Compute AI — Building Blocks

L8 — Application User Interface & API Gateway BOTH REST · WebSocket · gRPC

L7 — Model Foundation Model / Fine-Tune EDGE-VIABLE 7B → 405B params

L6 — Runtime Inference Engine EDGE-VIABLE vLLM · llama.cpp · TensorRT

L5 — Compute GPU / NPU / CPU Fabric CRITICAL FP16 · BF16 · INT8 · INT4

L4 — Memory HBM / VRAM / DRAM Hierarchy BOTTLENECK HBM3 · 3.35TB/s

L3 — Storage Model Weights & KV Cache EDGE-VIABLE NVMe · S3 · Ceph

L2 — Network Interconnect & Federation Fabric EDGE-VIABLE InfiniBand · Ethernet · 5G

L1 — Infrastructure Power · Cooling · Physical LEVERAGE POINT PUE · Tidal · Solar · Grid

◈

Select a layer
to inspect its
architecture profile

Node Topology // Right-sized compute spectrum

EDGE — R740

Rack Server Node

2× Xeon / 5060 GPU
128GB RAM · 4TB NVMe
Solar + Grid hybrid
The Garioch, Aberdeenshire

EDGE — RPi

Raspberry Pi Node

ARM Cortex-A72
8GB RAM · USB SSD
Low-power inference
Quantised 3B–7B models

EDGE — OIL PLATFORM

North Sea Ruggedised

Tampnet fiber backbone
350+ platform coverage
Subsea routing proven
Harsh environment rated

HYBRID — TIDAL

Renewable Anchor Node

MeyGen / SIMEC power
Pentland Firth capacity
2GW Peterhead HVDC
Grid-forming compute

HYBRID — MYCELIUM

Data Centre + Mushrooms

Server waste heat → grow
Medicinal cultivation
Dual revenue stream
Circular economy model

HYPERSCALE (reference)

AWS / Azure / GCP

Centralised extraction
A100/H100 clusters
Economic leakage model
Single jurisdiction risk

Edge vs Hyperscale // Key differentiators per block

Building Block	Edge Advantage	Edge	Hyperscale
Infrastructure / Power	Tidal, solar, waste heat re-use — lowest PUE possible
Network / Interconnect	Existing fiber (North Sea proven), no new builds needed
Storage	Local NVMe serves 7B–70B weights without round-trips
Memory Bandwidth	GPU VRAM scales linearly — rack servers viable for inference
Compute (GPU fabric)	Consumer / prosumer GPUs sufficient for quantised inference
Inference Runtime	llama.cpp / Ollama run on CPU — no GPU required for small models
Model (Foundation)	Open-weight models (Llama, Mistral, Phi) run edge-deployed
Data Sovereignty	Computation stays within jurisdiction — economic value local