Programming used to start with a pencil. You wrote the program on coding sheets, eighty columns wide, one character per box, and you checked it on paper too: played the machine at your desk, found the bugs before the computer ever saw them. The university mainframe did not care about your ambitions; you booked time on it, you got a quota, and your job went into a queue behind everyone else's. A wasteful program was not a style problem. It was your entire allocation gone by Tuesday, and an operator who knew your name for the wrong reason. An enormous amount of foundational computer science got done by people who could not afford to run their code twice.
Nobody designed that discipline. The meter did. It rationed carelessness and forced you to understand a thing before you executed it, and the habit followed the meter around: Steve Wozniak wrote Integer BASIC for the Apple I on paper, hand-assembling the 6502 machine code himself, because he could not afford an assembler to go with the twenty-dollar chip. Then compute got cheap, then effectively free, and the habit went with it.
Scarcity also made people invent. Time-sharing exists because whole departments were fighting over one machine and somebody had to make the sharing fair; virtualization exists because one expensive computer needed to pretend to be many. The machine room's rulebook (quotas, queues, tenancy, metering) is, one renaming later, the cloud.
I have written before about what happens when a cost filter disappears: in Judgement is the job now, the filter was effort, and removing it meant bad ideas stopped dying of natural causes. Compute was doing us the same favour, one layer down the stack.
The meter is back
Here is the part I did not expect: the meter came back. Tokens cost real money per unit of thought, and anyone running serious agent workloads has rediscovered mainframe-era habits without ever having touched a mainframe. I route dumb plumbing jobs to the cheapest model that can do them and save the expensive one for work that deserves it. I set budget caps. I cache aggressively. I look at a plan an agent proposes and ask whether the question is worth what the answer will cost, which is a sentence a 1970s operations manager would recognize instantly.
The habits are making the work better, not just cheaper. Deciding which model deserves the job forces you to say what the job actually is, and a budget cap is a crude summary of intent: it declares how much this question is worth before you ask it. The engineers who handle the meter well are the ones who were already good at the ask. People who plan rather than "vibe" produce better things; I've seen it time and time again.
Scarcity by itself is not the feature, though. A well-designed meter is. Allocation rules have to be published and legible, small experiments have to stay cheap enough that nobody writes a memo before trying something, and a capability that proves itself has to earn more compute without knowing somebody. Get that wrong and scarcity does not produce invention. It produces queues, patronage, and incumbents who learned how to work the paperwork, which is a failure mode this country has demonstrated before.
The pitch everyone is making
So it is strange to watch my own country treat a metered slice of compute as an embarrassment. The version of Canadian sovereign AI being pitched in every boardroom and briefing note right now is always the same shape: build sovereign datacenters, stand up a Canadian model hub, keep the data inside the border, and let patriotism handle adoption. Somewhere in the first five minutes comes the apology: of course, we cannot match American capex. It is an abundance play, made by a country that cannot afford abundance, delivered with an asterisk admitting it.
The arithmetic behind the apology is real. As I write this, in July 2026, Nvidia's market cap is just under US$5 trillion and the entire Toronto Stock Exchange (roughly 2,600 companies: every bank, every railway, every miner, Shopify and all) adds up to about US$4.9 trillion. One American chip company is worth our whole stock market. Anyone proposing that we can build the same way those tech giants do is not being ambitious; they are being silly.
I have been chewing on this since a long argument about what a "Canadian AI Commons" would actually be, and two convictions stuck: the model hub is the wrong organizing unit entirely, and the scarcity everyone apologizes for is the most interesting thing about the way forward.
Canada cannot win the spending race, so it should stop entering it. Innovation is the race where being small is not a handicap.
The capability is the asset
A model hub asks "what model do you want?", which is a question almost nobody in a government department or a mid-size Canadian company is actually asking. What they want is to reuse a job that is already known to be safe, and that job is a versioned, governed package: the model, the corpus it is grounded in, the prompts and orchestration, the evaluation suite with its results attached, the deployment template, and machine-readable metadata covering privacy, residency, retention, tenancy, and permitted use. Register that and a department can adopt it. Ship the model alone and you have given them a science project.
The strongest version of this is that the government runs the same capability it offers. When the CRA help desk and the small business owner are both leaning on the 2026 tax-year assistant, the answer you got at midnight is the answer their tooling gives the agent reviewing your file in the morning. A shared, versioned understanding of the rules between the state and its citizens is a new kind of public infrastructure, and no foreign vendor is going to build it for us.
Readers of my MCP servers report will recognize the shape: organize around the question someone is actually asking, and bundle the rest behind it.
The registry is also where the genuinely Canadian moat lives, and it is not patriotism. It is the boring, legalistic stuff nobody else will bother to test: bilingual parity as a hard gate, verified rather than asserted, and machine-readable answers to the questions every Canadian counsel asks about residency, Law 25, PIPEDA, and retention. Tenancy has to match how this country is actually structured, too, with universities, municipalities, federal departments, and Indigenous organizations governing their own data on their own terms. OCAP is the reference point there, and "appropriate" is theirs to define, not the platform's.
Privacy belongs in the foundation rather than on a settings page, and 2025 supplied the demonstration. Grok's share button quietly created public, indexable pages, and private conversations (medical questions, business plans, at least one password) turned up in Google search results. That was governance decided after launch. A capability registry runs in the opposite order: retention, sharing, permitted use, and refusals are declared, tested, and attached to the capability before anyone is asked to trust it. A Canadian's session with a public capability is not training data and not shareable by surprise.
