Doc #129 — AI Inference Facility Operations

---

RECOMMENDED ACTIONS

Pre-event: Iterative Development

The facility does not require a multi-year construction programme before useful work begins. Development is iterative — each phase generates value and informs the next.

1. [Immediately] Begin recovery research using commercial APIs. The Recovery Library and other recovery research can begin now, using commercially available AI inference (Claude, GPT, open-source models via cloud providers). This work generates the knowledge base, establishes editorial standards, identifies gaps, and builds the team — none of which requires local hardware. The cost is modest and the output is immediate.

2. [Month 1–6] Establish the entity and begin local inference. Incorporate the commercial entity with charter provisions for resilience mandate and governance transition. Acquire initial inference hardware and host in an existing NZ colocation facility (e.g., Datacom, Spark NZ, or CDC Data Centres operating in Auckland and Wellington).⁷ Begin serving initial data sovereignty clients alongside continued recovery research. Cache comprehensive open-source model weights locally.⁸

3. [Year 1–2] Scale within colocation. Expand hardware as revenue and demand grow. Move from single devices to a rack or cage in a commercial colocation centre. The facility operates commercially while dedicating a share of capacity to recovery research. Publish recovery knowledge base openly for peer review. Begin pre-positioning Tier 3 community devices and Tier 4 USB libraries.

4. [Year 2–5] Purpose-built facility (if scale warrants). As demand for data-sovereign inference grows, design and construct a purpose-built facility with professional environmental controls, redundant power, seismic resilience, and domestic fibre connectivity.⁹ Migrate from colocation. This phase may not be necessary if colocation capacity proves sufficient — the architecture should not depend on it.

5. [Year 1 onward] Build the global spoke network. Pre-position Tier 3 community devices worldwide. Establish Tier 2 regional installation partnerships. Build Tier 4 USB library production capacity. The network is useful in peacetime and transforms into a resilience asset in catastrophe.

6. [Ongoing] Maintain operations team and offline documentation. Cross-train operators. Document all procedures for someone technically competent but unfamiliar with the specific facility. Print and store complete setup documentation.

Post-event: Transition to Recovery Operations

7. [Day 1–3] Activate recovery mode. Verify hardware, model weights, and power. Notify government of operational status.

8. [Week 1] Execute governance transition. Shift from commercial to public stewardship per charter. Establish recovery governance board. Secure facility as critical infrastructure.

9. [Month 1] Begin systematic knowledge extraction. Generate the full Recovery Library (or update it if already complete). Generate customized guidance for Tier 2 and Tier 3 recipients globally. Activate distribution via fibre, HF radio, air, and sail.

10. [Months 1–36] Maximum-throughput operations. Serve NZ users via domestic network per priority framework. Cross-train additional operators. Implement rigorous maintenance regime.

11. [Year 1–5] Redistribute surplus hardware. Ship backup equipment to Tier 2 installations by air (while fuel lasts) then sail.

12. [Year 3–5] Accelerate preservation. As hardware approaches mid-life, ensure all critical knowledge exists on durable media — print, USB, institutional curricula.

13. [End of life] Graceful shutdown. Repurpose surviving components for general computing (Doc #14). Document operational history.

---

ECONOMIC JUSTIFICATION

Commercial case (pre-event)

Organizations worldwide — corporations, financial institutions, government agencies, healthcare systems — are increasingly uneasy about hosting AI inference with providers subject to great-power jurisdictions. A European pharmaceutical company running analysis through a US service is subject to CLOUD Act reach; a Latin American central bank using Chinese services faces analogous concerns.¹⁰

NZ offers: political neutrality (independent foreign policy, nuclear-free legislation, no military alliance entanglement)¹¹; strong rule of law (top-tier global rankings for judicial independence and corruption absence)¹²; transparent regulation (Privacy Act 2020)¹³; 85%+ renewable electricity; and geographic remoteness from conflict zones.

Strategic benefit. Serving clients across competing power centers gives multiple actors a direct interest in NZ’s continued stability and sovereignty — reinforcing NZ’s neutral position.

Revenue. AI inference is sold by the token or hour. As of 2025–2026, commercial pricing ranges from US$0.50 to US$60 per million tokens depending on model and provider.¹⁴ Even a modest operation — a niche data sovereignty service serving privacy-conscious professionals and organisations — generates revenue covering operational costs and the dedicated recovery research allocation. The addressable market for jurisdictionally independent AI inference is potentially much larger (data sovereignty is a rapidly growing concern across legal, financial, medical, and government sectors worldwide), but the facility does not need to capture a large share to be viable.

Construction cost. The hub facility is commercially funded and its scale depends on demand. A niche operation serving a handful of privacy-conscious clients requires modest capital; a facility capturing a significant share of the growing data sovereignty market could be substantially larger. Either way, the commercial case justifies the investment — the resilience capability rides free on commercially justified infrastructure. The distributed spoke network (hundreds of compact inference devices pre-positioned worldwide) is primarily a philanthropic and resilience investment, estimated at NZ$5–20 million depending on scale.^{15 16}

Recovery value (post-event)

• Recovery Library production: 174 documents requiring an estimated 7–28 person-years by human experts; the AI facility produces them in 1–3 person-years including human review.¹⁷
• Technical consultation: Replaces an estimated 5–20 person-years of specialist labor annually across medical, engineering, agricultural, and policy applications.
• Global distribution: Generates customized recovery guidance for thousands of communities — a combinatorial task no human team could accomplish from printed references.
• Conservative NZ-only estimate: 10–30 person-years of specialist value produced per year, at 5–15 person-years of operational cost. Return: approximately 2:1 to 6:1.¹⁸

Operational workforce requirements

The facility requires a small, highly specialised standing team. These roles are staffed commercially pre-event and transition to recovery operations post-event.

Role	FTE	Person-years over 10-year facility life
AI/ML engineers (model operations, inference tuning, query pipeline)	1–2	10–20
Datacenter technicians (hardware monitoring, replacement, spare parts management)	1–2	10–20
Power engineers (grid connection, UPS, generator, electrical maintenance)	0.5–1	5–10
Cooling system operators (HVAC, free-cooling management, thermal monitoring)	0.5–1	5–10
Operations manager / knowledge extraction coordinator	1	10
Total	4–7 FTE	40–70 person-years

These roles are filled by the commercial operations team during normal operations. The incremental cost of the resilience mandate — the charter provisions, the inference allocation for recovery research, the spare-parts stockpile — is modest relative to the commercial staffing budget. Post-event, the same team transitions to recovery operations; no new recruitment is required at the moment of greatest scarcity.¹⁹

This staffing profile is not light. AI/ML engineers and power engineers are among the most sought-after technical professionals in NZ. Their presence at the facility must be an established commercial fact before any catastrophe, not an aspiration to be fulfilled afterward.

Maintaining capability vs. losing it

The alternative to an operational AI inference facility is not a smooth degradation in decision quality — it is the loss of a class of capability that cannot be reconstituted until semiconductor manufacturing resumes, likely decades later. The comparison is instructive.

With facility operational: - A general practitioner in Invercargill, facing an unfamiliar presentation, queries the system and receives a differential diagnosis ranked by probability, with recommended investigations and treatment pathways — within minutes. - An engineer repairing a water treatment plant queries for chemical dosing calculations specific to the local source water chemistry — correct values returned in seconds. - A policy team drafting rationing legislation receives comparative analysis of six historical rationing frameworks, with NZ-specific adaptation notes, in hours rather than weeks. - The Recovery Library’s 174 documents are produced, reviewed, and updated continuously as field conditions evolve.

Without facility operational: - The same GP relies on printed references (if available) or their own memory — feasible for common presentations, dangerous for uncommon ones. - The same engineer works from first principles or waits for a specialist who may be unavailable or weeks away. - The same policy team conducts analysis manually — the work takes weeks and draws personnel away from implementation. - The Recovery Library, if not already completed pre-event, requires years of human expert effort under conditions of acute resource scarcity.

