Status: a proposed framework offered for public discussion — not the policy of any party, and not the position of any organisation. It is an AI-generated demonstration, produced by a simulated deliberation: twenty-eight fictional composite members, written from real sources, arguing across ten policy teams inside a working deliberative platform. Every factual claim in it carries a source reference [S-n] to a published register; every figure New Zealand does not have is declared as a gap [G-n] rather than invented; every position was tested at a decision point, cross-checked against the others, and ratified by an anonymous all-member vote whose records were cryptographically sealed and can be independently verified. The document is written as a neutral proposed framework, not in any individual’s voice. How the deliberation worked, and how to check it, is set out in the provenance annex (section 14).
New Zealand does not have an AI policy. Neither, for the most part, do the parties asking to govern it. New Zealand was the last country in the OECD to publish a national AI strategy, and the one it finally got is a strategy for adopting other people’s AI, not for holding any authority over it [S-10]. Meanwhile the technology has already moved into the country’s hospitals, classrooms, councils and workplaces, mostly on terms written somewhere else.
AI is the clearest test yet of whether New Zealand’s politics can lift its attention from the electoral cycle to the country’s foundations. It is not a left-versus-right question. It is a foundations question: who holds authority over the systems that will run public services, what happens to information about New Zealanders, whether people move up or get moved out, and whether the electricity and water these machines consume are planned for or discovered after the fact.
Foundations questions deserve a particular way of being answered: from evidence, in the open, and without pretending New Zealand is a bigger country than it is. This framework answers them that way. There is no sovereign frontier laboratory in here, no fantasy chip factory, no invented numbers. There are instruments that existing New Zealand machinery can carry this term, gaps stated plainly where the data does not exist, and — where the framework departs from the emerging cross-party consensus — recorded reasons, on the record.
There is also something rarer: a policy document that shows its working, cryptographically. Every position in this framework survived argument between participants who started apart. One team’s plan was blocked by another team and rewritten. One team’s conclusion was sent back by the full membership and strengthened. Those collisions are not embarrassments to be edited out; they are the reason to trust the result — and every one of them sits in a sealed record that any reader can verify independently of the platform that produced it (§14). Call it what it is: policy you can audit. No pamphlet, manifesto or ministry strategy in this country currently offers that. This document is offered to every party, and to everyone outside the parties, as a demonstration that it can be offered — and an invitation to check.
New Zealand was built on the courage to change. The machines are here. The only question is whose terms they run on — and that is a question a modern, evidence-based country can answer for itself.
Good policy starts with an inventory, not an ambition. Here is New Zealand’s, in three columns: what the country has, what it depends on, and what it simply does not know.
What New Zealand has is better than the anxiety suggests. The country’s electricity is the envy of most of the developed world: 43,879 GWh generated in 2024 at 85.5 per cent renewable, with the December 2025 quarter setting a record of 96.4 per cent [S-11]. Its privacy law is technology-neutral and already reaches AI across its whole lifecycle — the Privacy Act 2020 and its thirteen information privacy principles, with the Privacy Commissioner having twice issued AI-specific guidance on impact assessments, transparency, Māori engagement, human review and data minimisation [S-16]. The Algorithm Charter (2020) has six commitments and more than twenty signatory agencies [S-17]. The Data and Statistics Act 2022 gives system-level footing for official data [S-18]. And on Māori data, the hardest design work is already done and Māori-led: the Māori Data Governance Model, co-designed by Te Kāhui Raraunga with sixteen Crown agencies in 2023, with its eight Pou and its proposal for a Māori Chief Data Steward [S-22].
What New Zealand depends on is mostly offshore. The country’s first hyperscale cloud region, opened December 2024, is owned by a US corporation through an Irish subsidiary — the racks are here, the control is not [S-15]. In June 2026 a single overseas legal order disrupted access to a major AI tool worldwide; New Zealand was simply downstream [S-1]. Every GPU and AI accelerator in the country is imported from a fabrication base concentrated in East Asia; New Zealand makes none [S-21]. Forty-five per cent of the tech workforce is on work visas, while domestic digital-technology enrolments have fallen 33 per cent since 2010 [S-19]. And at the human surface, 48 per cent of workers have entered sensitive information into public AI tools, and 56 per cent have used AI at work unsure whether they were allowed [S-1].
What New Zealand does not know is itself a finding. There is no AI-specific procurement rule — the Government Procurement Rules are technology-neutral and carry no jurisdiction, residency or auditability test aimed at AI [S-23][G-6]. Transpower has stated that its load forecasts do not include any hyperscale data centres, even as roughly 32 GW of prospective load queues for connection and the north-Auckland grid faces exit-point overload by 2028 without upgrades [S-12][G-8]. There is no per-facility water-draw disclosure standard for data centres; the only public number in the country is one Southland consent artefact [S-14][G-7]. Nobody counts AI specialists as distinct from ICT generally, and nobody measures AI-attributable job displacement here [S-19][G-5]. There is no NZ-specific quantified AI threat metric [G-4], no sovereign frontier programme of any kind [S-10][G-3], no zetta- or exascale compute — the national research ceiling is a single ageing machine at roughly 1.4 petaFLOPS [S-20][G-2] — and no domestic chip fabrication, with none plausible at the country’s scale [S-21][G-1].
