GenEye AI Labs — White Paper
A self-contained, road-transportable 96×B200 GPU data centre with deployable solar, liquid cooling, and battery storage—designed to convert stranded electricity into AI compute anywhere on Earth.
The global AI compute market is constrained not by GPU supply alone, but by the years-long timelines required to build traditional data centres, secure grid connections, and provision cooling at scale. Grid connection delays routinely stretch 3–7 years. Meanwhile, stranded electricity—at hydroelectric dams, renewable installations, mining operations, and remote industrial sites—sits unused because there is no data centre nearby to consume it.
GenEye AI Labs proposes a fundamentally different approach: manufacture a complete, road-transportable AI data centre inside a 53-foot semi-trailer, drive it to any site with available power and fibre connectivity, deploy solar wings, and begin selling GPU compute within days—not years.
Our first product, the GENEYE-MDC-1, packs 96 NVIDIA B200 Blackwell GPUs (17.3 TB HBM3e), direct-to-chip liquid cooling, 500 kWh LFP battery storage, and 175 kW deployable solar into a single 53-ft trailer with a total CAPEX of $10–12M. At 80% utilization and $5.50/GPU-hour, a single trailer generates approximately $3.7M in annual gross compute revenue against ~$540K in operating costs, yielding a simple payback period of approximately 3.6 years.
The AI infrastructure industry is undergoing a phase transition. Anthropic has committed a reported $35B to Lambda and $45B to Nscale for dedicated AI compute1,2, validating that long-term contracted GPU capacity is the foundation of the next computing era. Yet the fundamental bottleneck has shifted: AI data centres are increasingly a power procurement problem, not merely a GPU procurement problem.
GenEye's mobile data centre inverts the traditional model:
Traditional model: Build $100M+ permanent data centre → wait years for grid connection → deploy GPUs.
GenEye model: Manufacture GPU trailer → drive to site with available power → connect → compute in days.
Across the globe, gigawatts of electricity generation capacity sits underutilised—hydroelectric dams in rural Canada, solar farms in the American Southwest, wind installations in Scandinavia, geothermal plants in Iceland, and industrial generation at mining and oil & gas sites. These locations have abundant, cheap electricity ($0.02–$0.06/kWh) but no data centre infrastructure.
Our proposition is direct: find unused electricity, deploy a trailer, connect fibre, and convert kilowatt-hours into GPU-hours. We do not build data centres. We manufacture mobile AI compute infrastructure that follows the energy.
The GENEYE-MDC-1 is a complete, self-contained AI data centre manufactured inside a reinforced 53-foot semi-trailer. Every subsystem—compute, networking, storage, cooling, power, solar, battery, fire suppression, and physical security—is integrated at the factory and road-tested before delivery.
| Subsystem | Specification | Notes |
|---|---|---|
| Trailer | 53 ft × 8.5 ft × 9.2 ft interior | Reinforced, insulated, vibration-isolated |
| Compute | 12 × DGX B200 (8-GPU ea.) | 96 × B200, 17.3 TB HBM3e |
| GPU Power | ~172 kW peak | 14.3 kW per 8-GPU system |
| Total Facility Load | ~213 kW typ. / 250 kW engineered | Includes cooling, networking, overhead |
| Networking | 400/800G InfiniBand | NVIDIA Quantum switches |
| Storage | 500 TB – 1 PB NVMe | High-speed local tier |
| Cooling | Direct-to-chip liquid + dry cooler | 250 kW heat rejection, N+1 pumps |
| Solar | 175 kW DC deployable | Folding wing arrays + roof panels |
| Battery | 500 kWh LFP | Solar smoothing, UPS ride-through |
| Backup Power | 250 kW grid/generator interface | Automatic transfer switch |
| Connectivity | Fibre primary + Starlink backup | Dual-path redundancy |
| Fire Suppression | Clean-agent + battery-specific | Dual-zone detection |
| Security | Biometric access, cameras, sensors | 24/7 remote NOC monitoring |
The trailer is divided into three primary zones: Power & Battery (front), Compute Zone A (centre), and Compute Zone B with Networking & Cooling (rear). The following isometric diagrams show the complete system at deployment scale.
The compute subsystem is built around 12 NVIDIA DGX B200 systems, each containing 8 B200 Blackwell GPUs. This provides a total of 96 B200 GPUs with 17.3 TB of HBM3e memory—sufficient for large-scale AI training, fine-tuning, and high-throughput inference workloads.
The B200 represents the current state of the art in AI accelerator design. Each GPU delivers 180 GB of HBM3e at 8 TB/s bandwidth, with support for FP4/FP8 precision that dramatically accelerates inference workloads. The DGX B200 system interconnects 8 GPUs via NVLink 5.0 at 1.8 TB/s bisection bandwidth, enabling the kind of all-reduce operations critical to distributed training.