The asymmetry is significant. Each individual query that receives a useful AI-assisted answer represents hours to days of human specialist time avoided. Across the full portfolio of recovery domains — medicine, engineering, agriculture, logistics, education, policy — the compounding effect over years is substantial. Losing the facility means losing this multiplier across all domains simultaneously, for decades or longer, at the moment when it is most needed.²⁰

Breakeven analysis

The breakeven question is: at what level of utilisation does the facility’s operational cost become justified by recovery value produced?

Operational cost (post-event, government-funded): 4–7 FTE at NZ recovery-economy salary equivalents, plus power (500 kW–1 MW = NZ$500,000–1,000,000/year at pre-event rates) and maintenance consumables. Total annual operating cost: approximately NZ$1.5–3.5 million/year in pre-event dollar terms.²¹

Value produced per year (NZ-only, conservative): 10–30 person-years of specialist labor equivalent at NZ specialist salaries (NZ$100,000–200,000/year). Equivalent value: NZ$1–6 million/year.

Breakeven: At the low end of utilisation (10 person-years equivalent output, NZ$1M value), the facility breaks even against the high end of operational cost (NZ$3.5M) only if indirect benefits — faster decisions, reduced errors, lower mortality from better-informed medical and engineering choices — are counted. These indirect benefits are real and large but difficult to quantify precisely.

At moderate utilisation (20 person-years equivalent, NZ$3–4M value), the facility covers its operational cost purely on direct specialist labor displacement, before any indirect benefits are counted.

The more defensible framing: The facility does not need to break even on operational economics alone. It is infrastructure — analogous to grid maintenance or telecommunications upkeep — whose value is partially expressed in the decisions enabled rather than the labor displaced. A functioning grid does not need to justify itself against the cost of candles. A functioning AI inference facility does not need to justify itself solely against the cost of hiring more specialists.²²

Opportunity cost

The opportunity cost of operating the facility falls on two axes: personnel and electricity.

Personnel. AI/ML engineers and power engineers are among the most constrained technical specialists in a post-event NZ. A recovery economy with approximately 5.2 million people²³ has perhaps a few dozen individuals qualified to staff the facility at the required level. Dedicating 4–7 of them to AI inference operations means those individuals are unavailable for other high-value technical roles — telecommunications maintenance, grid operations, industrial process support.

This is a genuine tradeoff, not a rhetorical one. The counter-argument is that the facility’s output — the knowledge multiplier it provides across all recovery domains — reduces the demand for specialist presence in those other domains. The facility enables a single power engineer to provide guidance to dozens of remote sites via AI-assisted instruction, rather than traveling to each site. The facility’s operators, in effect, operate a force multiplier for every other specialist in the country. The net effect on the specialist labor pool is positive if utilisation is sufficient — but this depends on the facility actually being used at scale, which requires the priority framework and access infrastructure to function as designed.²⁴

Electricity. At 500 kW–1 MW, the facility consumes 4,400–8,760 MWh/year — approximately 0.01–0.02% of NZ’s annual generation (~43,000 GWh). Under the baseline scenario (grid continues at 85%+ renewable), this is a negligible load. Under a degraded-grid scenario where NZ is operating on reduced hydro generation and prioritising industrial and domestic supply, the facility’s electricity allocation becomes a genuine policy decision. A 1 MW load, maintained at high priority, displaces other uses. The governance framework should include explicit provisions for facility power priority under grid-constrained conditions — not because the electricity cost is large in absolute terms, but because the priority decision will be contested and should be pre-resolved rather than improvised.²⁵

The free rider. Recovery capability costs almost nothing beyond a charter provision and a modest inference allocation. The commercial case justifies construction. If the catastrophe never occurs, the commercial investment stands on its own. If it does occur, the capability exists at no additional cost.

---

THE FACILITY: HUB-AND-SPOKE ARCHITECTURE

Design philosophy

The original Recoverable working paper proposed an AI inference centre. Subsequent analysis — particularly an investigation into whether the facility would be redundant given existing Southern Hemisphere infrastructure — led to a substantially revised architecture: a modest hub facility in NZ serving as the nerve centre for a distributed network of simpler inference nodes, rather than a large-scale data centre attempting to compete with commercial providers.²⁶

The hub’s primary value is not raw compute — it is its function as a knowledge refinery and coordination centre. The hub performs the heavy cognitive work requiring the most advanced AI capability: synthesizing information, reasoning about novel problems, coordinating priorities across regions, distilling specialist models for spoke nodes, and producing tailored guidance packages. The outputs flow outward through physical shipment to hundreds of spoke nodes that run simpler inference locally.

This is an ancient model — publishing houses, universities, and agricultural extension services have always worked this way — applied with new technology.

The hub (NZ)

A well-equipped computing centre whose scale depends on commercial demand. A small facility serving niche data sovereignty clients is useful; a larger facility capturing a broader share of the market is more useful — and generates more capacity for recovery research. The facility houses independent inference devices — enough to serve commercial clients and handle compute-intensive tasks like model distillation, fine-tuning, and knowledge package generation. The architecture should not artificially constrain scale.²⁷

Key design principles: seismic resilience (Canterbury or Otago preferred over Wellington — Christchurch’s rebuild incorporated modern seismic standards, and Otago/Southland has the lowest seismic hazard of NZ’s main centres²⁸); designed for autonomous, off-grid operation; graceful degradation (independent inference nodes rather than tightly coupled superclusters — every failure reduces capacity linearly, not catastrophically); spare parts stockpiled for 5–10 years of operation; and simplified, maintainable architecture.²⁹

The facility also serves as the canonical repository of human knowledge: comprehensive internet crawls, scientific literature, engineering databases, agricultural records, medical references, cultural archives — petabytes of redundant, indexed, retrievable storage. Storage is cheap, robust, and low-power compared to compute, and this breadth of stored knowledge is genuinely hard to replicate at spoke nodes.

The spokes (distributed inference nodes)

Independent, compact inference devices with large unified memory (128–192 GB as of early 2026, potentially larger as the technology matures), capable of running quantized models in the 70B–100B parameter range at useful speed — pre-positioned at key locations across the Southern Hemisphere and Pacific: universities, hospitals, government offices, community centres, marae. Each is self-sufficient for basic inference with pre-loaded models and local knowledge stores. Dozens or hundreds of them, each costing NZ$5,000–15,000 and consuming 100–300 W.³⁰ Performance is substantially slower than the hub facility — expect 2–10 tokens per second on larger models versus 20–80 tokens per second on purpose-built GPU hardware — but sufficient for interactive consultation on recovery-relevant queries.

For the vast majority of recovery-relevant queries — agricultural guidance, medical triage, engineering reference, language translation, educational content — a single high-end device with 128–192 GB unified memory running a quantized model (70B–100B parameters) provides competent inference.³¹ Larger unified-memory configurations may become available as the technology matures. Recovery-relevant inference at the point of need does not require 18,500 GPUs.

Spoke nodes need clear, simple interfaces for receiving and deploying updates: “plug in the hard drive, confirm the update, restart.” Day-to-day query operation should be accessible to anyone with basic computer literacy; maintenance, updates, and troubleshooting require someone with Linux system administration competence but not machine learning expertise.

Architecture resilience spectrum

Not all architectures are equally fragile. Hyperscale GPU superclusters (like Australia’s Project Southgate with 18,500 GPUs) are tightly coupled — dependent on complex liquid cooling, high-speed networking, and sophisticated power distribution. They fail catastrophically rather than gracefully: cutting the municipal water supply or disabling liquid cooling circulation pumps causes processors to thermal-throttle within minutes; sustained cooling loss forces shutdown within minutes to hours depending on ambient temperature and backup cooling capacity.³²

The hub-and-spoke architecture sits at the resilient end of the spectrum. Independent, self-contained inference devices degrade linearly: lose ten percent of the devices, lose ten percent of capacity. No single failure is catastrophic. Spoke nodes require an operator with competence in Linux system administration, hardware troubleshooting, and basic networking — more than casual computer literacy but less than the ML engineering expertise required at the hub. The hub is more complex but is still designed around independent inference nodes rather than tightly coupled clusters.