That inventory sets the shape of everything that follows. Where the machinery exists, this framework binds it. Where the data does not exist, the first instrument is the one that creates it. And where the gap is permanent, the gap is stated plainly and designed around. A country that knows exactly what it holds, owes and lacks is a country that can modernise its foundations rather than patch them.
In June 2026 an independent cross-party proposal, In Our Own Hands, was offered to every party in Parliament: seven commitments and a three-phase pathway for AI under New Zealand authority [S-1]. Where a consensus is sound, the evidenced move is to join it and say so; where the evidence points elsewhere, the difference belongs on the record with its reasons attached. The deliberation tested this framework against all seven commitments. The result: alignment with every one, extension of four, and difference — with recorded reasons from the deliberation record — in two, plus a third, catchment-level difference on water that no single commitment addresses (§7).
| # | Cross-party commitment [S-1] | Position | How |
|---|---|---|---|
| 1 | Authority stays here — New Zealanders keep rightful authority over the AI that runs public services and holds their data | Align + extend | Adopts the procurement test pair, register and audit standard; extends the register to local government from day one, because councils across the country buy algorithmic systems outside the Procurement Rules’ reach [S-23] (§3) |
| 2 | People move up, not out — AI lifts people into better work | Align + extend | Adopts the redeployment fund; extends it with a workforce-transition disclosure standard and a conditioned employer training levy-credit [S-19] (§5) |
| 3 | Sensitive information stays home — it should not have to leave the country, or its legal protection, to be useful | Align + differ on interpretation | Adopts the class-based handling standard and explicitly rejects any blanket residency mandate — the deliberation voted one down. A right to domestic processing, not a wall; with a protected lawful research pathway [S-16][S-18] (§6) |
| 4 | Te Tiriti and Māori data sovereignty — Māori data is a taonga; obligations met in how public AI is built and bought | Align + extend substantially | Goes beyond the commitment: a statutory Māori Chief Data Steward AND appropriated multi-year implementation of the eight Pou, drafted as a non-severable pair, with Māori data classification binding the handling standard [S-22] (§4) |
| 5 | People decide, machines assist — decisions that materially affect people stay with accountable humans | Align | Adopted as drafted: every automated decision path ends in proceed, refuse, or escalate to a human [S-1]; carried into the audit standard (§3) |
| 6 | What public AI does is checkable — an auditable record of what it decided and why | Align + extend | Adopts the provenance-and-audit standard; extends it with a usability test — writable by a council with three IT staff, not just a ministry with a chief data officer (§3) |
| 7 | We build our own capability, with others | Align + differ on conditions | Adopts the capability fund and partnership posture; differs by binding every owned compute stage to grid milestones and evidence gates with named stop conditions — capability that the transmission planner cannot see is not capability, it is a stranded asset in waiting [S-12] (§7, §8) |
The three differences, stated as positions rather than footnotes:
First, the physical ledger is a precondition, not an afterthought. The cross-party proposal builds capability in Phase 2 without conditioning it on electricity and water reality. The deliberation blocked exactly that move internally — the grid team refused the compute team’s unconditioned staging, and the staging was rewritten (§7). This framework carries that discipline into national policy: no owned compute stage proceeds without disclosure into Transpower’s forecast, interruptibility contracts, and siting consistent with the 2028 constraint map [S-12][S-11].
Second, water cannot be governed by national milestones at all. The freshwater team held — and the full record preserves — that catchment is the unit of analysis: 220 million litres a year is trivial nationally and potentially significant against one Southland aquifer [S-14]. National frameworks, including the cross-party one, need a catchment-level instrument class they currently lack (§7).
Third, residency absolutism is rejected in terms. Commitment 3’s careful wording — should not have to leave — is the wording this framework defends, against harder readings. Health research runs on lawful, ethics-approved international collaboration [S-16][S-18]; a blanket onshore mandate would relocate the research offshore along with the researchers. Classes and safeguards, not walls (§6).
On the three phases — ground rules now, capability years one to three, govern-federate-sustain after — this framework adopts the sequence and the select-committee destination [S-1], with the extensions set out in section 12.
Who holds authority over public-sector AI today? On paper, several instruments; in practice, nobody. The Algorithm Charter is a signed commitment with self-reported compliance and no enforcement mechanism [S-17]. The Public Service AI Framework is explicitly advisory [S-24]. The Procurement Rules are technology-neutral, with no AI-specific test of any kind [S-23][G-6]. Nothing happens when an agency ignores any of them. Voluntary instruments have had six years to deliver checkability; they have delivered signatures.
