What Would It Actually Cost to Build a Frontier AI Model?
As governments around the world weigh “sovereign AI” strategies, New Zealand and Australia must each be clear-eyed about what frontier capability actually requires.
Tom Maasland is a Partner at MinterEllisonRuddWatts and one of New Zealand’s most informed observers at the intersection of technology, law, and public policy. Ahead of the AI Forum NZ’s 30 June panel, where he will serve as moderator, AI Forum NZ Executive Director Madeline Newman sat down with him to explore what it would genuinely cost New Zealand and Australia to build a frontier AI model, and what a more realistic sovereign strategy might look like.
MN Madeline Newman — Executive Director, AI Forum New Zealand · Interviewer
TM Tom Maasland — Partner, MinterEllisonRuddWatts · Moderator, 30 June Panel
ℹ Tom Maasland used Claude (Anthropic) to help research and structure some of the figures cited in this piece. Many figures are sourced in the footnotes – figures that are not cited are estimations on how much frontier AI would cost in NZ given the global costing from the cited figures; all views are his own.
| NZ$3.5–8.5b Frontier build, if NZ & AU combined | NZ$150–400m Annual access alternative | 80–90% Capability retained via access model |
MN
Let’s start with the headline question: what would it actually cost New Zealand and Australia to build their own frontier AI model?
The honest answer is that it depends entirely on what you mean by “frontier” – and the gap between the marketing version and the real thing is roughly two orders of magnitude.
And the gap is widening. Two things are happening at once. The frontier ceiling keeps rising: training a single GPT-4-class model cost more than US$100 million by 2023,[1] and a single GPT-5-generation run is rumoured to have cost more than 500 million dollars[2]. Credible industry voices, including Anthropic’s CEO, suggest frontier runs could reach US$10 billion by 2028.[3] For a country deciding whether to build or to access-and-adapt, that divergence is the whole argument – chasing the moving frontier gets exponentially harder, while standing one generation back gets cheaper every year.
To put numbers on the frontier itself: by estimates from research groups such as Epoch AI, amortised training costs are rising by nearly 3.5x per year.[4] Named-model anchors make it concrete – GPT-4 around US$78 million, Google’s Gemini Ultra around US$191 million, Meta’s Llama 3.1 405B around US$170 million.[5] A current frontier-class effort therefore sits in the US$200–500 million range for the GPT-5 and Gemini Ultra class and is projected to be 1 to 3 billion dollars for the late 2027 frontier.[6]
The training run is actually the cheap part. What breaks the economics is the organisation around it.
— Tom Maasland
Hardware accounts for roughly 47–67% of total cost, with R&D staff making up a further 29–49%.[7] Those staff costs reflect a brutal market reality: the world’s most experienced frontier researchers are concentrated in the San Francisco Bay Area, or elsewhere in the US, and through 2025 the bidding war pushed headline packages into the hundreds of millions. Meta reportedly offered up to US$300 million over four years to individual researchers.[8] That’s not a commentary on the quality of New Zealand’s AI talent, which is genuinely recognised internationally. It’s a question of critical mass: the pool of people who have actually built frontier-scale systems is estimated at only around 2,000 globally.[9] Attracting them, or retaining New Zealanders who could join that pool, would be one of the more consequential challenges any ANZ effort would face, and any credible programme would need to meet the global going rate. That’s not negotiable if you want people who have done this before.
Then there is the infrastructure. In 2026 the working unit of frontier compute is a cluster of around 100,000 advanced GPUs, costing roughly US$3–5 billion all-in and drawing 130–180 megawatts from the grid once cooling is included.[10] The scale leaders are larger still. xAI’s Colossus reached around 555,000 GPUs for an estimated US$18 billion, heading toward two gigawatts.[11] To put even the entry-level version in terms New Zealanders recognise: the Tiwai Point aluminium smelter, our single largest electricity user, draws about 572 MW, with the smelter consuming the equivalent of 20% of the North Island’s electricity use.[12] A single frontier cluster would need a meaningful slice of that. It’s not Tiwai Point, but it’s a noticeable new claim on a grid already being asked to do a great deal more.
To stay at the frontier, rather than touch it once, an ANZ entity would realistically need US$3–10 billion per year in sustained spending. That’s roughly 0.2–0.5% of combined trans-Tasman GDP, every year, compounding.
MN
If building a true frontier model is out of reach, what’s a realistic ANZ budget for a credible, one-generation frontier-adjacent model?
