Frore Systems Says Diamond Coldplate Can Lift AI Factory Output 35%

What LiquidJet Diamond adds

LiquidJet Diamond builds on the earlier Frore Systems LiquidJet coldplate. That design uses a 3D ultra short-loop multi-stage structure that can be customized to the power map of each GPU, including extreme hotspots. According to Frore, LiquidJet lowers GPU die temperature by 12 °C, which it says translates into 25% higher Tokens/Watt efficiency.

The Diamond version places an additional spreader with integrated diamond wafers on the bottom of the coldplate, positioned at the hotspot locations. Frore says this brings a further 10 °C reduction in hotspot die temperature and lifts the Tokens/Watt gain to 35%. The company states the coldplate can cool extreme hotspots of up to 770 W/cm².

Compared with a skived coldplate, Frore lists the following results:

• Over 150% higher max power density at the hotspot
• 75% higher heat transfer efficiency
• GPUs that run 16 to 22 °C cooler
• 30 to 35% more AI Tokens/Watt
• A seamless drop-in upgrade

The power constraint behind the pitch

Frore frames the product around energy scarcity. It says demand for AI Tokens is pushing hyperscalers to find more efficient ways to run AI Factories, while power is finite and expensive to expand. The company names utility interconnection, transformer capacity and physical power delivery infrastructure as limits on total power draw, and notes that local communities are increasingly concerned about them.

Founder and CEO Seshu Madhavapeddy argued that more efficient GPUs make each token cheaper to produce. "When each GPU is more efficient, each AI Token costs less to produce, and more AI Tokens can be generated within the available power budget, using the same AI Factory infrastructure," he said. He added that a 35% gain would let hyperscalers reach growth goals faster with fewer AI Factories, noting that each new one costs tens of billions of dollars to build.

Manufacturing scale and larger chips

Frore also positions LiquidJet as a manufacturing answer. It says traditional skiving cannot deliver the thermal performance that current AI GPUs demand, and that newer approaches such as ECAM (electrochemical additive manufacturing, or 3D metal printing) offer design flexibility but cannot meet hyperscaler quality, reliability, volume or timeline needs. LiquidJet, according to the company, combines the precision of semiconductor fabrication with high-volume manufacturing.

Die size is another factor. Frore points to a range from the 1,488 mm² NVIDIA Blackwell Ultra to the 46,225 mm² Cerebras Wafer-Scale Engine, roughly 30 times the footprint. The company says it has pioneered the only coldplate manufacturing process able to support both the power and the size requirements of such larger AI GPUs.

Demos at the OCP Global Summit

At the 2026 OCP Global Summit, Frore is showing live demos of LiquidJet and LiquidJet Diamond cooling a 2300W NVIDIA Rubin and a 770 W/cm² extreme hotspot density. It is also presenting LiquidJet Nexus, an integrated coldplate system for ½U compute trays that, per the company, eliminates hoses, connectors and manifolds. The booth is F4 at the San Jose McEnery Convention Center, October 12 to 16, 2026.

For data center operators, the pitch is that cooling gains can be turned into more output from existing power and infrastructure. The figures are Frore's own claims, and the announcement does not provide pricing or availability details.