A starter hardware assumption from the capsule deck: graphics processors tied to each capsule or nearby compute bay.
Let The Beds Think When No One Is Sleeping
Each capsule can be designed as a user-first compute berth: private and quiet while occupied, then routed into island workloads when vacant. The brave path is simple: start with per-capsule GPUs, prove real demand, then build toward a roughly $3M AUD NVIDIA-class supercompute rack for local models, media generation and scientific simulation.
Annual planning band for unused capsule capacity serving island workloads instead of sitting dark.
A future local NVIDIA-class rack ambition, treated as infrastructure to earn rather than shiny gear to worship.
Kiosks, noticeboards, digital twin, Aura, local LLMs, creative media and science-model experiments.
This is not a procurement quote. It is a planning target and capability ladder: capsule GPUs first, shared node discipline next, serious rack only when power, cooling, governance, operators and workloads are ready.
The hotel becomes an island brain after check-out.
Occupied capsules belong to people. Idle capsules belong to the island. That rule turns accommodation into a distributed compute commons without pretending every workload belongs in public, every dataset can be shared, or every model should run all the time.
User-first rule
When a capsule is occupied, comfort, privacy, noise, heat and power limits come first.
Idle compute pool
When vacant, GPUs can process local language models, image/video/song generation, maps, notices and simulation batches.
Survival fallback
Disaster workloads degrade to low-power kiosks, e-ink, mesh messages and cached maps when heavy compute is unsafe.
Rack escalation
The full rack target becomes credible only after smaller GPU-hours prove repeat use and local operators are trained.
What The Compute Powers
The compute layer should earn trust by doing useful work on ordinary days and holding shape on hard days.
Disaster kiosk intelligence
Pre-cache maps, run local LLM summaries, simulate evacuation pressure, prepare LoRa/mesh message packs and support kiosk operators without relying on the cloud.
Noticeboard network production
Render public notices into device formats, check stale data, build agent-readable Markdown, translate summaries and keep the island's public memory searchable.
Virtual Minjerribah
Run batches for visitor flow, transport, waste, events, weather pressure, microgrid demand and 4D island-history simulations before decisions hit the street.
Aura and health surge support
Host consent-bound local models for intake summaries, carer notes, public health education and privacy-preserving analysis, with clinical claims kept out of the public layer.
Local GenAI studio
Train residents on text, image, video, song and LLM workflows so Straddie can make its own public media, grant drafts, scenarios and education tools.
Science-model experiments
Prepare a path for materials-discovery and protein-folding style workloads, including GNoME and AlphaFold-class thinking where software access, licensing and review allow.
Web3 Sensorium bridge
Turn the island into a practical sovereign node for a larger simulation commons: signed data, CRDT sync, open debate, local-first identity and global model exchange.
The Rack Has To Earn Its Place
A roughly $3M AUD compute rack should be argued like civic infrastructure, not bought like a trophy.
Power and cooling
Measure heat, noise, solar, grid, backup power, battery reserve, cooling and maintenance before pretending the rack fits any site.
Operators
Train local people to run jobs, manage queues, protect data, document outputs and maintain hardware with external support only where needed.
Governance
Separate public workloads from private health, cultural, identity and commercial data. Use consent gates, source logs and withdrawal pathways.
Queue discipline
Rank jobs by public value: emergency, health, civic simulation, education, local enterprise, creative production and research.
Rack-ready proof
Use capsule GPU-hours to show demand, utilisation, training outcomes, public value, uptime and avoided cloud dependence.
Graceful failure
When heavy compute drops, kiosks, noticeboards and survival nodes keep serving core messages, maps and local coordination.