Moving Beyond the Lab
The AI infrastructure conversation has a familiar shape: power, density, cooling, cost, time to deploy. Everyone is facing the same constraints, but few have moved from talking about them to resolving them, in production, at scale, with real workloads.
That’s what this is about.
For more than two years, Denvr and UNICOM Engineering have been collaboratively innovating to design AI compute, fine-tuned for single-phase dielectric immersion environments. This isn’t a pilot or a lab experiment; it’s live Neocloud infrastructure.
The result: a thermally-, density-, and efficiency-optimized server platform that is designed and fully warrantied for Denvr’s operational environment. Boasting an unprecedented 14 servers per rack and 2x physical footprint reduction, the server demonstrates that thinking outside the box and collaborative innovation can deliver meaningful value for data center operators.
Since April 2024, the platform has been submerged in fluid, subjected to stress-testing at temperatures up to 60 °C, and simultaneous multi-component NVQual testing. We’ve published a deep-dive technical whitepaper on the methodology and outcomes, but this post is the shorter version – the "here's why it matters and what we learned" read.
The Heat Wall
The biggest limit on GPU density isn't power – it's heat.
You might have the electricity and the floor space, but if you can't strip the heat away, you're stuck. Operators are currently forced to either leave racks half-empty or accept thermal throttling, leaving expensive compute on the table. Five major constraints are converging:
- Density: Increasing server power and thermals worsen footprint in air or require challenging retrofits.
- Thermal Fatigue: Dynamic AI workloads cause rapid temperature swings that wear out hardware.
- Sustainability: Community and environmental impact of alternatives like evaporative cooling are severe.
- Location Limits: Traditional HVAC isn’t practical at edge locations or stranded power sites.
- Timing Gaps: Adaptability and speed to deploy have become a primary challenge for AI data centers.
The Way Over It
Liquid immersion changes this.
Denvr’s Modular Data Center (MDC) platform is what that change looks like in practice, purpose-built around immersion cooling from the ground up: modular, rapidly deployable, zero-water. The result is a 10x density improvement: 1 MW per 600 square feet versus 6,400 for air cooling. Lower costs, smaller footprints, and the ability to put high-density AI wherever power exists, including locations where conventional data center infrastructure simply isn’t viable.
That change shows up in three ways that matter operationally:
- Thermal Stability: Immersion suppresses the rapid junction temperature swings that AI workloads create, reducing wear and tear on the server components, eliminating thermal throttling under sustained load and delivering warranty confidence air cooling can’t match at these densities.
- Power Efficiency: A single-loop cooling architecture eliminates the ~40% overhead conventional cooling consumes. Optimization of server design recovers approximately 800W per server under load in normal usage conditions
- Signal Integrity: NVQual system-level qualification confirms that UNICOM Engineering’s H200 HGX node — a purpose-built immersion-ready redesign of the Dell PowerEdge XE-9680 platform — meets all H100 SXM signal integrity requirements. A hard-won engineering outcome, not a default hardware trait.
You Can’t Build It Alone
Here’s what wouldn’t have been done alone: everything.
UNICOM Engineering worked at the component and server level: purpose-built hardware engineering for an immersion environment, not an air-cooled adaptation. Denvr worked at the server and production level: validating UNICOM results and executing cluster-scale evaluation under sustained AI training workloads. Put those two together in a single program, and production-scale deployment becomes achievable, not theoretical.
That meant rebuilding the hardware from the ground up for immersion, not retrofitting an air-cooled platform. For the H200 HGX node, that involved:
- Form Factor: Reduced from 6U to 3OU, allowing 14 servers per tank, up from just 2-4 in air-cooled racks.
- Thermal Hardware: Immersion-optimized copper heat sinks with fins designed for fluid instead of air.
- Power Architecture: Inspired by OCP open rack configurations, adaptable to ORV2.2 and ORV3 designs.
- Firmware: Developed specifically for immersion environments and operational efficiency.
The impact of fluid on performance, reliability, and compatibility required the same rigor and the right partners to do it properly. Long-term interactions between dielectric fluid and components like polymers, elastomers, and PCB laminates aren’t often disclosed publicly, and we wanted a long-term practical benchmark to complement short-duration testing. A parallel workstream was conducted between Denvr and its fluid providers to evaluate the fluid quality and compatibility profile of critical materials in the operational environment, starting in July 2023. The work also drew on a broader foundation. Working within the OCP ecosystem gave both organizations a common technical language: Open Rack specifications, Immersion Cooling Environments requirements, and Signal Integrity guidance. Where those specifications hadn’t yet caught up, the 48V-54V bus bar departure being the clearest example, we documented the gap and the reasoning.
What Comes Next
The immersion architecture we've validated isn't anywhere near its thermal limits. This approach scales directly into next-generation platforms and reaches locations where conventional data centers simply can't go.
Deployments like this one are also how OCP specifications get stress-tested against reality: where they held, we built on them; where they hadn’t yet caught up, we documented why. The OCP Immersion Solutions, ITE, and Rack & Power workstreams are where that work continues and the best place to engage, contribute, and shape what comes next.
Immersion at scale is no longer a question of whether — the next chapter is already being written, and we’d welcome you in that conversation.
Amy Short is Executive Officer, Advanced Data Center Technologies at Denvr, and Chair of the OCP Immersion Solutions Technical Committee. Austin Hipes is Chief Technologist and VP of Engineering at UNICOM Engineering and Chair of the OCP Immersion IT Equipment Technical Committee. The full technical whitepaper, “Deploying Immersion-Native ITE in NeoClouds: Design, Validation, and Deployment at Scale,” is available at denvr.com and unicomengineering.com.