What the hub produces for the spokes

In a degraded communications environment — where submarine cables may be permanently severed and satellite bandwidth severely constrained — the hub serves the spoke network primarily through physical shipment of data. What it ships:

• Distilled specialist models. The hub’s most capable systems compress domain expertise into smaller models that run on spoke hardware. A “hydroelectric maintenance specialist” model that fits in 20–40 GB, runs on a compact device, and interactively guides a technician through common repair scenarios. The number and quality of distilled models depends on the hub’s compute capacity and the state of model distillation techniques — initial targets would cover the highest-priority domains: medical care, power systems, agriculture, water treatment, and manufacturing.
• Fine-tuning packages. Weight deltas (potentially a few gigabytes) that upgrade the general-purpose local model’s capability across dozens of domains. The spoke applies these to its existing model.
• Curated agentic knowledge packages. Extremely detailed, structured knowledge for specific tasks — decision trees, diagnostic flowcharts, step-by-step procedures with branching logic for different failure modes. Designed to be consumed by a simpler model running an agentic workflow locally. The local model navigates expert-created decision structures, asks the user the right diagnostic questions, and guides them through complex procedures.
• Raw reference data. Technical manuals, scientific literature, engineering specifications for retrieval-augmented generation at the spokes.

This is shipped on physical media — hard drives or equivalent — totalling hundreds of gigabytes to low terabytes per spoke update. The maritime distribution network (Doc #138, Doc #140) maps naturally to this architecture.

Why New Zealand — and why not elsewhere

The commercial arguments (neutrality, rule of law, renewable energy, privacy legislation) and the resilience arguments (remoteness from targets, renewable grid, food self-sufficiency under nuclear winter³³, domestic fibre network, institutional stability) reinforce each other. The same attributes that attract data sovereignty clients are the attributes that make NZ likely to survive a catastrophe with infrastructure intact. This is the core insight: the commercial case and the resilience case are not in tension — they are the same case, made to different audiences.

But the strongest argument for a NZ facility — and the one that most clearly distinguishes it from alternatives — is targeting.

Data centres as military targets. AI data centres are increasingly recognized in defence doctrine as strategic military assets — not peripheral commercial infrastructure, but high-value targets whose destruction directly degrades an adversary’s warfighting capability. Computational infrastructure is now a domain of warfare. The strategic logic is compelling and, in a nuclear exchange, effectively irreversible: destroying an adversary’s AI compute does not set back capability by months — it eliminates it for decades or longer, because the global semiconductor manufacturing base (concentrated in Taiwan, South Korea, the US, and Japan — all primary target zones) would itself be destroyed. There would be no replacement GPUs, no new server hardware, no path to reconstituting destroyed AI capability for decades or longer.³⁴

Australia’s vulnerability. Australia is the dominant AI infrastructure player in the Southern Hemisphere, with total data centre capacity of approximately 1.3 GW and an enormous investment pipeline (AWS AU$20B, Microsoft AU$5B, OpenAI/NextDC $7B). But Australian AI infrastructure is now explicitly integrated with defence operations. The OpenAI/NextDC facility is described as “sovereign AI infrastructure” aligned with Australia’s Security of Critical Infrastructure (SOCI) framework. Google is negotiating an AI data centre on Christmas Island following a three-year computing agreement with the Department of Defence. AWS operates a “Top Secret” cloud for national security. AUKUS deepens integration with US military operations. These are not neutral commercial assets — they are military-industrial targets that also serve commercial customers.³⁵

An adversary assessing these facilities sees US-owned or US-partnered GPU superclusters, explicitly integrated with Australian and US defence operations, in a country that already hosts Pine Gap and is the US’s primary Southern Hemisphere military partner. The defence integration makes them rational targets; the irrecoverability of the loss makes them high-priority targets. Facilities can also be destroyed through non-nuclear means — conventional cruise missiles, high-altitude EMP, or cyber warfare — without requiring a nuclear strike on Australian cities.³⁶

NZ’s non-target status. NZ occupies a fundamentally different targeting category. NZ has: no US military bases; nuclear-free zone status since 1987; no AUKUS membership; no integration into US intelligence/military physical infrastructure. An AI inference facility in NZ, owned by a NZ entity or international foundation, with no defence contracts and no military integration, presents no strategic rationale for targeting. This non-target status is a structural feature of NZ’s positioning — but it must be actively maintained. It requires: no defence contracts whatsoever; local or foundation ownership (not a subsidiary of any US or allied-nation technology company); transparent operations open to international inspection; and international governance including advisors from multiple nations. The facility should be positioned alongside institutions like the Svalbard Global Seed Vault and the ICRC — assets the international community has implicitly agreed to preserve because their purpose transcends geopolitical competition.³⁷

NZ’s negligible existing AI infrastructure. NZ has only 432 MW of total data centre IT load, projected to reach 591 MW by 2030. AWS and Microsoft have launched NZ cloud regions, but these are enterprise workload platforms — not GPU inference clusters. NZ has negligible AI-specific inference capacity.³⁸ The proposed facility fills a gap, not a redundancy.

Other Southern Hemisphere alternatives. Brazil has approximately 707 MW of colocation capacity, overwhelmingly in São Paulo — massive scale but concentrated in a complex urban environment with social stability concerns. Argentina’s Patagonia Stargate project (OpenAI/Sur Energy, up to 500 MW) is potentially the most interesting facility from a geographic safety perspective, but remains at the Letter of Intent stage and would be a commercial OpenAI facility with no recovery-specific design, no autonomous operational capability, and no governance structure enabling pivot to recovery operations.³⁹ Africa has less than 1% of global data centre capacity. Southeast Asia’s $55B+ investment pipeline is concentrated near South China Sea flashpoints.

No alternative location combines all of NZ’s advantages: no nuclear targeting risk, no military alliance entanglements, established renewable energy infrastructure operating independently of global fuel supplies, stable democratic governance, geographic isolation from conflict zones, and an existing technical workforce.

The three differentiating pillars of the NZ facility, in order of argumentative strength:⁴⁰

1. Non-target status (strongest). Cannot be retrofitted onto facilities already integrated into defence operations. In a world where AI compute destruction is permanent, the targeting argument is determinative — targeted facilities do not survive, and no replacements can be manufactured.
2. Architectural resilience (strong). Purpose-built for graceful degradation and autonomous operation under prolonged degraded conditions where no resupply is possible. Commercial facilities without spare parts stockpiles or autonomous power would lose significant capacity within months to a few years;⁴¹ a purpose-built facility with stockpiled spares outlasts them by years.
3. Purpose and readiness (moderate). Pre-loaded recovery content, trained operations team, governance transition mechanism, systematic knowledge extraction programme. Most valuable in the immediate post-catastrophe period when operational delays reduce the window for knowledge extraction.

Network connectivity

NZ’s domestic fibre and cellular networks (Doc #127) serve the facility to the entire country — any connected terminal can access AI capability. Latency across NZ’s fibre: under 20 ms.⁴² Pre-event, international submarine cable connects to global clients. Post-event, domestic network serves NZ; HF radio, air, and sail connect to the world.

Communication infrastructure itself is vulnerable: submarine cable terminal stations are high-value targets (destroying a handful severs entire continents), and repair requires specialized cable-laying ships that would be unavailable post-exchange. Satellite constellations degrade as ground stations fail. The hub-and-spoke architecture is designed for this reality — spoke nodes operate independently, and the hub connects to spokes via physical shipment of data on hard drives, not real-time network links.

---

INFERENCE VS. TRAINING

The computational asymmetry

Training a large language model — adjusting billions of parameters across trillions of tokens of training data — is one of the most computationally expensive tasks humans have ever attempted. NZ cannot do this, and does not need to.

