Canada Did It. Why Don’t We?

The question is never can a middle power build a frontier-class model. DeepSeek and Mistral settled that. The question is the one nobody wants on the board: what is the bill, in dollars and in people, and what does it cost every year after the first model ships.

Here is the answer for an Australian build, before the detail: roughly US$320 million in the best case and US$1.8 billion in the worst, to a first credible model — then US$110 million to US$460 million a year, forever, to stay in the game. Everything below is how those numbers are built, what drives them, and the one line item — people — that decides whether the whole thing is possible or fantasy.

First, kill the cheap number. DeepSeek’s famous “US$5–6 million” was the marginal cost of a single final training run, not the program. The cluster it ran on was assembled over years and cost hundreds of millions. Anyone waving the five-million figure at a minister is either confused or selling something. The real order of magnitude is the one above.

“Frontier-like” here means frontier-minus-one: a model in the class of the best open-weight and prior-generation frontier systems — Mistral Large, Cohere’s Command line, the DeepSeek and Qwen frontier — that is sovereign, controllable, fine-tunable on classified data, and air-gappable. It does not mean beating GPT-5.5, Opus, or Gemini at the bleeding edge. That target is a different and far worse number — a multi-billion-dollar-a-year standing treadmill that OpenAI, Anthropic, and Google are now running, and no single middle power can sustain alone.

Frontier-minus is the correct sovereign target. You do not need to win the leaderboard. You need a capable model you own, can audit, and can tune on data you would never send to a US or Chinese API. That is the bar Canada cleared with Cohere and France with Mistral, and it is the bar priced here.

Two scenarios, two honest extremes:

• Best case — lean, disciplined, private-sector execution. DeepSeek/Mistral efficiency. Smart procurement, few failed runs, a tight elite team, synthetic-data-heavy pipeline.
• Worst case — realistic-pessimistic. A bigger model, more failed runs, premium data licensing, a larger team on richer packages, infrastructure overruns, slower iteration.

And the framing that matters most and is always omitted: this is costed as program-to-first-model plus annual run-rate. A sovereign model is not a project that finishes. It is a standing capability — a subscription you pay in talent and power for as long as you want to stay sovereign.

Every dollar in this build sits in one of four buckets. Their order surprises people: the GPUs are not the dominant cost. The people are.

1. Compute. As of early 2026, an H100 costs roughly US$25–33k (PCIe) to US$35–40k+ (SXM5) to buy, and a full 8-GPU HGX system exceeds US$350k — so the all-in cost of a GPU deployed in a training cluster, once you add fabric, storage, and the node around it, runs 30–50% above the bare card: call it US$40–55k per GPU. Rental has collapsed from its scarcity-era highs and now clusters at roughly US$2–4 per GPU-hour, with Nvidia recently nudging rates back up about 20%. Buying beats renting only above sustained, high-utilisation use over multiple years — which is exactly what a standing sovereign capability is, and exactly why sovereignty forces you to own the iron eventually.

2. People. The binding constraint, and the answer to the only question that actually decides this. The pool of researchers who have actually contributed to a frontier model training run is a few hundred people on the planet, and OpenAI, Anthropic, Google DeepMind, and xAI are in a bidding war for all of them. Current total comp at those labs: OpenAI engineer median around US$555k, L5 near US$1.15M, research scientists US$770k to US$1.47M+; Anthropic median US$440–600k, senior researchers clearing US$1M with secondary tenders. At the apex, Meta’s Superintelligence Labs has reportedly written packages worth US$100M+ in signing alone and one deal near US$1.5B over six years. You cannot outbid that, and you do not need to — but you do need to clear the L5-researcher band, in cash, to pull people to Sydney. More on why “in cash” is the expensive word in a moment.