The deliberation’s answer is a hybrid — chosen over both a new Crown entity now (too slow, and a machinery-of-government fight before a single system is checked) and voluntary-plus (six years of evidence against it [S-17]):
1. Binding procurement tests, immediately. MBIE amends the Government Procurement Rules — an executive-level act, done before without legislation [S-23] — to insert the cross-party test pair: jurisdiction of inference (where does the AI actually run, and whose court order does it answer?) and data residency [S-1]. This is deliverable inside a parliamentary term, and it converts procurement from a purchasing function into the enforcement lever for everything else in this document: the data-handling standard (§6), the Māori data classification wiring (§4), and the audit standard below all bind at the point of purchase, where a system that fails the test is never bought.
2. A public register of significant government AI — including councils, from day one. Central agencies first, but with local government included from the start, initially by invitation. Councils up and down the country procure algorithmic systems — consenting triage, CCTV analytics, regulatory decision support — outside both the Charter’s signatory list [S-17] and the Procurement Rules’ mandate [S-23]. A register that stops at the Wellington boundary leaves the public meeting unaudited AI at exactly the level of government closest to their daily lives. The register begins administratively, as a published schedule under the existing Public Service AI Framework mandate [S-24].
3. An audit standard: a checkable record of what public systems decided and why [S-1] — with a usability test attached. The standard must be writable by a council with three IT staff, not only by a ministry with a chief data officer. The Charter worked where it was simple — a contact point, an appeal pathway [S-17] — and stalled where it demanded capability agencies did not have. Human decision-authority is carried inside this standard: every automated decision path that materially affects a person ends in one of three outcomes — proceed, refuse, or escalate to an accountable human [S-1].
4. A statutory authority decided by Parliament, not asserted in this framework. Registers maintained by circular are abandoned by reshuffle, so the destination is statutory — but the shape belongs to a select-committee inquiry, with the Economic Development, Science and Innovation Committee the natural host [S-1]. If an authority is established, it carries a support function for small agencies and councils, not only an inspection function, and Māori governance membership as of right (§4).
Rules first, Parliament decides the body. That is authority you can hold this term, with the durable architecture built by the institution designed to build it.
Proviso. This chapter draws on Te Kāhui Raraunga’s own published Māori Data Governance Model [S-22]. It is offered in that spirit — as a contribution for their review and for Māori-data leadership to shape, not as a settled position. Māori data sovereignty is for Māori to determine; this framework’s role is to make space for it in the architecture, not to define it.
This chapter sits early in the document because its subject sits early in the machinery. If te Tiriti obligations are met “in how public AI is built and bought” [S-1], then Māori data governance is not a consultation paragraph appended after decisions — it is upstream wiring, and it has to be installed before the rest of the system runs.
The foundation already exists, and it is Māori-led. The Māori Data Governance Model — co-designed by Te Kāhui Raraunga with sixteen Crown agencies in 2023 — sets out eight Pou spanning workforce, infrastructure, collection, protection, access and sharing, use and reuse, quality and integrity, and classification, and proposes a Māori Chief Data Steward [S-22]. The Crown’s favourite delay — “we must first develop a framework” — is unavailable. The framework is developed, co-designed, and signed. It is also, on the evidence available, largely unimplemented across signatory agencies. That gap is what this chapter learns from: endorsement at the top, without appropriated implementation underneath, has moved little.
The deliberation ran the real argument — authority first, or implementation first — and the Model’s own biography proved both sides half-right. An office without implementation is a flag on an empty building; funding without authority evaporates at the third budget cycle, because nobody with binding standing is positioned to contest the lapse. So the position is a pair, and the drafting makes it non-severable:
1. A Māori Chief Data Steward with statutory basis — the office the Model itself proposes [S-22] — in the room as of right, not by invitation, with standing in the AI authority architecture this document builds (§3).
2. Appropriated, multi-year funding for eight-Pou implementation across the sixteen signatory agencies, including direct funding for data infrastructure that iwi themselves control — appropriated funding, not contestable pilot pools that exhaust iwi capacity in grant-writing [S-22].
Non-severable means what it says: no future budget may deliver the office and defer the money “pending fiscal conditions”, or fund the Pou while leaving the Steward on memorandum legs. The Model’s unimplemented history is the proof that halves do not work [S-22].
Two pieces of structural wiring complete the chapter:
Classification before residency. The Model’s classification Pou determines what Māori data is, whose it is, and what protection travels with it — before any question of where it may go [S-22]. This framework’s data-handling standard (§6) therefore takes Māori data classification from the Model as a binding input, not advisory context. Follow the wiring: the handling standard governs procurement; procurement governs what public AI gets bought; so Māori-defined classification sits upstream of every public AI purchase in the country. That is “built and bought” [S-1] made mechanical.
Co-design standing from the start. The select-committee inquiry (§3) carries Māori co-design standing in the inquiry stage itself — not “the inquiry will consider representation” — and any statutory authority carries Māori governance membership as of right, through the Steward’s office or direct appointment [S-1][S-22]. The precedent is the Crown’s own signature on the Model. The ask is not for something new; it is for the Crown to inhabit what it already signed.