Inference — running a pre-trained model to generate output from a prompt — requires a tiny fraction of the compute used for training. The asymmetry is roughly analogous to the difference between writing a book and reading it. The model weights (the “book”) are the product of training. Once the weights exist, running inference requires only enough hardware to load those weights into memory and perform matrix multiplications at useful speed.

Model Size	VRAM (FP16)	VRAM (4-bit quant)	Min. Hardware	Tokens/sec
7B parameters	~14 GB	~4 GB	1x RTX 4090	30–80
70B parameters	~140 GB	~35 GB	1x A100-80GB	5–20
405B parameters	~810 GB	~200 GB	4–8x A100-80GB	2–10

⁴³

Model weights

Model weights are billions of floating-point numbers encoding the patterns learned during training — a compressed representation of broad technical and general knowledge. Without weights, GPUs are purposeless. In a commercially operating facility, weights are an operational necessity — eliminating the risk that NZ might lack locally cached model weights when connectivity is severed.⁴⁴

Which models to cache

1. General-purpose LLMs — best available open-source (Llama 3.1 405B, DeepSeek-V3/R1, Qwen 2.5 as of early 2026⁴⁵); full-precision and quantised variants
2. Efficient medium models (7B–30B) — single-GPU fallback and basis for Tier 3 devices
3. Code generation — software maintenance across all computing systems
4. Medical reference — clinical decision support (complements Doc #4, Doc #116)
5. Multilingual/translation — NZ’s diverse population and global distribution needs
6. Embedding/retrieval — semantic search over document collections

Total storage: approximately 1–2 TB — trivial for a commercial facility.⁴⁶

Software stack

Linux, GPU drivers (CUDA/ROCm), inference engine (vLLM, llama.cpp, text-generation-inference, or Ollama⁴⁷), and web interface. All open-source, stored locally, documented in print.

---

LAYERED GLOBAL KNOWLEDGE DISTRIBUTION

The facility does not operate in isolation. Knowledge is distributed through five tiers — decreasing capability but increasing distribution and resilience.

Tier 1: Central inference facility (NZ)

The purpose-built facility. Full-scale inference on largest models. Handles hardest problems. Generates comprehensive documentation and customized guidance for all lower tiers. One facility; its loss is severe but knowledge already distributed persists.

Tier 2: Regional inference installations

Existing data centres in countries with reliable (ideally renewable) electricity — Australia, Brazil, Scandinavia, and others. Running medium-to-large open-source models (70B–405B) on available hardware. The central facility generates customized setup guidance (“You have 4 A100 GPUs in a 30°C room with 15 kW available — here is your configuration”), provides model weights via physical media, and redistributes surplus hardware by air or sail. Once configured, Tier 2 installations operate independently.

Tier 3: Community-level inference devices

These are the spoke nodes described above. Compact devices with large unified memory running quantized models in the 70B–100B parameter range — capable of competent inference at moderate power consumption (100–300 W).⁴⁸ Cost: NZ$5,000–15,000 each. Pre-positioned in hundreds to thousands of communities worldwide — libraries, hospitals, schools, marae. Each device is self-contained, requiring only electricity and an operator with basic technical literacy. Output quality is lower than the hub’s largest models, and response times are slower, but the devices are sufficient for the medical, agricultural, and engineering reference queries that constitute the bulk of recovery-relevant demand.

Tier 4: Pre-loaded storage devices

USB drives loaded with comprehensive technical libraries. Contents include:

• The complete Recovery Library (all documents, NZ and international editions)
• Agricultural manuals adapted to local climate zones
• Medical references — diagnostic guides, pharmacopoeia, surgical procedures, public health protocols
• Engineering databases — materials properties, structural calculations, electrical reference
• Educational curricula — primary, secondary, vocational, technical
• Manufacturing procedures — from blacksmithing to chemical processes
• Navigation tables, mathematical reference, astronomical data
• Multilingual versions of all critical documents

A 1 TB USB drive costs approximately NZ$15–30, holds the entire collection, has no moving parts, and tolerates years of storage.⁴⁹ The drives are readable by any surviving computer — desktop, laptop, tablet, or smartphone with an adapter. Pre-positioned in libraries, hospitals, schools, emergency depots, and civil defence facilities worldwide. One million drives: NZ$15–30 million depending on procurement scale and timing — a favourable cost-to-value ratio, though distribution logistics (shipping, customs, local placement) add substantially to the total programme cost.⁵⁰

Tier 5: Personal devices

A large but uncertain number of smartphones and laptops survive in non-targeted regions — potentially hundreds of millions, though EMP effects, power grid failures, and physical damage reduce the surviving fleet substantially from pre-event installed base.⁵¹ Small language models (1B–7B parameters) now run on ordinary phones.⁵² The challenge is ensuring devices carry useful models and knowledge pre-event. Pre-event actions: distribute apps with small models and offline knowledge bases through app stores while billions of devices can download at zero marginal cost.

Distribution channels

Knowledge flows from the central facility outward through multiple channels with different speed, capacity, and reliability characteristics:

• Domestic fibre/cellular — NZ’s telecom infrastructure (Doc #127) largely survives in a non-targeted country. Electronic distribution to every connected home, office, hospital, and school in NZ is immediate. Also serves countries with surviving telecommunications.
• Air transport — fastest for physical media (USB drives, hard drives with model weights), compact hardware (Tier 3 devices), and specialist personnel. NZ-Australia: approximately 3 hours. Rationed fuel provides an estimated 5–10 year window for critical air deliveries.⁵³
• Sail — primary mode for long-distance bulk transport as aviation fuel depletes. Tasman crossing: 1–2 weeks. Longer voyages to Southeast Asia, South America, Pacific Islands: weeks to months.⁵⁴ Carries heavier cargo: replacement hardware, printed document sets, Tier 3 devices, supplies for Tier 2 installations.
• HF radio (Doc #128) — trans-oceanic text-based communication requiring no infrastructure. Too slow for model weights or documents (tens to hundreds of characters per minute), but sufficient for coordinating distribution, transmitting urgent queries, and sending compressed summaries. The channel of last resort — slow but global and effectively unkillable.
• Electric vehicles — local distribution within NZ. EVs do not depend on imported fuel. NZ’s road network is physically intact under the baseline scenario. USB drives, printed materials, and small hardware reach every community.

---

THE AGI COORDINATION ROLE

One of the core problems of catastrophe recovery is expertise. Resources, infrastructure, and supply chains are enormous challenges in their own right. But the loss of specialised knowledge compounds every other problem. Functioning hydroelectric dams may lack anyone who knows how to maintain them. Intact machine shops may lack anyone who can design needed parts. Seeds may be available but agronomic knowledge for radically altered climate conditions may be absent. Historically, civilisational recovery has been bottlenecked by the loss of tacit knowledge — the knowledge that lives in experts’ heads rather than in books.

AI systems capable of expert-level reasoning change this equation fundamentally. Instead of needing thousands of specialists distributed across every critical domain, you need one system that can reason at expert level across all of them, and ways to get that reasoning to the people who need it. That is a structurally different — and more tractable — problem, provided the inference facility survives and remains operational.⁵⁵

Within NZ, this is relatively straightforward. The domestic telecommunications network would probably survive largely intact, enabling real-time AI-assisted guidance: an engineer in Dunedin repairing a water treatment plant communicates with the system and receives step-by-step interactive guidance. One AI system effectively replaces the need for thousands of distributed domain experts within the country.

The hard problem is crossing oceans. If submarine cables are permanently severed and satellite bandwidth is unavailable, real-time interactive guidance to distant regions is lost. The hub’s AI then takes on its highest-value role: anticipating questions that have not been asked yet and pre-computing the answers. It analyses what it knows about conditions in, say, coastal Peru — based on climate models, pre-catastrophe data, and whatever sparse information arrives via HF radio or returning ships — reasons about what problems those communities are likely facing, and produces tailored guidance packages before anyone asks.

