Frore Systems Shows How Advanced Cooling Can Boost NVIDIA Rubin
As AI accelerators become increasingly powerful, thermal management is emerging as one of the most important constraints on data center performance.
At WAIC 2026, Frore Systems founder and CEO Dr. Seshu Madhavapeddy presented the company’s latest thermal strategy in a keynote titled “Chip to Chiller: The Thermal Stack’s Impact on AI Factory Efficiency.” The presentation argued that future AI infrastructure cannot optimize compute independently from cooling, power delivery, and facility-level thermal systems.
Frore Systems develops semiconductor-manufactured thermal solutions built around two core technologies: AirJet solid-state active cooling and LiquidJet high-density liquid cooling. Together, they are designed to address thermal challenges ranging from edge AI devices to hyperscale AI factories.
🌡️ 2026 Marks the “Year 1 of Thermal Management” #
According to Frore Systems, the AI industry is moving beyond a narrow race for individual GPU performance or model parameters.
As accelerator power consumption rises, power delivery, electricity consumption, and cooling infrastructure are becoming equally important design considerations.
Frore describes AI performance as the interaction of three interconnected layers:
- Software Stack — AI frameworks, models, and optimization software.
- Compute Stack — CPUs, GPUs, accelerators, memory, and interconnects.
- Thermal Stack — The technologies responsible for extracting and rejecting the heat generated by computation.
The thermal layer ultimately determines how consistently the compute layer can operate at peak performance.
Frore therefore characterizes 2026 as “Year 1 of Thermal Management,” reflecting the growing importance of thermal engineering in AI infrastructure design.
The company’s strategy spans the complete thermal path, from AirJet solid-state cooling for compact systems to LiquidJet and LiquidJet Nexus technologies designed for high-density servers and AI factories.
⚡ Thermal Optimization Can Increase Tokens per Watt #
Frore Systems’ 2026 thermal white paper presents Tokens/Watt as a critical efficiency metric for AI factories.
Once a GPU has been manufactured, its fundamental silicon efficiency is largely fixed. However, the thermal environment in which that GPU operates can still be optimized.
Frore argues that a sufficiently optimized Thermal Stack can increase overall AI factory Tokens/Watt by more than 30%.
The optimization opportunity extends across three major layers:
Chip-Level Heat Extraction #
GPU package construction has a direct impact on thermal resistance.
An un-lidded package can eliminate thermal resistance associated with the integrated heat spreader and certain interface layers, although it introduces additional mechanical and manufacturing considerations.
The Thermal Interface Material (TIM) between the GPU package and coldplate is equally important. Frore highlights liquid indium combined with gold-plated surfaces as a particularly effective approach for minimizing thermal resistance.
NVIDIA’s Rubin platform uses liquid indium TIM technology as part of its thermal design.
Device-Level Heat Transfer #
At the coldplate level, three variables dominate thermal performance:
- Microchannel geometry.
- Coolant flow paths.
- Coolant flow rate.
High-performance coldplates can use extremely short microchannels measuring less than 1 mm, combined with three-dimensional structures and split-flow, multi-inlet designs.
Manufacturing such structures is difficult with conventional machining techniques. This is where Frore’s semiconductor-style copper wafer manufacturing approach becomes important.
The LiquidJet platform uses wafer-based processing to create complex copper coldplate structures that would be difficult to reproduce using traditional manufacturing methods.
Coolant selection and flow rate also influence efficiency. Deionized water and PG25 coolant can be used depending on system requirements, but increasing flow rate eventually produces diminishing thermal benefits while increasing pump power requirements.
This creates an optimization problem constrained by the capabilities of the Coolant Distribution Unit (CDU) and data center piping infrastructure.
Data Center-Level Heat Rejection #
The final stage is moving heat away from the cooling loop and into the surrounding environment.
Data centers can combine free-cooling technologies such as evaporative cooling towers with mechanical chillers. The optimal combination depends heavily on environmental conditions and system efficiency.
Frore highlights Coefficient of Performance (COP) as a key decision metric for determining when mechanical chilling becomes worthwhile.
The result is a system-level tradeoff: additional cooling electricity must produce enough reduction in GPU temperature and enough additional compute efficiency to justify its energy cost.
