Nexalus and RENDA Launch Nova8, a Liquid-Cooled AI Server Built for Retrofit
Nexalus and RENDA's Nova8 puts eight Blackwell GPUs in a 4U chassis with self-contained liquid cooling, claiming up to 20% more AI throughput — figures buyers must verify on their own workloads.

Why it matters
- Nexalus and RENDA launched the Nova8 liquid-cooled AI server on November 17.
- Nova8 integrates eight NVIDIA RTX PRO 6000 Blackwell GPUs in a 4U chassis, each running at 600 watts.
- Nexalus claims up to 20% greater AI throughput and 22% better energy efficiency — unverified company figures.
- Internal Llama 3.3 testing reported 55,000 tokens per second and 8.6 tokens per watt.
- The cooling system uses 10 pumps and 16 fans in a redundant configuration, with no building-level piping required.
Nexalus and RENDA launched the Nova8 on November 17, a self-contained liquid-cooled AI server that puts eight NVIDIA RTX PRO 6000 Blackwell GPUs into a standard 19-inch rack without requiring building-level liquid piping. The claim that comes attached is specific: up to 20% greater AI throughput and up to 22% better energy efficiency than comparable air-cooled systems. Those numbers come from the companies, distributed in a release by ICR. They are not independently verified, and the announcement itself frames them as claims rather than guaranteed results for every workload.
The product targets a question that has moved from theoretical to urgent in AI deployment: how much useful performance can an existing facility actually support? Data center operators face a real constraint. They can buy accelerators, but operating those accelerators productively inside the electrical, thermal, networking, and service envelope of an existing site is a separate decision entirely. Nova8 is an attempt to collapse those two decisions into one purchase.
What does the hardware actually contain?
According to the launch announcement, Nova8 packs eight NVIDIA RTX PRO 6000 Blackwell GPUs into a 4U chassis. Each GPU is designed to run at its full 600-watt power level. That is a dense thermal load for a machine meant to slide into racks originally designed for air cooling.
The cooling technology is Nexalus jet-impingement cooling, applied to both the GPUs and the CPU. Instead of relying solely on coolant flowing along channels, the system directs jets of coolant at a plate above the hottest parts of a chip. The partners say they developed the compute and thermal components together, treating cooling as part of the server architecture rather than something bolted on afterward.
For an infrastructure buyer, that co-design claim matters. A server that physically fits an existing rack still has to meet the site's broader requirements. A self-contained cooling loop can simplify one part of that assessment, but it does not eliminate the need for a full facilities review.
What do the performance claims show?
The launch release reports internal Llama 3.3 testing at 55,000 tokens per second and an efficiency result of 8.6 tokens per watt. It also specifies redundant cooling hardware comprising 10 pumps and 16 fans.
Those are concrete numbers, and they give prospective customers concrete questions to ask. But a throughput figure alone does not establish what a production application will deliver. The companies' own framing acknowledges this, positioning the specifications as starting points for investigation rather than substitutes for testing.
Buyers evaluating Nova8 — or any server making similar claims — should ask pointed questions:
- Which model configuration, precision, input length, and output length were used in the benchmark?
- What concurrency level and serving software produced the headline figure?
- Does the reported power measurement cover the whole server, a subset of components, or a broader infrastructure boundary?
That last question is easy to overlook. Comparing unlike measurement boundaries can make an apparently simple efficiency percentage deeply misleading. A 22% efficiency gain measured at one boundary can shrink or vanish when measured at another.
Why latency belongs in the evaluation
The most useful evaluation, per the framing around the launch, would pair throughput with latency targets. An organization serving interactive requests values predictable response time differently than one running large batches overnight. Both could run on identical hardware and reach opposite conclusions about cost and capacity. That is an evaluation framework, not evidence that Nova8 has passed or failed either requirement.
The distinction is becoming central to AI infrastructure purchasing generally. Throughput numbers dominate launch announcements because they are simple. Production economics depend on the harder combination of throughput, latency, power draw, and sustained operation over time.
How can buyers test it before committing?
The companies are offering qualified prospects remote access through a Test Drive program. Global after-sales support runs through UNICOM Engineering. That access channel gives interested organizations a route to investigate the product before making assumptions based solely on announced specifications.
A meaningful trial, on the announced product's own terms, would look like this: reproduce a representative workload, hold software settings constant, and compare results against the operator's existing deployment. Recording power use, latency, throughput, and sustained operation together would help distinguish a genuine infrastructure improvement from a benchmark advantage that never translates into the intended application.
What about serviceability?
Serviceability deserves attention too. Operators evaluating any liquid-cooled system should understand how components get replaced, how faults are detected, which maintenance tasks their own teams can perform, and how support arrangements fit their uptime requirements. Those details carry particular weight when a product is positioned as a retrofit rather than as part of a new facility design.
A rack of pumps and coolant loops in a building whose staff has only ever handled air-cooled servers is an operational change, not just a hardware swap. The 10-pump, 16-fan redundancy scheme is a hedge against single points of failure, but redundancy and serviceability are not the same thing.
Why this launch matters beyond one product
Nova8's proposition is broader than putting more GPUs in a rack. It is an attempt to make liquid-cooled AI compute usable inside infrastructure that organizations already operate. Liquid cooling has historically implied new buildings or major retrofits — raised floors, coolant distribution units, building-level piping. Nova8 tries to package the entire thermal loop inside a 4U chassis.
Whether that approach delivers the advertised advantage will depend on customer workloads and site-level validation. What the launch does settle is the framing: retrofit practicality now sits at the center of the AI server buying decision, and vendors are building products specifically for operators who cannot rebuild their facilities from scratch. The next signal to watch is whether independent customers report throughput numbers from their own deployments that resemble the claimed 20% gain.
Source: unite.ai
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
708 articles
Related articles
- ASUS Ramps Vera Rubin NVL72 Shipments as Rack Race Shifts to Execution
- AMD launches 256-core EPYC 9006 CPUs to fight Intel and Nvidia
- NVIDIA Open-Sources Run:ai Under Apache 2.0
- Nvidia launches open-weight Nemotron models for enterprise agents and AI PCs
- GIGABYTE AI TOP AERO fits 128GB in a roughly one-litre desktop