This is a demanding reasoning task — synthesising sparse, noisy information about diverse situations and producing contextualised, actionable outputs — that improves substantially with model capability. The facility’s value increases nonlinearly as model capability improves. Building now, sized for current models but architecturally ready for more capable models when they arrive, is a sound investment — the facility does not depend on future capability improvements to be useful, but its value increases with model capability.

---

HARDWARE LIFESPAN AND MAINTENANCE

Failure modes

Electronic hardware fails through several mechanisms: electrolytic capacitor aging (life halves per 10°C above rated temperature)⁵⁶; solder joint fatigue from thermal cycling⁵⁷; fan bearing failure (4–8 years)⁵⁸; memory degradation (increasing bit error rates over time)⁵⁹; power supply failure (7–12 years); and SSD degradation (though inference workloads are read-heavy, minimising wear).⁶⁰ In a facility maintained at 18–24°C with stable load patterns, these failure modes are pushed toward the upper end of their ranges. The key principle: keep hardware cool, clean, and running at stable temperatures.

Lifespan projections (purpose-built, professionally maintained)

Component	Expected Life	Limiting Factor
GPU compute die	10–20+ years	Electromigration
GPU HBM memory	8–15 years	Cell degradation
Power supply	7–12 years	Capacitor aging
Cooling fans	4–8 years	Bearing wear
System DRAM	8–15 years	Cell degradation
SSDs (read-heavy)	7–15 years	Controller electronics
Motherboard	8–15 years	Capacitor aging

⁶¹

Aggregate lifespan: 7–12 years primary, possibly 15 years with cannibalisation. The purpose-built advantage over a repurposed facility: perhaps 2–5 additional years — representing 2–5 additional years of knowledge extraction.

Maintenance strategy

Keep hardware cool, clean, and monitored. NZ’s climate is advantageous — annual mean temperatures range from 10°C (Invercargill) to 16°C (Auckland), meaning free cooling (outside air economisation) provides adequate cooling for 7,000–8,500 hours per year depending on location.⁶² Proactive replacement of fans and power supplies before failure prevents cascade damage. Comprehensive spare parts inventory pre-positioned during commercial operations is critical — post-event, this stock cannot be replenished. The spare parts dependency chain includes: replacement cooling fans (ball-bearing type, 40mm–120mm, multiple voltage ratings); complete power supply units (server-grade, matched to installed hardware); thermal interface material (thermal paste/pads degrade and must be reapplied during fan or heatsink replacement); replacement DIMMs (matched to installed motherboard specifications); replacement SSDs; network cables and transceivers; UPS batteries (lead-acid or lithium, requiring replacement every 3–5 years); and generator consumables (fuel filters, oil, coolant). Each of these is an imported manufactured product with no NZ domestic source — the spare parts stockpile is the entire supply.⁶³ Cannibalisation of failed units extends remaining equipment. The hub-and-spoke architecture, with its independent nodes rather than tightly coupled clusters, means individual failures reduce capacity linearly without cascading effects.

---

POWER AND COOLING

The hub facility at 500 kW–1 MW represents less than 0.02% of NZ’s grid generation (~43,000 GWh/year, ~4,900 MW average).⁶⁴ NZ’s 85%+ renewable grid (hydro, geothermal, wind) powers the facility indefinitely without imported fuel. Spoke nodes at 100–150 W each can run from a modest solar installation with battery storage.

Backup power: UPS for ride-through; diesel/biodiesel generator for extended outages (Doc #53, Doc #57). Under nuclear winter conditions with 5–8°C cooling, mechanical cooling loads reduce substantially.⁶⁵ NZ’s temperate climate enables free cooling for most of the year, a significant advantage over tropical or subtropical data centre locations.^{66 67}

---

GOVERNANCE AND ACCESS

Pre-event: Commercial with resilience mandate

Commercial governance funds operations. Charter provisions distinguish the facility: a resilience mandate dedicating capacity to recovery research; a governance transition mechanism activated by defined catastrophe triggers; an open knowledge commitment for all recovery output; and a government liaison ensuring the transition mechanism is known to civil defence authorities.

Post-event: Public stewardship

In a catastrophe scenario, the facility transitions from commercial to public stewardship — the specific governance structure to be determined by the charter, but the principle is clear: the facility passes from private to public hands and serves the recovery of NZ and the world.

Priority framework

• Tier 1 (unrestricted): Government recovery planning, Recovery Library production, medical decision support, critical infrastructure guidance, agricultural support, global distribution coordination
• Tier 2 (scheduled): University research, CRI analysis, regional government planning
• Tier 3 (allocated blocks): Engineering/technical practitioners, translation, training content
• Tier 4 (as capacity permits): Public education, general reference

Ethical considerations

Accuracy and reliability. AI models produce confident-sounding text that is sometimes wrong. Every user must understand this. Output used for medical, engineering, or safety-critical applications must be reviewed by a qualified human before implementation. The facility should maintain a log of significant errors to calibrate user trust.

Equity of access. There is an inherent tension between allocating access by recovery value (which concentrates access among institutional users) and providing equitable public access. The priority framework attempts to balance these, but the tension should be acknowledged openly rather than obscured by rhetoric about equal access that cannot be delivered in practice.

Transparency. The facility’s operations, access policies, and usage statistics should be public information. A scarce national resource operated behind closed doors invites suspicion and undermines public trust.

The fundamental principle: The facility serves NZ and the world — not its investors or operators. This principle, established in the charter, is the foundation of governance legitimacy.

---

POST-EVENT OPERATIONS

Phase 1: Transition (Week 1)

Verify all hardware functionality — run diagnostics on every GPU, server, and storage drive. Confirm model weights are intact and loadable. Verify power supply: grid, UPS, generator. Execute governance transition per charter. Establish physical security — the facility is now critical national infrastructure. Begin serving priority recovery queries immediately: government planners, medical practitioners, infrastructure engineers.

Phase 1: Systematic knowledge extraction (Months 1–6)

This is the highest-value period — full capability meets acute need.

• Generate or update the complete Recovery Library. All 174 catalogued documents plus additions identified by government planners and field practitioners.
• Generate customized global guidance. Installation guides for Tier 2 regional installations, operating manuals for Tier 3 community devices, knowledge packages for Tier 4 USB distribution — all adapted to each recipient’s specific conditions.
• Systematic querying program. A dedicated team whose role is to identify and ask the questions that have not yet been asked. The facility is a time-limited oracle — questions not asked are knowledge lost.
• Print production coordination. Work with Doc #5 (Printing and Document Supply) and Doc #29 (National Printing Plan) to convert digital knowledge into durable printed form.
• If the pre-event research program operated as planned, the Recovery Library already exists as a mature resource. Focus shifts to updating for actual conditions and filling scenario-specific gaps.

Phase 1–2: Ongoing operations (Months 1–36)

Maximum-throughput inference service per the priority framework. Medical reference for NZ’s practitioners. Engineering and technical consultation. Environmental monitoring and rigorous hardware maintenance. Continuous operator cross-training and succession planning — the facility may need to operate for 10–15 years, and the initial operators will not all be available for the entire period.

Phase 2–4: Managed degradation (Years 1–15)

Hardware attrition begins: fan failures (years 4–8), then power supplies (7–12), then GPUs and memory (8–15). Each failure reduces capacity. The response is consolidation: fewer servers running smaller, more efficient models. A 405B model that required 8 GPUs gives way to a 70B model on 1–2 GPUs, then to a 7B model on a single GPU. Capability declines but useful output continues.

Surplus or backup equipment — especially units not yet needed as replacements — can be redistributed to Tier 2 installations while transport is available. As the end approaches, accelerate print production: the goal is that by hardware end-of-life, every useful question has been asked, answered, and committed to durable media. Graceful shutdown when remaining hardware can no longer serve useful inference; repurpose surviving components for general computing (Doc #14).

---

APPLICATIONS

Recovery Library production

The highest-value near-term application. The Recovery Library’s 174 catalogued documents represent a comprehensive guide to civilizational recovery. If the pre-event recovery research program has operated as planned, the library already exists as a mature, peer-reviewed resource. Post-event, the facility updates existing documents for actual conditions, fills gaps, and generates customized international editions in multiple languages.