If the goal is a one-shot frontier-adjacent model, with top-tier open weights, a generation behind the true frontier, then the numbers look more like this:
Indicative ANZ Frontier-Adjacent Budget (USD)(calculated from above information)
| Compute infrastructure or multi-year cloud commitment | US$1–3b |
| Training runs plus experimentation | US$300–800m |
| Team of 100–300 researchers & engineers (2–3 years) | US$300m–1b |
| Data licensing, safety, evals, deployment | US$100–300m |
| Total indicative range | US$2–5b (NZ$3.5–8.5b) |
That’s broadly comparable to what France has effectively backed with Mistral, which secured a €2 billion investment valuing it at €12 billion (14 billion USD).[13] Mistral is a strong model family that is nonetheless not at the absolute frontier, which is exactly the point: this budget buys a credible, sovereign, one-generation-back capability, not a seat at the frontier. And critically, the estimate is for one generation. By the time it ships, relevance is not guaranteed.
Each cost category carries its own complexity, too: data licensing is increasingly contested, safety and evaluation work is still an emerging discipline, and deployment at scale brings operational costs that are easy to underestimate.
MN
What’s actually happening locally — how do Australia’s recent sovereign AI investments illustrate the challenge?
Sovereign Australia AI – a privately founded venture, in partnership with Sharon AI, rather than a government programme – is instructive precisely because it’s honest about its scope. It launched with 256 Nvidia Blackwell B200 GPUs hosted in NextDC data centres, plus a minimum A$10 million earmarked for licensing copyrighted training material.[14] Its models, Ginan and Australis, are small relative to the global frontier. Ginan is positioned as an open research model rather than a ChatGPT competitor.
To be clear on scale: 256 GPUs is roughly 1/100th to 1/400th of a frontier training cluster. It’s a legitimate sovereign capability play, but it makes no claim to compete with the global frontier models – the founders say so explicitly. It’s a small-scale training, fine-tuning and deployment capability, not a frontier-scale one.
Sovereign AI doesn’t have to mean building the engine. It can mean understanding the engine well enough to drive it safely.
— Tom Maasland
The distinction matters enormously for policy. If the goal is genuine strategic independence, you need frontier training. If the goal is resilience, localisation, and the ability to adapt and deploy responsibly, a well-resourced access and fine-tuning capability may deliver 80–90% of the practical value.
MN
Where does this leave New Zealand in terms of strategic choices?
The total picture for a credible access-based strategy is roughly NZ$600 million to NZ$1.5 billion upfront, which is mostly attributed to the compute facility. And then it’s NZ$150–400 million per year ongoing, covering frontier access, model refresh, operations, and talent. That’s 1–3% of the cost of the frontier-build option, for roughly 80–90% of the practical capability.
What you give up is control over the frontier itself, which is precisely the strategic question governments are wrestling with. The UK is instructive: it has committed around £2 billion to expand public compute twentyfold by 2030, with a £500 million Sovereign AI Unit layered on top to take stakes in domestic AI companies.[15] It has gone further still. A coalition including BAE Systems, BT, Lloyds, NatWest and PwC is now building “Lumen Sovereign,” Britain’s first fully sovereign frontier model, trained on home soil.[16] France has gone bigger again, committing €109 billion in 2025 including a one-gigawatt AI supercomputer.[17] Once you’re in, the costs compound.
For New Zealand, the more important question may be less about the price of the compute and more about the conditions under which we want to be a sovereign user of AI systems built elsewhere – who owns the models, who controls the infrastructure, and what obligations come with access. Those are fundamentally governance questions, not engineering ones.
And those questions are no longer hypothetical. In June 2026, a US export-control directive forced Anthropic to abruptly disable two of its most capable models (Fable (aka Mythos)) for every customer worldwide – not because of price, performance or uptime, but because the US government treated foreign access to a hosted model as a controlled “export.”[18] That episode is the sharpest illustration yet that an access-based strategy carries its own sovereignty risk: a frontier model can be withdrawn for reasons entirely outside our control, and even hosting a commercial model onshore does not insulate us if the provider’s home government intervenes. That is precisely why the conditions of access – contractual exit rights, multi-vendor resilience, the ability to fall back to open-weight models, and diplomatic arrangements with trusted partners, matter as much as access itself. The cheaper path is not the risk-free path; it simply relocates the risk from the balance sheet to the geopolitical domain.