We selected the DGX form factor specifically because it provides a factory-integrated, tested compute node at a known power envelope (14.3 kW maximum input). This eliminates custom integration risk and provides NVIDIA's full software stack, including Base Command and DGX OS.
The trailer is designed for GPU replacement cycles. As next-generation accelerators become available (B300, Rubin, and beyond), the rack infrastructure, cooling, power, and networking can accommodate drop-in upgrades. The liquid cooling manifold uses standard quick-disconnect fittings, and the power distribution supports up to 300 kW peak, providing substantial headroom for denser future GPU nodes.
| Parameter | Per System | Total (12 Systems) |
|---|---|---|
| GPUs | 8 × B200 | 96 × B200 |
| HBM3e Memory | 1.44 TB | 17.3 TB |
| GPU Interconnect | NVLink 5.0, 1.8 TB/s | — |
| Max Input Power | 14.3 kW | 171.6 kW |
| Weight | 313.9 lb | ~3,767 lb |
| Form Factor | 10U | 120U (+ switches, cabling) |
| Network Fabric | 400/800G IB per GPU | InfiniBand Quantum |
| Load Category | Power (kW) |
|---|---|
| GPU Compute (12 × DGX B200) | 172 |
| InfiniBand / Ethernet Switches | 8 |
| NVMe Storage | 4 |
| CDU Pumps | 5 |
| Dry Cooler Fans | 12 |
| UPS Conversion Losses | 5 |
| Lighting / Security / Controls | 2 |
| Miscellaneous | 5 |
| Maximum Facility Load | ~213 |
The system is engineered for 250 kW continuous / 300 kW peak, providing headroom for transient loads and future GPU upgrades.
The GENEYE-MDC-1 operates as a solar-assisted hybrid system, not a fully off-grid installation. This is an honest engineering distinction: 175 kW of solar at ~25% capacity factor produces approximately 43.75 kW average—roughly 20% of the total facility load. The system is designed to maximise renewable utilisation while maintaining reliable operation through grid or generator backup.
Daytime: Solar feeds GPUs directly; excess charges battery.
Night / overcast: Battery provides partial load + grid/generator supplements.
Low battery: Automatic transfer to grid/generator via ATS.
A customer requiring complete off-grid capability would need approximately 800 kW–1 MW of solar and 2–5 MWh of battery storage. This exceeds a single trailer but is achievable with a dedicated solar field adjacent to the trailer—a viable configuration for remote industrial sites where grid access genuinely does not exist.
At its ~213 kW typical facility load, the system must reject approximately 727,000 BTU/hour of heat (213 kW × 3,412 BTU/hr per kW). Air cooling is not viable at this density. The GENEYE-MDC-1 uses direct-to-chip liquid cooling throughout.
Primary loop: Propylene glycol/water circulates through cold plates mounted directly on each B200 GPU die, absorbing heat at the source. This coolant flows through rack-level manifolds to the CDU.
Secondary loop: The CDU transfers heat through a liquid-to-liquid heat exchanger to the external dry cooler, which rejects heat to ambient air via forced convection.
Redundancy: Dual CDUs (one primary, one standby) with N+1 pumps and fans. The system continues operating with any single pump or fan failure.
| Component | Specification |
|---|---|
| Heat Rejection Capacity | 250 kW (headroom over 213 kW load) |
| CDU Count | 2 (primary + redundant) |
| Cold Plate Inlet Temp | ~25°C |
| Hot Return Temp | ~45°C |
| Flow Rate | ~200 L/min per CDU |
| Dry Cooler Position | Roof/rear-mounted, external |
| Coolant | Propylene glycol / deionised water |
| Vibration Isolation | Flexible hoses + quick-disconnect at every joint |
A GENEYE-MDC-1 unit goes from arrival to live AI workloads in a matter of days, not months or years.
The following represents planning-grade estimates. Actual procurement pricing, particularly for B200 systems, varies materially by volume, configuration, and supply agreement. GPU servers dominate the economics.
| Subsystem | Key Components | Budget |
|---|---|---|
| B200 Compute | 12× DGX B200, rails, spare PSUs, fans, warranty | $5,600,000 |
| Networking | IB switches, NICs, transceivers, cabling, routers | $750,000 |
| Storage | 500 TB–1 PB NVMe, storage servers, RAID | $350,000 |
| Liquid Cooling | 2–3 CDUs, pumps, heat exchangers, cold plates, dry coolers | $650,000 |
| Electrical | 300 kW transformer, switchgear, PDUs, UPS, ATS, busbars | $800,000 |
| Battery | 500 kWh LFP modules, BMS, PCS/inverter, HVAC | $425,000 |
| Solar | 175 kW PV, folding structures, inverters, deployment mechanisms | $450,000 |
| Trailer | Reinforced 53-ft, raised floor, insulation, racks, vibration isolation | $425,000 |
| Fire / Security | Detection, clean-agent suppression, cameras, access control | $175,000 |
| Software | GPU monitoring, Kubernetes, Slurm, billing, remote management | $300,000 |
| Engineering / Cert. | Electrical, mechanical, thermal, structural, transport compliance | $600,000 |
| Spares / Contingency | ~7% contingency reserve | $750,000 |
| Total CAPEX | ~$11,275,000 |
Working investment number: $10–12M per 96-B200 trailer. GPU servers represent approximately 50% of total CAPEX. As GPU pricing evolves, total unit cost will track accordingly.