The second cross-party commitment is the one most New Zealanders will judge any AI policy by: people move up, not out — AI takes the drudgery and leaves the judgement [S-1]. The deliberation held itself to a discipline worth stating: it argued only from data New Zealand actually has, and refused to invent the data it lacks.
What New Zealand has: roughly 98,290 ICT professionals as at 2023, growing 3.4 per cent a year; 45 per cent of the tech workforce on work visas; domestic digital-technology enrolments down 33 per cent since 2010; and an acute shortage at the senior end — AI, cyber and data science — rather than a raw junior shortage [S-19]. The system produces growth without producing New Zealanders to staff it. What it lacks: any count of AI specialists as distinct from ICT generally, and any NZ dataset on AI-attributable displacement [G-5]. So this framework makes no numeric training guarantee and no displacement forecast; either would be invented.
Four instruments, in order:
1. A workforce-transition disclosure standard — significant employers and agencies adopting AI at scale report role changes: redeployments, exits, new roles. This is the gap-targeting instrument [G-5]: in three years, the next version of this policy argues from evidence rather than anecdote.
2. A redeployment fund on the cross-party pattern — funding that moves people into better work as AI adoption changes roles, not compensation for moving them out [S-1].
3. A conditioned employer training levy-credit. The 45 per cent visa figure [S-19] is employers solving their problem the fast way — importing skills is rational for each firm and hollowing for the country, a textbook collective-action failure. The credit changes the incentive at the firm, but pays out only against training that ends in a portable, recognised micro-credential delivered through the vocational network. Employers set the demand signal; polytechnics deliver credentials a worker can carry to any job; poaching stops paying. Aimed first at employed adults — the administrator whose routine tasks are automating, the technician who now supervises a system rather than performs the task — because that is where the transition risk actually sits [S-19]. This instrument existed in neither side’s opening position in the deliberation; the collision produced it.
4. The pipeline, rebuilt around how this workforce actually forms. The 33 per cent enrolment decline [S-19] is not evidence that institutional training is obsolete; it is evidence that three-year degrees front-loaded before employment no longer fit. Short, stacked, work-integrated micro-credential pathways for employed adults are the vehicle, with the senior-skills shortage [S-19] the priority target.
This chapter builds the ladder. Beneath it sits a question the framework names but does not settle: the income floor. A transition this fast tests the design of the whole welfare system, and whichever instrument a government prefers — benefit reform, a universal income, something else — that floor is a system-level choice, not an AI line-item. The deliberation’s position is narrower and holds either way: build the ladder from the evidence New Zealand has, create the evidence it lacks, and do not ask retraining policy to do the welfare system’s job.
New Zealand’s data-protection law is stronger than the AI debate gives it credit for. The Privacy Act 2020 is technology-neutral by design — the thirteen IPPs govern collection, use, disclosure and security wherever an AI sits in the lifecycle, generative or otherwise — and the Privacy Commissioner has already told agencies what good looks like: leadership sign-off, privacy impact assessments, transparency, engagement with Māori, human review of consequential outputs, data minimisation [S-16]. Above it sit the Data and Statistics Act 2022 [S-18] and the Algorithm Charter [S-17].
And yet: 48 per cent of workers have entered sensitive information into public AI tools, and 56 per cent have used AI at work unsure whether they were allowed [S-1]. All of that happened under the Act, under the guidance, under the Charter. The law did not fail on paper; it failed to reach the person at the keyboard on a Tuesday afternoon. That fact disciplines the design: whatever is proposed must be usable by a school office manager and a parish treasurer, not just by counsel.
The instrument is a class-based data-handling standard for externally-controlled AI, built in this order:
Classification first. Sort information into classes before writing rules about movement. For Māori data, the classification layer is not this framework’s to design — it exists, Māori-led, in the Māori Data Governance Model, and it binds this standard as an input (§4) [S-22].
Handling rules second. For each class: which destinations are permitted, under which safeguards, and which are not [S-1]. A rule a human can hold in their head — “this class of information never goes into that class of tool” — where a forty-page code of practice is not.
Enforcement third, and upstream. The standard applies through the Government Procurement Rules [S-23], because procurement binds before the data moves: the contract fails the test, the system is never bought. A Privacy Commissioner AI code of practice under the Act is the backstop — enforceable, complaint-driven, and the reach into private-sector conduct that procurement cannot touch [S-16]. Primary and backstop, in that order.
Two clauses the deliberation insisted on, and the draft keeps:
The research pathway. Ethics-approved, appropriately de-identified research data may move to partner institutions under the Privacy Act’s existing disclosure framework [S-16] and the system-level governance of the Data and Statistics Act [S-18], with the classification standard determining what de-identification each class requires — and Māori data taking its requirements from the Model, not from defaults [S-22]. Without this clause, the first casualty of a well-meant standard is the clinical collaboration that catches the next pandemic variant. The deliberation put a blanket “everything stays onshore” mandate to a vote and rejected it; the commitment is that data should not have to leave to be useful — a right to domestic processing, not a prohibition on movement [S-1].
