The First Software That Can Sell Out

Software’s economic identity rested on one miracle: the millionth copy costs what the first one did, which is to say nothing. Every scaling playbook of the last thirty years is a footnote to that fact. Intelligence is the first software in history that can sell out.

Two years ago that sentence would have read as a category error, like saying a poem ran out of stock. And 2026 is the year the receipts arrived: for the first time since the AI era began, the machinery of intelligence got more expensive, breaking a decade-long trend of relentlessly falling costs.

Seventy-two hours to empty shelves

Consider what happened last week. On July 16, Moonshot AI released Kimi K3, by most accounts the largest open-weight model ever published, priced at roughly half of what comparable closed models charge per task and, by the company’s own evaluation, ahead of most of the field on the coding work that founders actually buy models for.

Three days later, Moonshot stopped accepting new paid subscribers. Not a price hike, not a throttle: a closed door. Demand had pressed against the ceiling of the company’s GPU capacity. A software product, seventy-two hours after launch, hanging a sign in the window that no software product was ever supposed to hang: sold out.

It is tempting to file this as one startup’s growing pains. It is more honest to read it as a data point on a curve, and the curve is the story.

The year the curve bent

For roughly a decade, the price of a unit of intelligence did only one thing: it fell, and it fell fast. Every year the same benchmark score cost a fraction of what it had cost the year before. Falling cost was not a fact about the industry; it was the industry’s weather, and we all dressed for it. In 2026, the curve bent the other way for the first time. Memory, the binding constraint of AI hardware, entered a supercycle: industry analysts project DRAM prices up more than seventy percent this year, with demand for high-bandwidth memory growing roughly twice as fast as supply.

Here is the part that hides the story from casual view. Token sticker-prices kept sliding, so the inflation never appeared where we were trained to look for it. It surfaced as rationing instead: weekly usage caps at the top labs (anyone who has bumped against a frontier model’s limits on a Thursday knows the feeling), waitlists for capacity, and now a frontier lab turning away paying customers. Rationing is a price increase wearing a disguise. When a merchant cannot raise the price, the queue raises it for him, and we pay in waiting what we no longer pay in dollars. Nobody raised our price. They lowered our availability, which is the same thing said politely.

The recipe and the oven

The strangest twist is that K3 is open-weight. Anyone on earth may download the model. Moonshot gave away the recipe and ran out of bread. Any Montrealer who has stood in line outside a bagel shop at one in the morning understands the situation perfectly: the recipe was never the scarce part. The oven was. For thirty years the source was the secret and the running was trivial; now the source is public and the running is the moat. You can have the model; you cannot have it run.

What the old books knew

There is a certain comedy in watching an industry rediscover the oldest chapter of the economics textbook. Marginal cost, capacity planning, queues, thrift: we spent thirty years proudly exempt from all of it, and the exemption has now expired. The most immaterial artifact in history has rediscovered matter, and matter, patient as ever, has sent its invoice.

The industry’s most interesting response so far is not bigger spending but better housekeeping. DeepSeek recently published serving techniques, speculative decoding with confidence-scheduled verification, reportedly delivering about fifty percent more throughput from scheduling alone. No new chips, no new model: efficiency, that dowdy old virtue we retired when compute was free, walks back in wearing the lab coat. In a world of abundant compute, efficiency was a rounding error. In a world of rationed compute, efficiency is a moat.

Scarcity at the summit, tailwind in the valley

Now for the turn that matters most to those of us who fund and serve Canadian entrepreneurs, because the news here is genuinely good. The scarcity is at the summit, not in the valley. Roughly one Canadian business in eight has adopted AI at all; the national ambition is six in ten by 2034. The work of closing that gap, the unglamorous work of putting working intelligence into a trucking dispatcher in Laval or a claims desk in Halifax, is orders of magnitude less compute-hungry than frontier training. The labs are fighting over ovens the size of power stations; most of our clients need a toaster that works every morning. Scarcity at the summit is a tailwind for the practical middle, and Canada’s national strategy already treats compute as infrastructure, which is precisely the right instinct: you do not leave the roads to chance and hope the trucks sort themselves out.

So when we sit across from a founder, we might start asking about compute the way we ask about payroll: not as a technical curiosity but as a cost of goods, with a plan behind it. We can prize the product that runs beautifully on a small, cheap, fine-tuned model over the demo that depends on a frontier model staying both available and underpriced, because one of those assumptions just failed in public. And we can read a vendor’s capacity story as carefully as their capability story, remembering that a dependency that can sell out is a supply chain, and supply chains deserve second sources.

None of this is a reason for gloom. It is a reason for craft. The zero-marginal-cost miracle gave us thirty years of businesses that scaled by wish; the new scarcity will give us businesses that scale by skill, and skill has always been the more durable endowment. The door that closed last week was a frontier lab’s; the door that stands open is ours, and the great unserved middle of this country is on the other side of it. We should walk through while the giants are still waiting in line.