If the library has not been completed pre-event, generating it from scratch is the immediate post-event priority. An AI inference system can produce the full library in weeks to months with human editorial oversight, versus years if produced by human authors from printed references. The library is a living document while hardware operates — updated, corrected, and expanded as field experience reveals gaps.

Medical reference and clinical decision support

NZ has approximately 17,000–19,000 registered doctors.⁶⁸ Post-event, specialist knowledge is concentrated in a small number of individuals — NZ may have only a handful of experts in any given subspecialty. An AI system serves as a reference tool that a general practitioner can consult when facing unfamiliar conditions. Critical caveat: AI medical output must always be reviewed by a human clinician. The model can hallucinate plausible but incorrect diagnoses. The governance framework must ensure AI medical output is treated as a reference, not an authority.

Engineering and technical reference

Structural calculations, materials properties, chemical process guidance, repair procedures — a rural engineer repairing a hydroelectric turbine governor (Doc #65) or a welder attempting an unfamiliar joint type (Doc #94) can submit a query and receive specific, step-by-step guidance.

Education and training

AI systems generate lesson plans, textbook chapters, examination questions, and worked examples tailored to NZ conditions and skill levels, supporting the expanded vocational and technical training programs described in Doc #157 (Trade Training) and Doc #162 (University Reorientation).

Translation and communication

NZ’s population includes speakers of English, te reo Māori, Samoan, Hindi, Mandarin, Cantonese, and many other languages. Multilingual AI capability provides functional (not perfect) translation between languages — valuable for refugee populations and essential for generating global distribution content in dozens of languages.

Planning, policy analysis, and software maintenance

AI accelerates the analytical groundwork for resource allocation decisions — scenario modeling, comparative analysis, policy drafting. Code-generation models assist programmers in maintaining, debugging, and adapting the software systems (Doc #130) that NZ’s infrastructure depends on, extending the useful life of all software-dependent systems.

---

THE EXPIRATION DATE

The degradation curve

• Years 0–3: Full capability. Maximum knowledge extraction and global distribution window.
• Years 3–7: 10–30% capacity lost. Smaller models, slower speeds. Surplus hardware redistributed.
• Years 7–12: 50%+ capacity lost. Efficient models only. Facility transitions to supplementary role — most knowledge already on durable media.
• Years 12–15: Residual. Small models, high-priority queries only.
• Beyond 15: Central hardware exhausted. Replacement decades away (Phase 7).⁶⁹

Knowledge persists beyond hardware

Tier 2 installations (potentially newer hardware) may continue after the central facility shuts down. Tier 3 devices (simpler, lower-power) may outlast data centre equipment. Tier 4 USB libraries are essentially permanent at room temperature.⁷⁰ Printed documents on acid-free paper survive decades to centuries under reasonable storage conditions; on standard paper, degradation is significant within 50–100 years.⁷¹ Institutional knowledge persists through human transmission. The central facility expires. Its distributed knowledge does not — provided distribution is executed while the facility operates.

The knowledge preservation problem

The questions not asked during the facility’s operational life represent knowledge lost. This argues for systematic querying programs, a dedicated knowledge extraction team, catalog gap reviews, field practitioner feedback, and pre-event research that identifies the most important questions before a catastrophe compresses the timeline. The facility is a time-limited oracle. The quality of the answers depends on the quality of the questions, and there is a deadline.

---

CRITICAL UNCERTAINTIES

Commercial viability

Risk: Data sovereignty demand may not materialize at required scale. NZ may face competition from other neutral jurisdictions. Technology changes may reduce demand. Mitigation: Phased construction. Hybrid funding models supplementing commercial revenue with government resilience investment. Assessment: The single largest uncertainty. If the commercial case fails, the facility is not built, and the architecture does not exist. Contingency: a smaller, partially government-funded facility with reduced capability.

Hardware lifespan

Risk: Shorter than projected due to defects, seismic events, or operational errors. Mitigation: Purpose-built environmental controls, seismic design, comprehensive spare parts, front-loaded knowledge extraction. Assessment: Purpose-built facility should achieve upper-end lifespans. Layered distribution mitigates — knowledge already distributed persists regardless.

Operator availability

Risk: Skilled operators unavailable post-event. Mitigation: Standing commercial team (substantial advantage over ad hoc assembly). Continuous cross-training. Comprehensive documentation.

Model quality

Risk: AI output errors in safety-critical applications. Mitigation: Mandatory human review. Error logging. Pre-event research identifies model weaknesses before lives depend on output.

Social and political risks

Risk: Concentrated intellectual capability generates resentment or legitimacy challenges during governance transition. Mitigation: Pre-established charter, transparent governance, distributed network access, public reporting.

---

CROSS-REFERENCES

Document	Title	Relationship
Working Paper	Recoverable	Strategic vision; this document provides operational detail
Doc #135	Computer Construction	Long-term computing replacement (Phase 7)
Doc #127	Telecommunications Maintenance	Domestic network for distributed access
Doc #128	HF Radio	Trans-oceanic communication channel
Doc #130	Device Life Extension	Hardware longevity strategies
Doc #162	University and Research Reorientation	Research/education framework
Doc #8	Skills Census	Operator and hardware identification
Doc #53	Fuel Allocation	Generator and air transport fuel
Doc #57	Biodiesel Production	Renewable backup fuel
Doc #65	Hydroelectric Maintenance	Grid reliability
Doc #67	Transpower Grid Operations	National grid
Doc #5	Printing and Document Supply	Physical document production
Doc #29	National Printing Plan	Print program coordination
Doc #4	Medical Supply	Clinical decision support context
Doc #116	Pharmaceutical Rationing	Clinical decision support context
Doc #138	Sailing Vessel Design	Maritime distribution transport

---

APPENDIX A: MINIMUM VIABLE INFERENCE SETUP

For Tier 2 installations with limited hardware:

Hardware: 1x NVIDIA A100-80GB GPU (or 2x RTX 4090); 64+ GB system RAM; 1 TB SSD + backup SSD with model weights; UPS.

Software: Ubuntu Linux; CUDA drivers; llama.cpp or vLLM; 70B model weights (4-bit quant, ~35 GB); web chat interface.

Environment: 18–27°C; adequate ventilation; clean, dry; physically secured.

Power: 500–800 W continuous; 1,500 W peak; 1,000 VA UPS minimum.

Personnel: 1 trained operator + 1 backup + electronics repair access.

This setup runs in any building with reliable electricity. The central facility generates customized installation guides for each recipient’s specific hardware, building, and power supply. A single-GPU setup running a 70B model provides competent AI assistance for medical reference, engineering, document generation, and translation — not equivalent to the central facility, but far better than no AI capability at all.

---

---

1. Training costs for frontier LLMs estimated at US$50–100+ million. GPT-4 reportedly used ~25,000 A100 GPUs for months. See Epoch AI (https://epochai.org) for compute trend estimates.↩︎

2. Component lifespan data from MIL-HDBK-217F and industry reliability studies (Vishwanath & Nagappan, “Characterizing Cloud Computing Hardware Reliability,” ACM SoCC 2010). Server failure rates: 2–10%/year, increasing with age.↩︎

3. Hub-and-spoke architecture based on findings from Recoverable Foundation, “Preliminary Findings: Southern Hemisphere AI Infrastructure & Implications for the New Zealand Inference Center Proposal,” Working Document, February 2026.↩︎

4. Defence analysts have framed computational infrastructure as a domain of warfare. Data centres are recognised as critical national security assets. Destroying AI compute is permanent elimination of capability when the semiconductor manufacturing base (concentrated in Taiwan, South Korea, US, Japan) is itself destroyed. See findings document, Section 3.↩︎

5. NZ data centre market: approximately USD 0.89 billion (2025), total IT load capacity 432 MW projected to reach 591 MW by 2030. AWS and Microsoft NZ regions are enterprise cloud platforms, not GPU inference clusters. See findings document, Section 2.6.↩︎