The one-page worker standard. A plain usage standard for the 48 per cent problem [S-1]: what may go into which tools, on one page, in plain language, in every workplace that adopts AI. Cheap, fast, and aimed at the keyboard where the law currently is not.
Every AI ambition in this document draws electricity and, depending on cooling choices, water. This chapter is where those ambitions meet the meter — and it is the chapter where the deliberation visibly changed its own policy, so the story is told straight.
Electricity: the constraint is visibility and delivery, not energy. The generation picture is genuinely good — 85.5 per cent renewable in 2024, a record 96.4 per cent quarter in December 2025 [S-11]. But three facts from the system operator’s side reframe everything: Transpower’s load forecasts do not include any hyperscale data centres; without upgrades, north-Auckland grid exit points are overloaded by 2028; and a single hyperscale AI facility draws 300–600 MW continuously [S-12] — 3 to 6 per cent of the roughly 10.6 GW installed fleet [S-11], committed to one site around the clock, including the winter evening peak when hydro storage is the constraint and every flexible megawatt matters. Around 32 GW of prospective load is queued seeking connection — roughly three times today’s system [S-12]. The country’s transmission planner is flying without instruments on this load class [G-8]. The only forward demand estimate that exists is market research — 432 MW of data-centre IT load in 2025, rising to 591 MW by 2030 [S-13] — and this document flags it as market research: usable for scenario-shaping, never for commitment-sizing [G-8].
Demand growth is not the enemy of a renewable grid; firm, creditworthy, long-duration demand is what gets new generation financed, and the marginal electron for a large load here is among the cleanest on earth [S-11]. So the position is a gateway, not a wall — large compute loads are welcome if they:
What the deliberation changed. The compute team’s first staged expansion plan named stages and capacity with none of these conditions attached — proposed against a forecast that cannot see the load class at all [S-12]. The electricity team formally blocked it. The compute team accepted the block as correct and rewrote the staging: every stage’s go-decision now requires its load disclosed into Transpower’s forecast before commitment, interruptibility contracted for every workload class that tolerates it, and siting or scheduling consistent with the constraint map, with the 2028 northern constraint treated as binding until its upgrade programme delivers [S-12][S-11]. That is not a concession; it is a strengthening — a compute stage the grid has planned for is a stage that cannot be stranded in the connection queue. The block, the revision and the adoption are sealed in the deliberation record (§14).
Water: catchment by catchment, or not at all. The entire public evidence base for data-centre water draw in New Zealand is one consent: Datagrid’s Makarewa site in Southland — 280 MW, roughly NZ$2 billion, phase one targeted around 2028 — consented to draw 220 million litres of groundwater a year for cooling, with full consent still pending [S-14]. That number is public only because the consent process forced it out; no disclosure standard exists for any facility [G-7]. And 220 million litres is trivial against a national water budget while potentially significant against one aquifer’s sustainable yield — in most productive catchments there is no spare water, only allocated water, over-allocated water, and ecological minimum flows already under pressure [S-14]. Any policy sentence about data-centre water written nationally is a sentence about nothing. Meanwhile Microsoft’s Auckland region operates entirely water-free cooling [S-15] — proof, running an hour up the motorway from anywhere, that taking nothing is a real engineering choice.
Four instruments:
One disagreement kept rather than smoothed over. The water team holds that national milestones — the kind the compute and electricity coupling trades in — cannot answer a catchment-level question, and declined to fold that view into a consensus it did not share. The difference is recorded openly in the cross-check record as an acknowledged disagreement [S-14] (§14). A deliberation that can hold a disagreement without smoothing it is a deliberation whose agreements mean something.
Two permanent facts bound this chapter, and the policy is designed on top of them rather than in denial of them.
Fact one: New Zealand has no frontier compute, and will not. The national research ceiling is Māui — a Cray XC50 delivering about 1.4 petaFLOPS, commissioned around 2018 and ageing — alongside Mahuika, more than 33,500 cores between them, operated since 1 July 2025 under REANNZ after NeSI’s consolidation [S-20]. The international frontier sits three to four orders of magnitude beyond that ceiling [G-2]. This framework cites no FLOPS target it cannot source and implies no training-scale ambition the country cannot power or fund. Sovereign compute at New Zealand’s size means three things the country can actually own: custody (whose hands the machine sits in), access (terms for New Zealand researchers and firms), and jurisdiction (whose court order the inference answers to) [S-1][S-15].
The architecture is a floor plus a ladder:
The floor: an owned research baseline. Māui’s replacement under REANNZ is due on ordinary research-infrastructure grounds regardless of anything this policy says — climate modelling, genomics, national science capability [S-20] — sized with headroom for the workload classes the data-handling standard requires to stay home (§6). Custody unambiguous. A jurisdiction clause in a services contract does not survive a conflict with the supplier’s home-state law; the June 2026 disruption is the live exhibit [S-1][S-15]. Where custody is load-bearing — health data, justice, anything classified home-only — the workload runs on the owned baseline, full stop.
