The LiquidJet Nexus coldplate is Frore Systems’ direct-to-chip liquid cooling solution designed for next-generation NVIDIA AI hardware, announced in December 2025 and engineered to deliver measurable throughput gains where data centers need them most. As GPU power envelopes push past 1,950 watts, traditional coldplates are hitting thermal limits—and Frore is betting that a redesigned cooling architecture can squeeze 10% more token generation out of the same silicon.
Key Takeaways
- LiquidJet Nexus delivers 75% higher cooling performance for 1400W NVIDIA Blackwell Ultra, or 7.7°C lower max die temperature
- For 1950W NVIDIA Rubin GPU, cooling performance increases over 50%, reducing max die temperature by 7.5°C
- Coldplate weight reduced by over 50% while retaining full reliability
- Frore claims the design improves GPU performance, training time, token throughput, and total cost of ownership
- Product announced December 29, 2025, with direct-to-chip jetchannel microstructure architecture
Why LiquidJet Nexus Matters for AI Infrastructure
Token throughput is the metric that matters in large language model training. A 10% boost translates directly to faster inference, shorter training windows, and lower operational costs per token generated. Frore’s LiquidJet Nexus coldplate targets this bottleneck by attacking the root problem: thermal throttling. When a GPU die hits temperature limits, clock speeds drop, and token generation suffers. By reducing max die temperature by 7.7°C on Blackwell Ultra, Frore claims the coldplate removes that performance ceiling.
The timing is critical. NVIDIA’s Vera Rubin and Rubin Ultra GPUs are landing in production deployments as training clusters scale. These chips consume 1,950 watts or more per accelerator—nearly 40% more power than the previous generation. Existing coldplate designs were engineered for lower thermal loads, which means they struggle to maintain stable operating temperatures on Rubin-class hardware. LiquidJet Nexus addresses this head-on with a redesigned internal structure.
LiquidJet Nexus Architecture and Performance Claims
Frore first unveiled LiquidJet at the Open Compute Project summit in October 2025, featuring a 3D short-loop jetchannel microstructure. The newer Nexus variant adds multistage cooling architecture and 3D hybrid cell structures—engineering changes that Frore claims increase cooling performance by 75% for Blackwell Ultra or over 50% for Rubin GPUs. On Blackwell Ultra, that translates to a 7.7°C reduction in maximum die temperature; on Rubin, a 7.5°C drop.
The mechanical redesign matters too. Frore reduced the coldplate weight by over 50% while maintaining reliability—a win for data center operators managing rack density and power distribution. Lighter hardware means easier deployment, faster maintenance cycles, and less structural load on GPU mounting hardware.
How LiquidJet Nexus Compares to Existing Solutions
Current direct-to-chip coldplates for Blackwell Ultra are the baseline. Frore’s 75% cooling performance improvement over those designs is substantial, though the company’s own claims lack independent third-party verification. The broader liquid-cooling market for AI data centers includes passive cold-plate designs and immersion-cooling approaches, but LiquidJet Nexus positions itself specifically as a direct-to-chip solution that preserves existing GPU and motherboard form factors. This compatibility matters—data center operators can retrofit LiquidJet into existing racks without redesigning infrastructure.
Frore’s approach targets infrastructure buyers who already committed to liquid cooling but need better thermal performance. If the 10% token generation claim holds in production deployments, the value proposition becomes hard to ignore: faster training, shorter model iterations, and lower power-per-token costs.
Real-World Impact on AI Operations
The promised improvements affect multiple operational metrics simultaneously. Better cooling reduces power draw per GPU, improving power usage effectiveness (PUE) across the data center. Lower water usage follows from more efficient heat transfer. Token throughput gains compound across training clusters—a 10% boost on a 10,000-GPU training run saves weeks of compute time and translates to significant cost savings.
Whether Frore’s claims survive real-world testing remains an open question. Thermal performance in lab conditions often differs from production environments with variable workloads, ambient temperatures, and pump degradation over time. Data center operators will demand independent benchmarks before committing to large-scale deployments.
What Happens Next
Frore announced LiquidJet Nexus in late December 2025, positioning it for Vera Rubin and Rubin Ultra deployments that are ramping throughout 2026. The company framed the product as addressing a critical cooling wall in AI infrastructure—a problem that only intensifies as GPU power envelopes continue climbing. If major cloud providers and AI labs adopt LiquidJet Nexus at scale, Frore shifts from a niche cooling vendor to an essential component in next-generation AI infrastructure.
Does LiquidJet Nexus work with existing data center infrastructure?
Yes. LiquidJet Nexus is designed as a direct-to-chip liquid cooling coldplate compatible with NVIDIA Rubin and Rubin Ultra GPUs in standard liquid-cooled data center setups. It does not require motherboard redesigns or new mounting standards, making retrofit deployments feasible.
What is the 10% token generation boost based on?
Frore claims the 10% improvement results from reduced thermal throttling enabled by the coldplate’s 75% higher cooling performance on Blackwell Ultra and over 50% improvement on Rubin GPUs. Lower die temperatures allow sustained clock speeds during inference and training workloads. The company presented these figures in its product announcement but has not released detailed benchmarking methodology.
Is LiquidJet Nexus available now?
Frore announced the LiquidJet Nexus coldplate on December 29, 2025. The company has not disclosed specific shipping dates, pricing, or regional availability through the sources provided. Data center operators interested in early access should contact Frore directly for deployment timelines.
Frore’s LiquidJet Nexus coldplate arrives at a moment when AI infrastructure is hitting genuine thermal limits. Whether the company’s 10% token generation claim holds up under production scrutiny will determine whether this becomes a standard component in the next wave of AI data centers or remains a niche solution for operators willing to bet on unproven cooling gains.
Edited by the All Things Geek team.
Source: Tom's Hardware