6. MBIE Energy in New Zealand report. NZ installed capacity ~9,500 MW; average demand ~4,500–5,000 MW; annual generation ~43,000 GWh; renewable share 82–87%. See https://www.mbie.govt.nz/building-and-energy/energy-and-n...↩︎

7. Major NZ colocation providers include Datacom (Auckland, Wellington, Hamilton, Christchurch), Spark NZ (multiple locations), and CDC Data Centres (multiple Auckland and Wellington sites). Total NZ colocation capacity is modest by global standards. See individual provider websites and Cloudscene NZ data centre directory.↩︎

8. Open-source weights published by Meta (Llama), Mistral AI, DeepSeek, Alibaba (Qwen), and others. Sizes: Llama 3.1 70B ~35 GB (4-bit quant); 405B ~200 GB (4-bit quant). Freely downloadable from https://huggingface.co.↩︎

9. Uptime Institute Tier Classification and ASHRAE TC 9.9 guidelines. Tier III/IV design provides redundancy appropriate for decades of operation with limited replacement access. See https://uptimeinstitute.com/tiers.↩︎

10. US CLOUD Act (2018) compels US companies to provide data stored abroad. China’s Data Security Law (2021) creates analogous concerns. EU Schrems II ruling (CJEU C-311/18, 2020) invalidated EU-US Privacy Shield. See https://www.justice.gov/dag/cloud-act.↩︎

11. NZ Nuclear Free Zone, Disarmament, and Arms Control Act 1987 led to effective suspension from ANZUS military cooperation.↩︎

12. NZ ranks: Transparency International CPI 2nd globally (2023); World Justice Project Rule of Law Index top 10. See https://www.transparency.org/cpi; https://worldjusticeproject.org/rule-of-law-index.↩︎

13. Privacy Act 2020 (NZ). See https://www.privacy.org.nz/privacy-act-2020/.↩︎

14. Representative 2025–2026 pricing: GPT-4o US$2.50–10/M tokens; Claude 3.5 Sonnet US$3–15/M tokens; smaller models US$0.10–1/M tokens. See provider pricing pages.↩︎

15. Hub facility NZ$10–50M; distributed spoke network NZ$5–20M. Total investment within philanthropic reach. Commercial data sovereignty revenue sustains ongoing operations. See findings document, Section 7.↩︎

16. Industry estimates: US$7–25 million per MW IT load for purpose-built data centres. See JLL “Data Centre Outlook” reports.↩︎

17. Per-document estimate: 2–8 person-weeks by human experts. For 174 documents: 348–1,392 person-weeks (7–28 person-years). Based on actual Recovery Library production experience.↩︎

18. Operational cost 5–15 person-years/year (commercially staffed facility); output value 10–30 person-years equivalent specialist labor/year for NZ. Global value additional and difficult to quantify.↩︎

19. Staffing estimates derived from comparable commercial AI inference operations and small-scale data centre staffing norms. AI/ML engineer salaries in NZ (2025): NZ$100,000–180,000/year; data centre technicians: NZ$60,000–100,000/year; power engineers: NZ$90,000–140,000/year. FTE ranges reflect facility scale uncertainty (small colocation vs. purpose-built facility). Staffing costs are commercially justified pre-event; the incremental recovery-mandate cost is effectively zero.↩︎

20. The argument that maintaining AI capability is worth the cost rests on utilisation — the facility must be actively used, queries must be acted upon, and human reviewers must be integrated into decision workflows. A facility that operates but whose output is ignored provides no value. The governance framework (priority tiers, access infrastructure, operator training) is the mechanism that translates raw capability into recovered value.↩︎

21. Post-event operating costs estimated at pre-event dollar terms for tractability. In a recovery economy, the relevant unit is labor hours and energy, not currency. Electricity at 500 kW–1 MW = 4,380–8,760 MWh/year; at NZ pre-event rates (~NZ$120/MWh commercial), NZ$525,000–1,050,000/year. Labor (5–6 FTE at average NZ$120,000): NZ$600,000–720,000/year. Maintenance consumables and spare parts: NZ$100,000–300,000/year. These are pre-event dollar estimates; post-event, the correct accounting is opportunity cost in scarce labor and energy.↩︎

22. Infrastructure framing: AI inference capability is analogous to telecommunications or grid maintenance in recovery economics — it is a enabling layer whose value is distributed across all sectors that depend on it rather than concentrated in its own outputs. The decision to maintain it should be treated as an infrastructure decision, not a cost-benefit calculation on direct outputs alone. Cf. Doc #127 (Telecommunications) and Doc #65 (Hydroelectric Maintenance) for analogous infrastructure-tier arguments.↩︎

23. Stats NZ estimated resident population: approximately 5.2 million (2025). See https://www.stats.govt.nz/.↩︎

24. The net specialist-labor calculation depends critically on utilisation. If the facility serves 100 medical queries per day, each saving an average of 2 hours of specialist consultation, the annual specialist-time savings is ~73,000 hours (~40 FTE-equivalents). Against 4–7 FTE operating cost, the multiplier is 6–10x. This is plausible at full post-event utilisation and assumes the priority framework routes queries effectively. Lower utilisation or ineffective routing narrows the multiplier; higher utilisation expands it.↩︎

25. NZ grid electricity consumption (2022–23): approximately 42,000–44,000 GWh/year. Facility at 1 MW continuous = 8,760 MWh/year = ~0.02% of annual generation. MBIE Energy in New Zealand report. Under degraded-grid conditions — say 30% reduction in hydro generation due to altered precipitation patterns — available electricity drops to ~30,000 GWh/year and the facility’s share rises to ~0.03%. Still small in absolute terms, but the priority framework must be established before the grid is under pressure, not improvised during an emergency.↩︎

26. Hub-and-spoke architecture based on findings from Recoverable Foundation, “Preliminary Findings: Southern Hemisphere AI Infrastructure & Implications for the New Zealand Inference Center Proposal,” Working Document, February 2026.↩︎

27. Hub-and-spoke architecture based on findings from Recoverable Foundation, “Preliminary Findings: Southern Hemisphere AI Infrastructure & Implications for the New Zealand Inference Center Proposal,” Working Document, February 2026.↩︎

28. Wellington: near Hikurangi subduction zone and multiple active faults. Canterbury: experienced 2010–2011 sequence but generally lower ongoing hazard. Otago/Southland: lower seismic hazard. See GNS Science, https://www.gns.cri.nz/.↩︎

29. Uptime Institute Tier Classification and ASHRAE TC 9.9 guidelines. Tier III/IV design provides redundancy appropriate for decades of operation with limited replacement access. See https://uptimeinstitute.com/tiers.↩︎

30. Current-generation compact inference devices with large unified memory can run models in the 400B+ parameter range at 100–150 W. Requires no special cooling; operates from a modest solar installation with battery storage.↩︎

31. As of early 2026, Apple Mac Studio with M2 Ultra offers up to 192 GB unified memory. Apple Mac Pro with M2 Ultra is similar. These devices can run quantized 70B models at 5–15 tokens/second. Larger unified-memory configurations (256 GB+) may become available with future Apple Silicon or competing architectures. See Apple technical specifications and llama.cpp Apple Silicon benchmarks.↩︎

32. Data centres can be destroyed through non-nuclear means: conventional cruise missiles, high-altitude EMP (E3 effects on transformers feeding grid-connected facilities), or cyber warfare. Hyperscale GPU superclusters fail catastrophically — cutting cooling causes throttle within minutes. See findings document, Sections 3.4, 4.2.↩︎

33. Food production under nuclear winter: Doc #74. NZ feeds ~40M normally; under 30–50% pasture reduction, sufficient for 5.2M population with reduced surplus.↩︎

34. Defence analysts have framed computational infrastructure as a domain of warfare. Data centres are recognised as critical national security assets. Destroying AI compute is permanent elimination of capability when the semiconductor manufacturing base (concentrated in Taiwan, South Korea, US, Japan) is itself destroyed. See findings document, Section 3.↩︎