Assumes deployment at a site with $0.08/kWh electricity, 70% average GPU utilisation, and centralised remote operations (no full-time on-site staff). The headline revenue scenario in §12 models 80% utilisation; the additional draw adds roughly $1,250/month ($15K/year) to electricity, which does not materially change payback.
| Category | Monthly | Notes |
|---|---|---|
| Electricity | $8,640 | 150 kW avg × 720 hrs × $0.08/kWh |
| Compute Maintenance Reserve | $10,000 | GPU, PSU, fan, memory, NVMe failures |
| Cooling Maintenance | $1,800 | Coolant, filters, pumps, dry cooler |
| Electrical Maintenance | $1,400 | UPS, PDU, switchgear, battery inspect. |
| Trailer Maintenance | $650 | Tires, brakes, suspension, structural |
| Solar Maintenance | $750 | Panel cleaning, inverter, deployment hardware |
| Network / Software | $2,750 | Fibre, monitoring, cloud mgmt, security |
| Insurance | $4,000 | Equipment, liability, transport |
| Labour (Centralised) | $12,500 | Infra eng., SRE, NOC, field tech, admin |
| Miscellaneous | $2,500 | Travel, consumables, permits |
| Total Monthly OPEX | ~$44,990 |
| Electricity Rate | Monthly Cost | Annual Impact |
|---|---|---|
| $0.04/kWh | $4,320 | Saves ~$51K/yr vs. baseline |
| $0.06/kWh | $6,480 | Saves ~$25K/yr vs. baseline |
| $0.08/kWh (baseline) | $8,640 | — |
| $0.10/kWh | $10,800 | Adds ~$26K/yr vs. baseline |
| $0.15/kWh | $16,200 | Adds ~$91K/yr vs. baseline |
Site selection at $0.04–$0.06/kWh electricity provides meaningful cost advantage but does not fundamentally change the unit economics. The dominant variable is GPU utilisation, not electricity.
Current public B200 pricing provides the benchmark: Lambda lists B200 instances at ~$6.69–$6.99/GPU-hour, while CoreWeave’s equivalent HGX B200 rate is ~$8.60/GPU-hour3,4 (rate cards as of 8 October 2026). We model three price points that position below established providers while maintaining strong unit economics.
Maximum annual GPU-hours: 96 GPUs × 8,760 hours = 840,960 GPU-hours.
| Utilisation | $4.00/GPU-hr | $5.50/GPU-hr | $7.00/GPU-hr |
|---|---|---|---|
| 50% | $1.68M | $2.31M | $2.94M |
| 70% | $2.35M | $3.24M | $4.12M |
| 80% | $2.69M | $3.70M | $4.71M |
| 90% | $3.03M | $4.16M | $5.30M |
$5.50/GPU-hour, 80% utilisation:
Customer controls workloads via API. Standard GPU cloud model at a discount to hyperscalers.
Customer leases entire 96-GPU trailer. Private, single-tenant infrastructure on their site.
GPU + Kubernetes + storage + model deployment + monitoring + SLA. Turnkey AI infrastructure.
The killer proposition is not "rent our GPUs." It is: "We deploy a private 96-B200 AI data centre at your site in weeks, not years. You supply land + electricity + network. We supply everything else."
AI Startups & Scale-ups — LLM companies, video/voice AI, robotics, computer vision, generative media, and AI agent platforms that need hundreds of GPUs but lack the capital or timeline to build their own infrastructure. A mobile trailer provides dedicated capacity without the $100M+ data-centre commitment.
Pharmaceutical & Biotech — AI drug discovery requires massive compute for protein modelling, molecular simulation, and generative chemistry. Lambda publicly cites Iambic Therapeutics using HGX B200 clusters for AI drug discovery. These customers require private, on-premise infrastructure where their data never leaves their environment.
Government & Defence — Military bases, intelligence agencies, emergency response, and government AI laboratories. The value is not cheap GPU-hours; it is: "Bring the compute to us." Security-cleared, air-gapped configurations are possible.
Universities & Research Labs — Researchers with time-limited grants who need burst compute. A 12-month trailer deployment is far faster than a university procurement cycle for permanent infrastructure.
