Ultra-low power
10-30 kW
SambaNova racks operate in an average envelope of 10 kW to 30 kW per rack, drawing up to 4x to 10x less power than traditional high-density GPU racks that routinely exceed 45 kW.
High-Performance AI that
Doesn’t Drain the Grid.
We built the SCX.ai Factory to solve the two biggest constraints in modern computing: Power availability and Water scarcity.
The Challenge: AI’s Physical Footprint
Artificial Intelligence is colliding with real-world constraints. Independent analyses project that data-centre electricity demand will more than double by 2030, with AI as the primary driver. This demand is often concentrated in regions where the power grid is already at capacity.
To cool ultra-dense GPU clusters, many hyperscale facilities rely on evaporative cooling, consuming millions of litres of freshwater annually. As AI models scale—from training to mass inference—the "water cost per prompt" is becoming a critical environmental liability.
The Reality: If AI is to scale responsibly, we must radically shrink the kWh per token and Litres per token.
Projected Demand vs Capacity
Traditional infrastructure cannot support exponential growth without unsustainable resource consumption.
SambaNova RDU architecture
Sustainable AI inference without compromising performance.
SCX Labs has partnered with SambaNova to provide highly sustainable, energy-efficient inference for large-scale AI and autonomous agentic workloads.
Standard air-cooling
No liquid loop
SambaRack and SambaStack platforms run on standard data centre air-cooling, avoiding high-density liquid cooling, immersion cooling, complex plumbing, and large water consumption.
Existing data centres
No overhaul
Low power draw and air-cooled operation allow direct deployment into traditional data centres without expensive structural retrofits or new physical plant construction.
Intelligence per watt
Dataflow
The RDU architecture minimises power-hungry data movement through assembly-line-style dataflow and a three-tier memory design for fast frontier-model inference with lower operational footprint.
Model bundling
Less hardware
Multiple large models, up to trillions of parameters, can remain resident on a single node. Dynamic model switching avoids expensive reloads and reduces hardware bloat in multi-step agentic workflows.
The SCX.ai Approach: Efficiency by Design
We didn't just build a cloud; we engineered an AI Factory designed to maximise useful work per unit of energy.
1. Inference-First Silicon (ASICs)
Our orchestrator routes every workload to the most efficient silicon. Specialised accelerators for LLM inference deliver markedly better performance-per-watt than general-purpose GPUs.
Result: Significantly lower energy cost per token.
2. Air-First Cooling Strategy
We utilise high-quality facilities with moderate rack densities (~12kW/rack). This allows us to stay within standard air-cooled envelopes for the majority of the year, avoiding water-intensive evaporative cooling.
Result: Targeting zero Water Usage Effectiveness (WUE) for much of the year.
3. Grid Pairing & Renewables
We place compute capacity where low-carbon power is abundant or schedulable. By analysing the grid’s carbon intensity, we help customers understand the real-world emissions of their workloads.
Result: Lower grams of CO₂e per token.
4. Algorithmic Efficiency
The greenest energy is the energy you don't use. We right-size context windows, utilise RAG to reduce waste tokens, and apply parameter-efficient fine-tuning (LoRA).
Result: Higher accuracy without power-hungry full model retrains.
SUSTAINABILITY DASHBOARD
REAL-TIME METRICS
Measurement & Transparency
SCX.ai / Sust.
Energy Intensity
0.04 kWh/1M
Water Intensity
0.01 L/1M
Site PUE
1.25
Carbon Intensity
140 gCO₂/kWh
You cannot manage what you do not measure. Traditional clouds hide these metrics; SCX.ai exposes them directly in your dashboard.
PUE (Power Usage Effectiveness): We separate IT energy from facility overhead. Lower PUE means your budget pays for compute, not air conditioning.
WUE (Water Usage Effectiveness): We track cooling water use normalised per unit of IT energy.
Per-Token Intensity: The metric that matters for AI. We expose kWh/1M tokens and L/1M tokens directly in your dashboard alongside latency.
At-a-Glance: SCX.ai vs. Conventional Cloud
| Feature | Conventional GPU Cloud | SCX.ai AI Factory |
|---|---|---|
| Primary Silicon | General-purpose GPU | Next Generation Efficient ASICs |
| Cooling Profile | Water-Intensive (Evaporative) | Air-First cooling |
| Density Strategy | Ultra-dense (Requires complex cooling) | Optimised Density (Uses standard cooling) |
| Metrics Provided | Billable Hours & Storage | kWh/Token, L/Token, gCO₂e/Token |
| Deployment Speed | Years (New build dependent) | Weeks (Deploys in existing Tier-3 sites) |
Frequently Asked Questions
Sources & Standards
ISO/IEC 30134-2: Definition and usage of PUE.
The Green Grid: Standards for Water Usage Effectiveness (WUE).
IEA & Academic Research: Benchmarks for data centre electricity growth and AI water footprints.
Ready to lower your AI footprint?
Don't guess your impact—measure it. Book a consult to run a baseline on your current prompts and see the energy, water, and cost savings of the SCX.ai architecture.