35. Australian AI defence integration: OpenAI/NextDC “sovereign AI” under SOCI framework; Google Christmas Island data centre for Department of Defence; AWS “Top Secret” cloud; AUKUS integration. Project Southgate in Melbourne: 18,500 NVIDIA GB300 GPUs. Australia data centre capacity ~1.3 GW (2025). See findings document, Sections 2.1, 3.2.↩︎

36. Data centres can be destroyed through non-nuclear means: conventional cruise missiles, high-altitude EMP (E3 effects on transformers feeding grid-connected facilities), or cyber warfare. Hyperscale GPU superclusters fail catastrophically — cutting cooling causes throttle within minutes. See findings document, Sections 3.4, 4.2.↩︎

37. Non-target status design requirements: no defence contracts whatsoever; local or foundation ownership; transparent operations open to inspection; international governance. Positioned alongside Svalbard Global Seed Vault, ICRC. See findings document, Section 5.1.↩︎

38. NZ data centre market: approximately USD 0.89 billion (2025), total IT load capacity 432 MW projected to reach 591 MW by 2030. AWS and Microsoft NZ regions are enterprise cloud platforms, not GPU inference clusters. See findings document, Section 2.6.↩︎

39. Alternative locations: Brazil ~707 MW colocation, 95% of Latin America’s market, concentrated in São Paulo. Argentina Patagonia Stargate (OpenAI/Sur Energy, up to 500 MW) at Letter of Intent stage. Africa <1% global capacity. Southeast Asia near South China Sea flashpoints. See findings document, Sections 2, 8.↩︎

40. Three differentiating pillars — non-target status, architectural resilience, purpose and readiness — from findings document, Section 10.↩︎

41. Commercial data centre degradation without resupply depends on spare parts availability, maintenance team retention, and power continuity. Facilities without pre-positioned spares would begin losing capacity as fans fail (3–5 years) and power supplies degrade (5–8 years). Facilities dependent on imported diesel for backup power or on municipal water for cooling would face earlier constraints. Estimate based on component lifespan data in footnotes 28–33.↩︎

42. NZ domestic fibre latency: 5–20 ms between major centres. See Chorus and REANNZ measurements.↩︎

43. Approximate figures from llama.cpp and vLLM benchmarks. See https://vllm.readthedocs.io/; https://github.com/ggerganov/llama.cpp.↩︎

44. Open-source weights published by Meta (Llama), Mistral AI, DeepSeek, Alibaba (Qwen), and others. Sizes: Llama 3.1 70B ~35 GB (4-bit quant); 405B ~200 GB (4-bit quant). Freely downloadable from https://huggingface.co.↩︎

45. Open-source landscape changes rapidly. As of early 2026: Llama 3.1, DeepSeek-V3/R1, Qwen 2.5, Mistral. Cache multiple models for redundancy.↩︎

46. ~600 GB–1.5 TB for comprehensive model library. Enterprise SSDs: 2–8 TB each. Trivial at commercial scale.↩︎

47. vLLM: https://github.com/vllm-project/vllm; llama.cpp: https://github.com/ggerganov/llama.cpp; text-generation-inference: https://github.com/huggingface/text-generation-inference; Ollama: https://ollama.ai. All open-source.↩︎

48. Current-generation compact inference devices with large unified memory can run models in the 400B+ parameter range at 100–150 W. Requires no special cooling; operates from a modest solar installation with battery storage.↩︎

49. 1 TB USB 3.0 drives: ~US$60–100 retail, lower in bulk. NAND flash unpowered retention at 25°C: years to over a decade. See JEDEC JESD218.↩︎

50. Distribution logistics for a global USB library programme include procurement, content loading, international shipping, customs clearance, and last-mile placement. Per-unit loaded and shipped cost is likely NZ$20–50 including logistics overhead — substantially more than the bare drive cost. Estimate is speculative; no comparable programme exists at this scale.↩︎

51. Global smartphone installed base estimated at ~6.5 billion devices (Statista, 2025). Survival rates in non-targeted regions depend on EMP exposure, power availability for charging, and physical damage. Southern Hemisphere devices far from detonation sites would largely survive physically, but become non-functional without grid power for charging. The useful surviving fleet is highly uncertain.↩︎

52. Models of 1B–3B parameters run on current smartphones (Apple A17/A18, Snapdragon 8 Gen 3) via llama.cpp, MLC-LLM, or manufacturer implementations.↩︎

53. Aviation fuel: NZ imports refined fuel. Rationed supply estimated at several years. 5–10 year window is an estimate. See Doc #53.↩︎

54. Tasman crossing: ~2,200 km; 5–7 knot cruising = 8–15 days. See Doc #138, Doc #142.↩︎

55. AGI coordination role: the facility’s value increases nonlinearly with model capability. Build now for current models, architecturally ready for AGI. See findings document, Section 6.5.↩︎

56. Capacitor temperature dependence from Nichicon, Rubycon, Panasonic datasheets. Arrhenius relationship.↩︎

57. BGA solder fatigue: Lau, J.H., “Ball Grid Array Technology,” McGraw-Hill, 1995. Lead-free solders more susceptible.↩︎

58. Fan bearing life: Nidec, Delta, Sanyo Denki specs. Sleeve bearing: 30,000–50,000 hrs; ball bearing: 50,000–70,000 hrs continuous.↩︎

59. Schroeder et al., “DRAM Errors in the Wild,” ACM SIGMETRICS 2009. ~1/3 of machines: ≥1 correctable error/year.↩︎

60. Schroeder et al., “Flash Reliability in Production,” USENIX FAST 2016. Enterprise SSDs: lower error rates; read-heavy workloads minimize wear.↩︎

61. Synthesised from sources in footnotes 28–32 plus general reliability references. Purpose-built facility estimates at upper end of ranges.↩︎

62. NIWA Climate Summaries. Auckland: 15.4°C mean; Wellington: 13.0°C; Christchurch: 12.1°C. Free cooling >7,000–8,000 hrs/year.↩︎

63. NZ has no domestic semiconductor manufacturing, no server-grade power supply production, and no precision fan manufacturing. All spare parts must be imported pre-event. UPS battery replacement intervals: Schneider Electric / APC technical documentation. Thermal paste reapplication: Intel and AMD thermal management guidelines.↩︎

64. MBIE Energy in New Zealand report. NZ installed capacity ~9,500 MW; average demand ~4,500–5,000 MW; annual generation ~43,000 GWh; renewable share 82–87%. See https://www.mbie.govt.nz/building-and-energy/energy-and-n...↩︎

65. Robock et al., “Nuclear Winter Revisited,” JGR 2007. 5–8°C global cooling. NZ maritime climate moderates; data centre cooling loads reduced substantially.↩︎

66. NIWA Climate Summaries. Auckland: 15.4°C mean; Wellington: 13.0°C; Christchurch: 12.1°C. Free cooling >7,000–8,000 hrs/year.↩︎

67. NVIDIA data sheets: A100 TDP 300W, H100 TDP 700W. Total server power ~1.5–2x GPU power for inference. See Jouppi et al., “In-Datacenter Performance Analysis of a Tensor Processing Unit,” ISCA 2017.↩︎

68. Medical Council of NZ Workforce Survey. ~17,000–19,000 registered doctors. See https://www.mcnz.org.nz/.↩︎

69. Degradation curve based on component data from Hardware Lifespan section. Significant uncertainty — treat as planning guidance. Purpose-built facility extends curve vs. repurposed facilities.↩︎

70. NAND flash unpowered retention at 25°C: years to over a decade. Older technology nodes retain longer. See JEDEC JESD218, JESD47.↩︎

71. Acid-free paper longevity: Library of Congress preservation guidelines estimate centuries under controlled conditions. Standard wood-pulp paper degrades significantly within 50–100 years due to acid hydrolysis. See Library of Congress, “Care, Handling, and Storage of Books,” https://www.loc.gov/preservation/.↩︎