🔥 Temperature Directly Influences GPU Efficiency #
One of Frore’s central arguments is simple:
Lower GPU temperatures can translate into higher computational efficiency.
GPU leakage power increases as silicon temperature rises. This creates a feedback loop in which higher temperature increases leakage, which increases power consumption and can produce additional heat.
Frore references the Arrhenius relationship to illustrate this behavior, noting that a roughly 10°C increase in junction temperature can approximately double leakage-related power loss under certain conditions.
Conversely, Frore estimates that reducing GPU junction temperature by 10°C can increase token efficiency by approximately 15%.
The company’s simplified thermal relationship can be expressed as:
$$ T_{\text{j max}} = \text{Coolant Inlet Temperature} + (\text{GPU Power} \times \text{Total Thermal Resistance}) $$
This highlights three major optimization variables: coolant temperature, GPU power, and total thermal resistance.
The thermal resistance itself accumulates across the package, TIM, and coldplate.
🧊 LiquidJet Targets NVIDIA Rubin-Class Thermal Loads #
Frore Systems positions LiquidJet as a solution for the increasingly demanding thermal requirements of next-generation AI accelerators.
According to the company’s testing and modeling:
- 6–12°C lower GPU junction temperature compared with conventional coldplate approaches.
- 10–25% improvement in Tokens/Watt depending on the system configuration.
- Short-loop microchannels below 1 mm for efficient heat transfer.
- Three-dimensional split-flow and multi-inlet structures enabled by wafer-based manufacturing.
The technology is particularly relevant to high-power accelerators such as NVIDIA’s Rubin platform.
For the Rubin Max-P, rated at approximately 2,300W, Frore reports that LiquidJet can maintain the target junction temperature using approximately 2 LPM of coolant flow.
That compares with roughly 3.25 LPM for conventional coldplate technology under the same thermal conditions.
Reducing required flow has another benefit beyond cooling performance: lower flow requirements reduce the burden placed on pumps, piping, and the broader cooling infrastructure.
📉 Better Cooling Can Lower Chiller Energy Requirements #
The implications extend beyond the GPU itself.
Frore reports that its LiquidJet solution can reduce the mechanical chiller break-even COP for a Rubin Max-P configuration from approximately 6.7 to 4.1.
In practical terms, this gives the data center operator greater flexibility when deciding how aggressively to deploy mechanical cooling.
A more thermally efficient coldplate can therefore influence the energy consumption of the entire cooling system rather than simply lowering the temperature of an individual accelerator.
This is why Frore frames thermal engineering as a chip-to-chiller optimization problem rather than an isolated coldplate design exercise.
🏭 AI Factories Are Becoming Thermal Optimization Problems #
The rapid increase in accelerator power density is changing the fundamental economics of AI infrastructure.
As more compute is packed into each rack, the ability to remove heat efficiently becomes increasingly important. Simply adding more cooling capacity can increase facility power consumption and reduce the efficiency gains delivered by faster accelerators.
The better strategy is to optimize every stage of the thermal path:
GPU die → package → TIM → coldplate → coolant loop → CDU → chiller → environment
Each stage introduces thermal resistance or consumes energy, creating another potential optimization point.
Frore’s approach is to address that entire chain rather than treating cooling as a facility-level problem alone.
🚀 The Future of AI Performance May Depend on Cooling #
The AI hardware race is increasingly becoming an energy-efficiency race.
Raw accelerator performance remains critical, but increasingly dense AI systems must operate within strict power and thermal envelopes. Under those conditions, Tokens/Watt can become just as important as peak compute throughput.
Frore Systems’ LiquidJet technology illustrates how semiconductor manufacturing techniques can be applied beyond the processor itself to improve the thermal infrastructure surrounding it.
The company’s core message is straightforward: once silicon efficiency is fixed, thermal engineering becomes one of the remaining levers for extracting more useful AI computation from the same power budget.
As accelerators such as NVIDIA Rubin push toward increasingly high power levels, chip packaging, TIMs, coldplates, coolant distribution, and chillers will increasingly need to be designed as a single system.
The future of AI performance may therefore depend not only on building faster chips, but also on finding better ways to keep those chips cool.