The ladder: subsidised access, tested. Above the floor, scale comes from access on the Canada/Singapore pattern — Canada’s compute access fund runs up to CA$300 million, Singapore’s Enterprise Compute Initiative up to S$150 million in credits and consultancy [S-1] — with every cloud component passing the jurisdiction-of-inference and data-residency procurement tests (§3). Owning tin is the expensive way to feel sovereign; accelerated hardware depreciates brutally while hyperscale operators refresh continuously. So: elasticity where custody is not load-bearing, owned capacity where it is.
Staging, grid-conditioned. Stage 1: the Māui replacement, its load disclosed into Transpower’s forecast at procurement [S-20][S-12]. Stage 2: a national AI research compute facility supporting the domain-capability work of section 10, grid-milestone-tied as section 7 requires [S-1][S-12]. Stage 3: demand-proven growth, inheriting the same conditions plus the capability fund’s review gates [S-1]. And a symmetry clause: the review gates evaluate both lines — is the owned baseline earning its capital, and is access dependency growing beyond what the custody routing rule permits? If the second answer trends wrong, that is a signal to accelerate the owned stages, not to shrug [S-1][S-15].
Fact two: New Zealand makes no chips, and never will. New Zealand has no domestic semiconductor fabrication — none operating, none planned, none plausible; the global fabrication base is concentrated in East Asia and the country is wholly import-dependent for GPUs and accelerators [S-21][G-1]. This framework excludes any domestic-fab programme as fantasy at the country’s scale, and manages the dependency as a repeating cycle, not a one-off purchase — because accelerators are three-to-five-year assets, and the full exposure re-runs at every refresh:
This position was strengthened by the deliberation’s own membership: the full-member ratification vote returned the first version because it treated a permanent dependency as a procurement event rather than a recurrence. The revision above is the answer to that return, and both the objection and the response are sealed in the record (§14).
This chapter refuses to inflate. There is no NZ-specific quantified AI threat metric, and this framework will not manufacture one to make its recommendations look load-bearing [G-4]. The discipline cuts both ways: no threat inflation to justify spend, and no complacency dressed as scepticism. The operative, evidenced risk is jurisdictional and dependency-shaped, not military: one overseas legal order disrupted a major AI tool worldwide in June 2026 [S-1]; the country’s only hyperscale region is US-owned through an Irish subsidiary [S-15]; every accelerator arrived by ship or plane from a concentrated fabrication base [S-21]; and the widest attack surface needs no data centre at all — 48 per cent of workers have put sensitive information into public AI tools, and synthetic media in an election year makes polluting what official information looks like the cheapest interference available [S-1].
Four layered instruments, the cheap and fast pair first:
The risk is dependency, not war; the fix is visibility plus fallback.
New Zealand’s official strategy, Investing with Confidence (July 2025), points at the right sectors — health, agriculture, education — but it is an adoption strategy, and it concedes its own ceiling: capability framed as adopting AI, a shortage of expertise acknowledged, infrastructure posture centred on foreign-owned hyperscale cloud [S-10]. Adoption matters — diffusion is where measured productivity gains show up, and in a low-productivity economy that push continues. But adoption-only has a structural problem: when everything the country adopts is built offshore, the margin, the capability and the exit options all sit offshore too.
The proposed posture: evidence-gated niche-builder within partnership.
Niche-builder, deliberately narrow. Not frontier-chasing — the starting stock is zero and this framework says so [G-3][G-2]. Domain capability: models and products fitted to problems New Zealand owns and accountable to people here [S-1], in two or three named domains where the country holds durable advantage — the natural candidates are agritech, health workflows, and te reo language technology. Denmark shows the shape: sovereign public-sector AI on home-or-EU-jurisdiction compute plus a regulatory sandbox, from a nation of comparable size, with nobody pretending to race the frontier [S-1].
Evidence-gated, without exception. Every funded domain carries a named review point and a stop condition. If the evidence turns, the programme stops — in public. Evidence-led policy has to mean being willing to be wrong in public about one’s own enthusiasms.
Within partnership, always. Small advanced economies do not choose between adopt and build; they pool. The fund’s compute line buys pooled access alongside partners rather than sovereign scale [S-1] (§8); the hardware line buys alongside Australia (§8); and the quiet lever is standards — New Zealand holds its seat at ISO/IEC JTC 1/SC 42 and resources it [S-1]. One alignment discipline throughout: New Zealand’s regulatory settings must not diverge from its trading partners’ regimes, because divergence is a non-tariff barrier the country would impose on its own exporters. Posture is also a foreign-policy instrument.
None of this requires a new philosophy of innovation spending. Research-and-development investment, technology investment credits and accelerator programmes are instruments already argued across New Zealand’s political spectrum; the posture proposed here is that existing machinery, applied with evidence gates attached. What it does require is the discipline above: narrow domains, named stop conditions, and partnership over pretence.
This framework publishes no fiscal number nobody has costed. What follows is the funding architecture, the sourced envelope, and a plain list of what remains for Treasury and MBIE to price.