The conditions of access are as important as access itself. That’s where New Zealand needs to invest its policy attention.
— Tom Maasland
That’s the conversation I’m looking forward to opening on 30 June. Sustainability in AI isn’t just about emissions – it’s about whether the systems we depend on remain trustworthy, accessible, and aligned with our values over time. Those are exactly the kinds of system-level questions the AI Forum’s Blueprint work is trying to surface.
AI FORUM NZ · 30 JUNE 2026 — Sustainable AI Series
Date: Monday, 30 June 2026 · Time: 5:00pm – 7:30pm · Location: Level 22, PwC Tower, Auckland CBD
Moderator: Tom Maasland, MinterEllisonRuddWatts · Panellists: Megan Tapsell (ANZ), Albert Bifet (Waikato), Richard Prowse (CDC), Isuru Fernando (Microsoft)
[1] “OpenAI’s CEO Says the Age of Giant AI Models Is Already Over” OpenAI’s CEO Says the Age of Giant AI Models Is Already Over | WIRED.
[2] “GPT-5: The Most Expensive AI Model Ever Built” GPT-5: The Most Expensive AI Model Ever Built – Blogs – Trixly AI Solutions.
[3]GPUnex, “How Much Does It Cost to Train an AI Model in 2026” — Dario Amodei: frontier models could cost $10bn to train by 2028; a GPT-4-equivalent has fallen from ~$79M (2023) toward ~$5–10M (2026). gpunex.com
[4]Epoch AI, current trends dashboard (accessed 2026): frontier training costs rising ~3.5×/year; the underlying 2.4×/year figure since 2016 is from Cottier et al. epoch.ai/trends; epoch.ai cost study
[5]“Machine Learning Model Training Cost Statistics 2026” (Feb 2026), citing Stanford AI Index & Epoch — GPT-4 ~$78M, Gemini Ultra ~$191M, Llama 3.1 405B ~$170M, Grok-2 ~$107M; DeepSeek V3 ~$5.6M. aboutchromebooks.com
[6]Deluair, “Frontier AI training cost trajectory 2026” — GPT-5/Gemini-class runs ~$200–500M; late-2027 frontier projected $1–3bn. deluair.com
[7]Epoch AI cost breakdown (hardware/interconnect 47–67%, R&D staff 29–49% incl. equity, energy 2–6%). epoch.ai
[8]The Batch / WSJ (Dec 2025): Meta offered packages up to ~$300M over four years; OpenAI retention bonuses to $1.5M. deeplearning.ai; Levels.fyi (May 2026) OpenAI L5 ~$1.15M total comp. pin.com
[9]Fortune (Jul 2025): est. ~2,000 people worldwide can push the frontier of advanced AI. fortune.com
[10]Deluair (2026): a 100,000-H100 cluster is the working unit of frontier compute, ~$3–5bn all-in, needing 70–100MW IT load (130–180MW grid draw). deluair.com
[11]Introl (Jan 2026): xAI Colossus reached 555,000 GPUs (~$18bn), expanding toward ~2GW — the world’s largest single-site AI training installation. introl.com
[12]Electricity Authority. ea.govt.nz
[13]Bloomberg Mistral Set for $14 Billion (€12 Billion) Valuation With New Funding Round – Bloomberg.
[14]itbrief.com.au / Business News Australia (2025): Sovereign Australia AI — with Sharon AI — ordered 256 Nvidia B200 GPUs at NextDC; A$10M for copyright licensing; Ginan open research model. itbrief.com.au
[15]UK Sovereign AI Unit (£500M, Apr 2026) sits atop £2bn committed to expand public compute 20× by 2030: European Business Magazine (Apr 2026) europeanbusinessmagazine.com; GOV.UK “One Year On” (Jan 2026) gov.uk
[16]GOV.UK (Jun 2026): Cosine coalition (BAE, BT, Lloyds, NatWest, PwC) to build “Lumen Sovereign,” the UK’s first fully sovereign frontier model, trained on Isambard-AI. gov.uk
[17]tech-insider.org (2026): France committed €109bn (Feb 2025), incl. €10bn for a 1GW AI supercomputer. tech-insider.org
[18]Anthropic statement (Jun 2026) on the US export-control directive suspending Fable 5 / Mythos 5 access for foreign nationals. anthropic.com; analysis: CIO (Jun 2026) cio.com
Stephan is the sports journalist for the Maple Grove Report.
