CZAR
The sovereign AI sandbox.
Train, fine-tune, and run bespoke models on your most sensitive data without it ever leaving the building. One sealed enclosure holds the compute, the data, and the resulting model — with no dependency on a network you do not own.
FIRST LOOK
The unit itself.
Rev A hardware in the fixed-site configuration. Every interface lives on one face; the rest of the enclosure is solid, machined, and closed — because what happens inside is a training run on data that cannot leave the room.
- DC IN
- Locking circular power connector with a captive dust cap — one feed, no external supply logic.
- 5 GbE
- Copper uplink for enclave LAN access, governed by default-deny egress.
- 4× 25 GbE SFP28
- Fiber cages sized for the traffic fine-tuning actually generates: bulk dataset ingest and checkpoint I/O at line rate.
- SERVICE PANEL
- Tamper-evident window, sealed at final assembly. Field and extreme tiers ship without serviceable openings.
- PASSIVE COOLING
- The full-width fin stack is the thermal system: fanless conduction, no moving parts, no filters — sustained through multi-hour training runs, not just inference bursts.
OVERVIEW
Your data, your model, one enclosure.
The Czar is the training-and-fine-tuning tier of the Element 31 hardware line — one of three sealed appliances, alongside Forge for coding inference and Chassis for bespoke resale deployment. It exists for the work that public model providers and cloud fine-tuning endpoints structurally cannot be trusted with: training or fine-tuning a model directly on proprietary, regulated, or classified data. Base weights, your training corpus, checkpoints, and the resulting fine-tuned model all live inside the enclosure and stay there — nothing is uploaded to a vendor’s training pipeline to produce the model you end up with.
That constraint drives every decision in the design. Compute is selected for sustained training and fine-tuning throughput within a power envelope a facility can actually supply, not just for short-burst inference. Memory and storage are sized with headroom for dataset staging, checkpoints, and multiple model variants in flight at once. Cooling is conduction-based, so the sealed configuration needs no openings. Storage is encrypted, the software image is signed and read-only, and integrity is verified continuously rather than once at install. What arrives is what runs, and what you train stays yours.
SPECIFICATIONS
What is in the box.
- Compute
- NVIDIA Jetson AGX Thor (T5000)
- GPU architecture
- NVIDIA Blackwell
- Unified memory
- 128 GB — headroom for training and multi-checkpoint workflows
- Storage
- Up to 8 TB NVMe, encrypted — dataset staging plus checkpoint history
- Power draw
- Under 200 W sustained (training load profile)
- Networking
- Multi-gigabit; air-gap capable
- Cooling
- Fanless conduction (sealed configuration)
- Enclosure
- Machined aluminum, sealed, tamper-evident
- Environmental
- Ruggedized; tiered for indoor, field, and extreme deployment
- Origin
- US integration, sealing, software load, and final assembly
- Software
- E31 Substrate — preloaded and sealed, with training/fine-tuning pipeline
Specifications subject to configuration. Detailed engineering documentation available under NDA.
CONFIGURATION TIERS
Three builds, one architecture.
The software, the sealing discipline, and the API are identical across tiers. What changes is what the enclosure is built to survive while it trains.
Fixed Site
- ENVIRONMENT
- Conditioned indoor space — data center, secure facility, or operations room.
- SEALING
- Sealed enclosure with tamper-evident closure, on standard facility power.
- USE CASE
- Long-running fine-tuning jobs in an accredited room, with the appliance staying put for the duration.
Field
- ENVIRONMENT
- Mobile and forward locations with wide temperature swings, vibration, and dust.
- SEALING
- Sealed and fanless, with no service openings once deployed.
- USE CASE
- Fine-tuning on data that cannot transit a network at all — collected, trained, and consumed at the same forward location.
Extreme
- ENVIRONMENT
- Uncontrolled conditions at the limits of temperature, shock, and contamination.
- SEALING
- Fully sealed and hardened; the enclosure is not opened after final assembly.
- USE CASE
- Model development in austere and contested environments with no maintenance reachback and no tolerance for data exposure.
SEALED BY DESIGN
Integrity is a property of the build, not a policy.
Sealed enclosure
The chassis closes once, at final assembly. The sealed configuration has no field-serviceable openings — your training data and the model it produces never have a physical exit.
Tamper-evident construction
Physical interference with the enclosure leaves permanent, inspectable evidence. Nothing depends on trusting the shipping path between your data and the trained model.
Signed, read-only software image
The training and fine-tuning pipeline is signed and mounted read-only. Nothing is installed, patched, or altered in place at runtime.
Continuous integrity verification
The appliance re-checks its own image and tamper indicators while running. Deviations are written to an append-only record, including during multi-hour training jobs.
DEPLOYMENT POSTURES
How it sits on your network.
ENCLAVE
On-premise, no internet
The appliance sits on your internal network with no route to the public internet. Training jobs are submitted and monitored the way you'd reach any other internal system.
TACTICAL
Disconnected, syncs on reachback
The appliance trains and fine-tunes fully disconnected for the duration of a mission. When a trusted link is available, base weights, datasets, and resulting model updates move in one controlled exchange.
FULL AIR-GAP
Offline provisioning only
No network path exists in either direction. Training data goes in and trained models come out only as signed bundles on physical media.
RELATED HARDWARE
Two other builds, same sealing discipline.
The Czar is one of three purpose-built Element 31 appliances. If your workload is narrower than full model training, one of these may be the better fit.
FORGE
The air-gapped coding copilot
A sealed appliance tuned for low-latency code-model inference and fast local repo retrieval — for teams whose source code can't reach a public LLM.
CHASSIS
Bespoke hardware, built to resell
Custom-made, built-to-order hardware manufactured to a specific customer deal — for orgs that resell the appliance downstream under their own brand.