Phase 1 is cheap by design. The ground rules — procurement tests, the register, the audit standard, the handling standard, the provenance standard — are regulatory instruments carried by existing machinery: MBIE’s Procurement Rules [S-23], the Government Chief Digital Officer’s (GCDO) framework mandate [S-24], the Privacy Commissioner’s code-making power [S-16]. Low cost is not a happy accident; it is why they come first [S-1].
Phase 2 carries the substantive spend, inside a sourced envelope. The cross-party proposal’s indicative sizing for the national AI capability fund is low tens of millions over three to four years, benchmarked against Canada’s compute access fund (up to CA$300 million) and Singapore’s Enterprise Compute Initiative (up to S$150 million), with the precise figure explicitly reserved for Treasury and MBIE [S-1]. This framework adopts that envelope and that reservation. The fund’s three uses — subsidised compute access, advisory support, and community/sector-scale model support [S-1] — map directly onto this framework’s ladder (§8), domain programme (§10) and redeployment architecture (§5). The deliberation ran a priority-weighting exercise across the fund’s resource strands — grid flexibility enablement, subsidised compute access, workforce pathways, and Māori data infrastructure under the eight Pou [S-12][S-1][S-19][S-22] — and that weighting is sealed in the record (§14).
Two items are explicitly appropriations bids, to be costed in government: the eight-Pou implementation funding (§4), which is non-severable from the Steward and therefore not deferrable, and the Māui replacement (§8), which is due on ordinary research-infrastructure grounds regardless of this policy [S-20]. The income floor under a fast workforce transition (§5) is a whole-of-welfare-system design choice and belongs to that debate; it is deliberately not an AI line-item here.
What this framework will not do is decorate this section with invented precision. Where a figure appears in this document it is sourced [S-1][S-11][S-12][S-14][S-19][S-20]; where none exists, the gap is declared (§13). A costing you cannot check is a hope, not a budget.
The cross-party proposal’s three phases are the right skeleton [S-1]. This framework adopts them and extends each, and holds one principle across all three: every step must be reversible until the evidence says otherwise — pilots stoppable, funded domains carrying stop conditions, and open standards over central registries so that no future government inherits a machine it cannot change [S-1].
Phase 1 — set the ground rules (now; regulatory; low cost). The procurement test pair into the Procurement Rules [S-1][S-23][G-6]. The public register begins administratively — councils included by invitation from day one (§3) [S-24]. The audit standard with the three-outcome human-authority rule [S-1]. The class-based data-handling standard, Māori classification binding [S-1][S-22]. The provenance standard and the one-page worker standard [S-1]. This framework’s extensions: the water-draw disclosure standard and catchment precedence rules enter here, not Phase 2 — disclosure instruments are ground rules [S-14][G-7]; the dependency register’s wave one is mapped within a quarter (§9); the non-severable Tiriti pair is introduced as legislation-plus-appropriation together (§4) [S-22].
Phase 2 — build capability (years one to three; substantive spend). The capability fund inside the sourced envelope (§11) [S-1]. The domain programme, two or three domains, each with review point and stop condition (§10). The compute floor and ladder, every owned stage grid-milestone-tied (§7, §8) [S-12][S-20]. Pooled accelerator procurement stood up; alliance talks opened within the first cycle (§8) [S-23][S-1]. The redeployment fund and the conditioned levy-credit (§5) [S-1][S-19]. This framework’s extension: the grid conditions are not Phase-2 decoration; they are prerequisites written into each stage’s go-decision — the deliberation blocked the alternative (§7).
Phase 3 — govern, federate, sustain. The select-committee inquiry decides the statutory authority’s shape, with Māori co-design standing built in (§3, §4) [S-1][S-22]. Open standards, not a central registry [S-1]. Standing international engagement at SC 42 [S-1]. Scheduled built-in review reporting to the committee — including the symmetry review of owned-versus-access compute dependency (§8) and the three-year re-examination of the workforce evidence the disclosure standard will by then have created (§5) [G-5].
Sequence is the argument. Rules that bind now, capability gated by evidence, institutions decided by Parliament — in that order, because each step creates the information the next one needs.
A policy that hides its ignorance ages badly. This register lists every gap this framework stands on — each one acknowledged during deliberation and sealed in the record, none papered over with an invented number. Where a gap can be closed, the closing instrument is named; where it is permanent, the design constant is named instead.
