Atlas is a tightly integrated system, not a collection of parts. Every subsystem inside the container exists because a conventional alternative fails somewhere we intend to operate. This post walks through the three decisions people ask about most: the cooling, the power, and the radio.
Why immersion cooling, and why we submerged the network too
Air cooling is the first thing to fail at the edge. Fans pull in dust, salt fog, and humidity. Filters clog. Above roughly 35 C ambient, conventional racks start throttling, and the noise signature of a fully loaded GPU server is impossible to hide.
Atlas uses single-phase dielectric immersion: all compute and all production networking equipment sit fully submerged in an engineered dielectric fluid. Heat rejects through an external radiator sized for 50 C ambient operation. The consequences stack up:
- No fans. The platform is near-silent, which matters tactically and matters again in a hospital or an office.
- No particulate intrusion. Dust storms, salt spray, and 100 percent humidity are non-events for submerged electronics.
- Thermal stability. The fluid mass smooths thermal cycling, which roughly doubles expected silicon service life and eliminates thermal throttling.
- A wider envelope. Atlas is designed for continuous duty from -40 C to +50 C ambient. That covers the desert, the Arctic, and the equipment room with broken HVAC.
Submerging the production switches, not just the servers, was a deliberate choice. A data center fabric that stays dry in a dusty container is the weakest link in the thermal design. The one exception is out-of-band management, which stays dry on purpose: the unit remains remotely serviceable even with the immersion bath drained.
Why the power plant is a hydrogen fuel cell
Generators are the default answer for off-grid power, and they are the wrong one for this mission. They are loud, they emit a thermal plume, they need a steady diesel logistics chain, and they hate altitude and extreme cold.
Atlas runs on a hydrogen PEM fuel cell, 10 kW at the base tier and scalable to 50 kW, with days of continuous off-grid runtime on internal hydrogen storage. Runtime extends with refueling or an external hydrogen skid. The fuel cell is NATO NSPA qualified, which means allied logistics systems already know how to supply it.
The power architecture is layered:
- Primary: hydrogen PEM fuel cell with battery support for transient smoothing.
- Management plane: an independent backup fuel cell keeps monitoring and control alive even while the primary is being serviced.
- Alternates: diesel generator, utility tie-in, and shore power, behind automatic source transfer with UPS ride-through.
If your site eventually gets grid power, Atlas becomes a very well-protected load. Until then, it is its own power plant with a minimal acoustic and thermal signature.
Why every Atlas ships with its own cellular network
Edge AI is only useful if sensors, vehicles, and people can reach it. Plant Wi-Fi is brittle, carrier coverage is absent exactly where Atlas deploys, and tactical users cannot depend on commercial infrastructure at all.
So Atlas carries its own: a private 5G network with O-RAN compliant radios, GPU-accelerated baseband processing, and a 1 to 5 km coverage radius supporting more than a thousand concurrent devices. SIM-based authentication and per-mission network slicing mean the drone feed, the crew handsets, and the sensor mesh ride separate, policy-isolated slices of the same radio.
Running the 5G baseband on GPU is one of our favorite design decisions: the same silicon family that serves model inference also processes the radio waveform, which keeps the part count down and the upgrade path unified.
The part you don't see
The individual technologies are provable on their own. The product is the integration: the power budget that closes at 50 C ambient, the fabric segmentation that keeps five enclaves on one switch, the vibration isolation that lets all of this survive transport, and the compliance mapping (MIL-STD-810H, MIL-STD-461G, DoD STIG and RMF, FIPS 140-3, and sector standards from HIPAA to NERC CIP to IEC 62443) that turns a clever container into something an accreditor will sign.
That integration is the company. The platform page has the full stack, tier by tier, and if you want the deeper technical brief, ask us for the whitepaper.