Run it here or it didn't happen
Evaluation has to be something the platform does, not something the publisher says. An eval that counts is one that ran on platform compute, with traces attached, repeatable by the next skeptic. Follow that backwards and you arrive at the meter: the platform has to own compute, scarce and shared and queued, with quotas and review, which is the mainframe model with better cooling. The thing every pitch treats as the budget compromise is what makes the evaluations worth reading.
Scarcity keeps earning its keep from there. Review-before-allocation means a capability gets the shelving conversation (is this worth the metal?) before it exists, enforced by a scheduler instead of a lead who happens to be in the room. And because everything ran on the same metered infrastructure, efficiency becomes a published, comparable property of every capability. Abundance tends to leave you with a registry full of untested uploads and a datacenter full of idle fine-tuning jobs.
An API key for every Canadian
If this is national infrastructure, it should be available and safe for every Canadian, not just for departments with a procurement officer. Concretely, I mean something almost boring: any Canadian can request an API key, the endpoint speaks the standard OpenAI-compatible interface that every tool, library, and agent already understands, and the key comes with a modest monthly allotment. A small business writing its first grant application, a kid in Moose Jaw building something weird at midnight.
The allotment is a floor rather than a ceiling. Pay commercial rates when your ambitions outgrow the slice, and nothing stands in your way. Universal access to unmetered compute is a fiscal fantasy; universal access to a metered slice is a line item. Nobody is promised infinity. Nobody starts at zero.
The "safe" half comes from everything above: the citizen key serves the capability layer rather than raw model weights with a maple leaf on them, so what a Canadian gets by default has already passed the bench. Sovereign infrastructure that requires bespoke integration is dead on arrival; speak the dialect the entire ecosystem already speaks, and the cost of choosing the Canadian option drops to changing a base URL. That is the whole adoption strategy, and I feel like it's a good one.
Publishing at the flip of a switch
The runtime side is not exotic either; the managed-agent platforms have already sketched the shape. Amazon calls it AgentCore: a serverless runtime, a memory store, a gateway that turns existing APIs into tools an agent can call, primitives you configure rather than servers you run. Put the same primitives inside the border and publishing becomes a switch you flip. You publish the capability, the platform runs its evals, you attach the governance metadata, and a small isolated instance comes up with the corpus mounted, the tools gated, and the meter attached. Consumers pay a small fee per execution, publishers get their cut, and usage pays for the metal.
Then the long tail shows up, because not every capability needs a company behind it. A university lab publishes its document-extraction pipeline as a free, open capability. A municipal clerk publishes the one that answers zoning questions. Somebody with a weekend, a hobby corpus, and a QLoRA fine-tune publishes because they had an idea, the same way people publish open source. Paid and free live in the same registry and pass the same bench; the price tag is metadata.
The raw material should be public too. If a corpus was assembled with public money (the legislation, the StatCan tables, the forms and their guidance), publish it machine-readable, under an open licence, documented well enough to train against. Then anyone can fine-tune on Crown data and put the result on the registry, charge for it or give it away, their call. The giants trained on the open web without asking permission; we could do the deliberate version.
Where assets go to die
Governance-first platforms have a distinguished history of becoming graveyards. If publishing to the registry is all cost (forms, evals, review queues) and consuming from it carries no advantage, then nothing good gets published, the metadata gets filled in cynically, and in three years an auditor general writes a report about it. WASTED.
The way out is to make the registry how vendors get paid. The supply side is not hypothetical: Canada already has model shops like Cohere, startups wrapping a model around one legal or payroll workflow, and university labs (Mila, Vector, Amii) full of work that never survives the hand-off out of academia. What they lack is a buyer worth the paperwork. The federal government buys more software than anybody else in Canada, and anyone who has sold into it can tell you about the grind; eighteen months from first meeting to signature is considered a win. Registry approval is what shortens it, because the evaluations and the trust metadata arrive already done and already trusted by the buyer. Approval unlocks the purchase, the purchase is the demand, and the demand is why a vendor fills in the governance fields and runs the evals. Nobody has to ask nicely.
A library card for the model era
Everything above sounds like infrastructure planning, but what it adds up to is a benefit, in the mundane, glorious way a library card is a benefit. It is for the businesses that need models they can point at customer data without a lawyer wincing, for the builders with an idea and no hyperscaler budget behind them, and for the merely curious, who want to see what the fuss is about without shipping their questions to a server in Virginia.
If that sounds like settling for less, remember January 2025. DeepSeek shipped a model close enough to the frontier that Nvidia lost almost six hundred billion dollars of market value in a single day, then gave the weights away. The cost and hardware numbers are still argued over and I would not lean on them. What holds is that architectural and training efficiency upset an assumption everyone else had priced in capital, and my reading is that the constraint had something to do with it: working under export controls, they could not count on the same scale of unrestricted access as the largest American labs, so they put more weight on efficiency. Canada does not need to beat OpenAI or Anthropic at their own game. It needs its own version of that, pointed at Canadian problems, with the results belonging to the citizens who paid for the compute.
AI is threading itself into every Canadian's life whether they realize it or not: the bank's fraud model, the kid's homework help, the chatbot the airline hides behind. Something that woven into daily life should not be gated by a corporate board in the States deciding what we get, what it costs, and what happens to everything we type into it. It should be a social benefit, and Canada already knows how to deliver one of those: every citizen gets a card, everyone pays for it together, and the limits are what keep it working for the next person. A metered slice of Canadian AI capability, grounded in our data and working in both our languages, is the same kind of promise.
Small and metered is not the compromise position.
Small and metered is the thesis.