| # | Gap | Consequence | Response in this framework |
|---|---|---|---|
| G-1 | No domestic semiconductor fabrication — none operating, none planned, none plausible at New Zealand’s scale [S-21] | Permanent import dependency for all accelerators, re-run at every 3–5-year refresh | Design constant: pooled multi-cycle procurement, staggered refresh, alliance protocols, funded second-life tier (§8) |
| G-2 | No zetta- or exascale compute — the national research ceiling is ~1.4 petaFLOPS and ageing [S-20] | No credible frontier-training ambition; sovereign compute means custody, access, jurisdiction | Floor-plus-ladder architecture; no unsourced FLOPS targets anywhere in this framework (§8) |
| G-3 | No sovereign frontier-capability programme of any kind [S-10] | Year-one deliverables are institutional (fund design, domain selection, partnership terms), not technical | Evidence-gated niche-builder posture; frontier-chasing excluded (§10) |
| G-4 | No NZ-specific quantified AI adversary-threat metric | Neither threat inflation nor complacency can be evidenced | Jurisdictional-dependency framing only; instruments that create visibility (register, disclosure) rather than assume threat levels (§9) |
| G-5 | No AI-specialist headcount (only broad ICT aggregates) and no NZ displacement dataset [S-19] | No numeric training guarantee or displacement forecast can be made without inventing it | Workforce-transition disclosure standard creates the missing data for the next policy cycle (§5) |
| G-6 | No AI-specific procurement rule — the Rules are technology-neutral [S-23] | The state’s strongest lever over AI is currently unbuilt | The fast path: jurisdiction-of-inference + data-residency tests inserted this term (§3) |
| G-7 | No per-facility data-centre water-draw disclosure standard — one consent artefact is the entire public evidence base [S-14] | Councils allocate blind; catchment accounting impossible | Per-facility disclosure standard, published, feeding allocation accounting (§7) |
| G-8 | No granular, official AI-attributable electricity-demand data — the system operator’s forecasts exclude hyperscale entirely; the only forward estimate is market research [S-12][S-13] | The transmission planner cannot plan for the load class this policy enables | Mandatory load disclosure as a condition of connection, feeding Transpower forecasting directly (§7) |
Notice the pattern: five of the eight gaps are answered by disclosure standards — instruments whose first job is to create the evidence New Zealand currently lacks. That is deliberate. Where you cannot see, the first policy is sight.
This document was produced by a full-membership deliberation run inside a working deliberative platform, as a demonstration of the process policy of this consequence deserves. The people are not real; the process, the argument, the votes and the records are.
Who deliberated. Twenty-eight fictional composite members — personas authored from the real sources in the published Source Register — organised into ten policy teams: institutional authority (core), strategic posture, security and resilience, workforce, data governance, electricity, water, hardware supply (accelerators), compute, and te Tiriti / Māori data sovereignty. Each team held three members, with two cross-memberships deliberately wired across the known fault lines: an energy-systems modeller sitting on both the electricity and compute teams, and a Māori data-governance practitioner sitting on both the te Tiriti and data teams. The te Tiriti strand argued only from published Māori-led sources [S-22]; no tikanga position was fabricated. No words were placed in the mouth of any real person: this document is written as a neutral proposed framework, no real person’s voice appears in it, and none appears in any deliberation transcript.
How positions formed. Every team opened with a framing grounded in the Source Register, ran a genuine multi-turn argument — each team contained at least one designed disagreement drawn from real constraint data, and several positions in this framework (the conditioned levy-credit in §5, the non-severable pair in §4, the two-stage procurement sequence in §8) exist only because those disagreements collided — and then moved through formal decision points: a framing lock, an instrument selection by ranked vote, and a readiness check by consent, with a priority-weighting exercise across the capability fund’s resource strands. Every factual claim in every turn carried its [S-n] citation; every acknowledged gap [G-n] was posted as an explicit turn and sealed into the record.
What the machinery changed. Two events matter most for judging this process, because in both the outcome moved:
A deliberative instrument that never changes an outcome is decoration. This one blocked a plan, forced a revision, sustained a recorded dissent, and had its own membership send a conclusion back for being too shallow. The policy in this document is measurably different from the policy its teams first drafted — and the difference is auditable.
How outcomes were ratified. Every team’s position went to the full membership with a plain-language briefing, and was ratified by an anonymous all-member vote under a quorum rule. No vote payload at any point exposed any voter’s identity. One position was returned and re-ratified at a second round, as described above.
How to verify. Every deliberation record — team threads, decision points, cross-check history including the block and revision, the return and re-ratification, and the final tallies — was cryptographically sealed at the time it was made, and the sealed records can be verified independently of the platform that produced them. The sealed record bundles are published as plain JSON, alongside a standalone verifier that trusts nothing on this site and re-checks every signature itself. The Source Register, with every [S-n] entry and its URL, and the GAP register (section 13) are published with the framework, and the terms are set out in a glossary. Any reader can confirm that the positions in this document are the positions the deliberation actually produced, unaltered — start at Check the record ».
This is the claim no other AI-policy document in New Zealand currently makes, and the one this framework stakes everything on: not “trust us, we consulted”, not “trust us, we deliberated”, but a sealed record and the means to check it without trusting anyone. Policy you can audit. That is the standard this framework asks to be held to — and the standard New Zealand could ask of anyone who claims a policy was built on evidence. Do not take this document’s word for it. Check the record.
Sources: see the published Source Register (S-1 through S-24) and GAP register (G-1 through G-8), published alongside this framework. This document is an AI-generated demonstration: a proposed framework produced by a simulated deliberation of fictional composite members, offered for public discussion. It is not the policy of any party, and not the position of any organisation.