Remote Industrial — This is where the mobile form factor creates genuinely unique value. Oil & gas (seismic analysis), mining (geological modelling), energy (grid optimisation), manufacturing (digital twins), and agriculture (satellite/drone AI) all have electricity and fibre at remote sites but no data centres. That is the sweet spot.
Renewable Energy Operators — Solar farms, wind installations, and hydroelectric facilities with excess generation capacity. We convert their stranded kilowatt-hours into GPU-hours, creating a new revenue stream from existing assets.
A customer with: 250+ kW available electricity, fibre connectivity, physical land, and no data centre—willing to sign a 24–36 month contract at $250K+/month. We deploy. They compute. We do not build the first trailer on speculation. We secure the anchor customer first, then build against the contract.
Competition falls into three tiers: GPU neoclouds (most customers would compare us against these), mobile/modular infrastructure providers, and hyperscalers.
| Company | Position | B200 Pricing | Threat Level |
|---|---|---|---|
| CoreWeave | Premium GPU cloud, massive scale | ~$8.60/GPU-hr equiv. | |
| Lambda | GPU cloud / clusters, strong brand | ~$6.69–6.99/GPU-hr | |
| Nscale | $45B Anthropic contract, massive infra | Contract-based | |
| Crusoe | Energy + AI infra (closest thesis) | Contract-based | |
| Nebius | Rapid AI infra expansion | ~$7.15/GPU-hr | |
| Fluidstack | Dedicated AI compute | Varies |
| Company | Relevance |
|---|---|
| Antimatter | Distributed/mobile AI infrastructure thesis. Closest emerging competitor to our concept. Watch closely. |
| Vertiv | Trailer-mounted modular data centres. Infrastructure supplier, not GPU operator. |
| Schneider Electric | Modular DC solutions for remote locations. Infrastructure supplier. |
| Exyte | Mobile high-density DC with immersion cooling. Demonstrates concept viability. |
The trailer itself is not the moat. Any well-funded competitor can put B200s in a truck. The moat is built from six interlocking advantages:
| Risk | Severity | Mitigation |
|---|---|---|
| Low GPU utilisation | Critical | Anchor customer signed before build. 24–36 month contracts. |
| GPU pricing compression | High | Multi-product revenue (cloud, dedicated, managed). Cost structure allows price flexibility. |
| NVIDIA supply constraints | High | Early procurement agreements. Design supports multi-gen GPU drop-in. |
| Vibration damage during transport | Medium | Vibration-isolated racks, quick-disconnect cooling, factory road testing. |
| Thermal failures in extreme climates | Medium | 250 kW heat rejection with N+1 redundancy. Desert/arctic environmental ratings. |
| Regulatory / transport compliance | Medium | DOT-compliant dimensions. Battery transport certification. Insurance. |
| Permanent DC built at same site | Medium | Mobility: we relocate. Contract terms protect minimum revenue. |
| Competitive neoclouds at lower price | High | Our value is location flexibility + privacy, not lowest $/GPU-hr. |
Each additional trailer is a capital-efficient unit of AI compute capacity. The business scales linearly:
| Fleet Size | Total GPUs | Facility Power | CAPEX | Annual Revenue (80% util, $5.50/hr) |
|---|---|---|---|---|
| 1 trailer | 96 | ~250 kW | $11M | $3.7M |
| 5 trailers | 480 | ~1.25 MW | $50M | $18.5M |
| 10 trailers | 960 | ~2.5 MW | $95M | $37M |
| 50 trailers | 4,800 | ~12.5 MW | $425M | $185M |
| 100 trailers | 9,600 | ~25 MW | $800M | $370M |
At 50+ trailers, procurement volume drives unit CAPEX below $9M. At 100 trailers, GenEye would operate a distributed 25 MW, ~10,000-GPU AI infrastructure network—comparable to a mid-sized neocloud, but deployable anywhere electricity and fibre exist.
The company GenEye is building is not "mobile data centres." It is mobile AI power-to-compute infrastructure—a platform that finds unused electricity, deploys compute in days, and converts kilowatt-hours into GPU-hours at scale.
GenEye AI Labs is seeking strategic investment to fund the design, manufacturing, and deployment of the first GENEYE-MDC-1 unit, alongside anchor customer acquisition and operational team buildout.
For investment enquiries, partnership discussions, or anchor customer conversations:
Raj
CEO, GenEye AI Labs
raj@geneye.ai
Third-party figures cited in this paper are drawn from public reporting and published provider rate cards. All figures are as of 8 October 2026 and are subject to change; GPU rate cards in particular move frequently.
GenEye AI Labs is not affiliated with, endorsed by, or partnered with any company named above. They are referenced solely as public market benchmarks.