. The company plans to demonstrate the system at the OCP Global Summit, running October 12 through 16, where it will showcase cooling for a 2,300-watt Nvidia Rubin accelerator.
Diamond is an unusual choice for data center hardware, but the underlying physics is straightforward. The material's thermal conductivity can exceed 20 W/cm·K, roughly five times that of copper at about 4 W/cm·K. By placing diamond spreaders on the bottom surface of the coldplate, Frore aims to pull heat away from concentrated hotspots before it reaches the liquid-cooling channels above.
The coldplate pairs those diamond spreaders with a three-dimensional, ultra-short-loop, multi-stage jet-channel microstructure. Coolant pathways are customized for the power distribution of a specific processor, an approach Frore previously refined with its LiquidJet Nexus product for Nvidia Blackwell and Rubin-based servers. Unlike Nexus, which integrates multiple coldplates into a larger assembly, LiquidJet Diamond is a bespoke unit for an individual GPU.
Frore's performance figures are based on the company's own simulation models rather than independent benchmarks. The firm says its original LiquidJet coldplate lowered GPU temperatures by approximately 12 °C and delivered a 25 percent gain in AI tokens per watt. Adding the diamond heat spreader contributes another 10 °C reduction in hotspot temperatures, according to Frore, bringing the full solution to its claimed 16 to 22 °C range.
The company also reports a 75 percent improvement in heat-transfer efficiency and more than 150 percent higher maximum hotspot power density compared with conventional coldplates. It has not disclosed how the diamond wafers are manufactured or integrated into the coldplate, nor whether the OCP demonstration involves production GPUs or a thermal test vehicle.
LiquidJet Diamond is engineered with future silicon in mind, including Nvidia's Rubin Ultra and Feynman platforms, AMD's Instinct MI455 and MI500 series, and custom accelerators from hyperscale cloud providers. The original LiquidJet technology managed roughly 600 W/cm², according to Frore, meaning the diamond-enhanced version represents a meaningful step up in thermal capability.
Whether the technology achieves broad adoption remains an open question. Frore has introduced several cooling innovations in recent years, but the company has not indicated how widely LiquidJet Diamond will be deployed. AI accelerators continue to grow more power-hungry with each generation, however, and thermal management has become a critical constraint for data center operators seeking more performance from every watt. If Frore's claims hold up under independent testing, diamond-infused coldplates could become a practical tool rather than an extravagant novelty.