3. Data. Cheap if you lean on open corpora, Common Crawl, and synthetic generation; expensive if you license premium news, books, and code and stand up a serious human-annotation operation for the RL stage. Best case low tens of millions; worst case north of a hundred million. The modern frontier increasingly comes from reinforcement learning on verifiable rewards — code, maths, reasoning — where you generate the data rather than scrape it, which is genuinely deflationary for exactly the capabilities a sovereign cares about.

4. Overhead. Power, colocation, networking, security, facilities, legal, administration. Australia’s industrial power is not cheap, but it has the renewable headroom and the colocation footprint (NEXTDC and peers) to host this without building a datacentre from scratch.

Lean and disciplined. A small elite core, an efficient 2,000-GPU cluster, a synthetic-heavy data pipeline, and the kind of execution that gets a credible model out in about two years.

Two-year program to first credible model: ~US$320 million. Sustaining run-rate thereafter: ~US$110M / year (opex plus hardware-refresh amortisation).

A bigger model, an 8,000-GPU cluster on newer silicon, premium data, a much larger team on richer packages, failed runs, and three years instead of two to a credible result. This is not the disaster scenario — it is the realistic one for an institution learning as it goes.

Three-year program to first credible model: ~US$1.8 billion. Sustaining run-rate thereafter: ~US$460M / year.

A government will hear “US$320 million” and file it as a capital project — build it, cut the ribbon, move on. That is the category error that kills sovereign AI before it starts.

The frontier is a depreciating asset measured in months. The model you ship is the previous frontier within two or three quarters. Staying useful means a continuous retraining cycle — which means the US$110M–US$460M annual run-rate is the real commitment, not the upfront number. Sovereignty in AI is a subscription, paid in talent and power, for as long as you want to remain sovereign. A minister who funds the build and not the run-rate has bought a cluster that ages into a paperweight.

This is the crux, and it is more expensive than it looks, for a reason that is not obvious until you read the comp structure.

OpenAI’s ~US$1.15M L5 package is roughly US$336k base plus ~US$774k in equity. That equity is paper — Profit Participation Units funded by investors betting on a US$750B-plus valuation. It costs OpenAI almost nothing in cash today. A sovereign Australian entity that “matches OpenAI wages” in cash is therefore spending more than three times the cash per head that OpenAI itself spends. Naively matching the visible packages means a 3–4× cash multiplier on the base salaries. That is how a people line quietly becomes the entire budget.

There are only two ways out, and a serious build uses both:

Pay in equity, not just cash. Structure the lab as an equity-bearing national champion — the Cohere and Mistral model — so you can pay people in upside the way the labs do, not only in salary. Cohere raised global capital; Mistral raised over a billion euros. A pure government cost-centre cannot offer that, and will haemorrhage cash trying to compete on salary alone. The institutional form is a strategic decision, not an afterthought: national-champion-with-equity, not government-department-with-payroll.

Run the repatriation play. You cannot outbid Meta for the apex, and you do not need to — you target the strong tier of that few-hundred-person pool. Australia’s unique lever is that a large share of the world’s frontier AI talent is Australian expats in the Bay Area who left because there was nothing to come home to. Offer them a sovereign mission, a clearance-grade role, a country worth building for, and a life that is not a one-bedroom in Mountain View, and you can pull them home for packages that match — not multiply — the US tier. That is the one move France and Canada cannot copy, and it is the difference between the best-case people number and the worst-case one.

Rent to start; buy to be sovereign. Renting the compute for one program cycle runs roughly US$40M (best) to US$210M (worst) — a clean way to de-risk the first model before committing capital, prove the team can run a stable training pipeline, and avoid a nine-figure cluster that ages on a loading dock. But renting your reasoning capability from someone else’s datacentre is the opposite of sovereignty. The sustained, high-utilisation profile of a standing capability is precisely the case where owning beats renting on cost and the only case that satisfies the sovereignty requirement. Start on rented iron; migrate to owned iron in your own colocation once the team and the pipeline are proven.

Put the build against what nations already spend without blinking.

• US$320M (best case) is less than one B-21 bomber (~US$700M a unit).
• US$1.8B (worst case) is about half a single Virginia-class submarine (~US$3.5B), and less than a seventh of one aircraft carrier (~US$13B).
• Australia spends roughly AU$33 billion a year on defence and is committing AU$268–368 billion to AUKUS submarines over their lifetime.

A sovereign frontier model — the entire program, best case — costs less than one bomber. The worst case costs less than half a submarine. The annual run-rate that keeps it alive is a rounding error against the defence budget. “We can’t afford it” is not an economic statement. It is a statement about what an institution has decided to classify as strategic — and right now AI capability is filed under IT procurement, not national security, which is the actual failure.

The cost analysis answers how much. It does not answer why bother — and the why is where the real value sits, because the model itself is the smallest thing you get.

A frontier model is not a product. It is a foundry. Once the weights and the pipeline are in-country, every vertical stops being a new build and becomes a fine-tune — a casting from the same furnace, at a fraction of the cost of the base. And the verticals that matter to Australia are precisely the ones the global frontier will never touch, because they are not its market and not its security clearance.

Take agriculture — an economy this country runs on. No foreign lab is ever going to fine-tune a model on Australian soil chemistry, rainfall patterns, paddock-level yield history, and pest cycles, because that data is sovereign, the market is “merely” national, and the problem is too specific to bother with at San Francisco altitude. For a country that owns the base model, it is one fine-tune. The same holds for mining and resource exploration, for defence systems trained air-gapped on classified data, for a health model bound by privacy law to never leave the country, for legal and government models that can only ever run on sovereign infrastructure. One roughly billion-dollar base capability spawns dozens of sovereign vertical models at marginal cost. That is the leverage the cost table cannot show: you are not buying a model, you are buying a machine that makes models, pointed at the problems no one else will solve for you.

This is where sovereignty stops being defensive and becomes generative. The argument up to here has been protective — don’t let a foreign power read your queries. Owning the capability flips it: you can now build things that would otherwise simply not exist, because the global labs optimise for the big horizontal markets and ignore the local, the sensitive, and the small. The sovereign lab goes deep exactly where the frontier goes wide.

And the deepest asset is neither the model nor the verticals. It is the people. The moment a few hundred Australians have actually run a frontier training pipeline, that knowledge is in-country — and unlike a model, knowledge compounds. They train the next cohort. They spin out companies. They advise government from competence instead of from a brochure. They stop leaving, because for the first time there is somewhere here worth building. The brain drain reverses, and the human-capital base grows on its own. Canada built its research base in the 1980s and is still cashing the dividend through Cohere and an entire ecosystem four decades on. The model depreciates in months; the capability to build models appreciates for decades.

Which is the whole point, said plainly: there is an enormous difference between a nation that uses frontier AI and a nation that builds it. A user watches the tutorial and stays dependent on whoever made it. A builder owns the workshop — and becomes the one others learn from. This is not about getting better at watching videos on how to hold the screwdriver. It is about building things with it, and then building the screwdrivers. The country that can do the second is sovereign in a way no licence agreement will ever grant the country that can only do the first.

A sovereign Australian frontier-minus model is a US$320 million to US$1.8 billion program to a first credible model, with a US$110 million to US$460 million annual tail to stay relevant. The dollars are defence-rounding-error. The compute is buyable. The data moat is real but porous and shrinking for the capabilities that matter.

The binding constraint is none of those. It is the few hundred people who can actually run a frontier training pipeline — and the institutional will to structure a national champion that can pay them in equity and pull the expats home. Get the form right and the funding right, and the model is not a fantasy. The fantasy is believing it gets built by a committee that has never run a bare-metal cluster.

Canada answered the talent question in 1985 by funding Hinton. France answered it by treating Mistral as an instrument of state. Australia has not answered it at all. The bill, it turns out, was never the hard part.