On June 4, Canada released its first national AI strategy, “AI for All.” It sits on top of the $2 billion Sovereign AI Compute Strategy from Budget 2024, and it puts health and life sciences at the front of the queue. There is a $200 million health mission, a $100 million expansion of VITAL, the pan-Canadian platform that connects hospital data across provinces while keeping it in jurisdiction, another $100 million for a Health Sector Data Space with CIHI, and a $300 million fund meant to put sovereign compute within reach of life sciences companies.
This is a serious and overdue commitment, and I am glad to see it. I also want to make a point that matters for anyone deciding where to put money or effort over the next few years. Compute and data are inputs. The outcomes still have to be built on top of them, and that space, between the capacity and the result, is where the value is actually going to be made.
A cluster of chips in Quebec and a federated data platform spanning provinces are necessary things to build. But neither one changes the advice an oncologist gives the patient in front of them. Megawatts and datasets do not become better care on their own. And quite frankly, I already have all the compute and data access I need. For a startup or a small business, that was never the binding constraint. Something has to sit on top of them and turn evidence into a decision a clinician can act on, and that is a different kind of work from pouring concrete and laying fibre.
The strategy diagnoses its own gap
The document is candid about this, more candid than I expected. In the prosperity pillar there is a sentence I would underline for anyone reading it as an investor. The path from breakthrough to deployment stalls, it says, “not because the technology is not ready, but because no single force connects innovators, industries, and institutions around the problems that matter most.”
That is the government describing the missing layer in its own buildout. The compute is being funded. The data platforms are being funded. The connections to the problems that actually matter, the decisions in the room, are not being funded at the same scale, because they are harder to fund. They look less like infrastructure and more like deliberate, specific, clinical work.
There is no single force, and that is the point
I want to pick at that phrase, because it gives something away. “No single force” assumes there is supposed to be one, a champion to identify and then stand behind. That instinct runs deep here. Canadian governments in particular tend to want to name the winning platform, the national flagship, the one connector, and concentrate the support behind it.
The problem is that nobody can predict what the winning force will be. If we could pick it reliably in advance, we would not need a market to find it. Breakthroughs in healthcare rarely come from the actor everyone expected. They come from someone who saw a specific problem clearly and was allowed to chase it. The honest answer to “which single force will connect the breakthrough to the problem” is that we do not know, and the moment we pretend to, we start quietly excluding the entrants who might actually have been the answer.
So the goal should not be to find the single force. It should be to get as many capable actors into the game as possible, and then to clear the road for them as new roadblocks surface along the way. Open the data instead of gating it behind one anointed partner. Make hospital procurement navigable for a small company and not only for an incumbent. Resolve the regulatory ambiguity that makes potential founders hesitate. Close the gap between a promising pilot and an actual purchase order, and give patients a way to reach those pilots through channels the traditional system does not yet offer. Every roadblock left standing is a filter, and it screens out precisely the unpredictable entrant the strategy says it wants. The connective tissue the document is searching for is not something you build once and appoint. It emerges when enough people are allowed to try, and when failing is cheap enough that many will.
Where value accrues in a buildout like this
We have seen this pattern before. Railways, the electrical grid, the internet. The infrastructure is capital-intensive, strategically essential, and, over time, largely commoditized. The firms that earned durable returns were rarely the ones that laid the rails or ran the power stations. They were the ones that built what customers actually used on top. Cloud compute became a low-margin utility. The software written on top of it has arguably created the most prosperity in the history of the world.
AI infrastructure will follow the same logic. A data centre above 100 megawatts is a remarkable asset and a thin-margin one, and many builders may not even need one. The latest open-source models offer a glimpse of a very different future from the one this strategy anticipates, where capability is cheaper and far more distributed. The lasting value will accrue to the layer that converts great ideas and data into a specific outcome someone will pay for. In healthcare, that outcome is a better decision at the point of care, or greater responsiveness to patient need, or something I have not yet imagined. The compute is the cost. The decision is the product.
None of this is an argument against the spending. The foundation has to exist, and a government is the right entity to underwrite it, to a point. My point is about access, simplicity, sequence, and where the returns sit. The public money de-risks a certain kind of infrastructure and, in the same motion, commoditizes it. What stays scarce is the layer above.
The hard part is knowing what to ask of the data
Turning scattered evidence into something computable is real work, and it is necessary, but it is fast becoming table stakes. The part that does not commoditize, the part no model can do for you, is knowing which clinical problem is worth pointing all of that machinery at in the first place.
The personalization itself sits beyond the model, for a reason worth spelling out. What actually tailors a recommendation to one patient is the combination of a dozen loosely related facts about her: her genomic signature, her anatomic stage, her age and menopausal status, the comorbidity that quietly narrows her options, how much of the planned regimen she can realistically tolerate, her values, her risk tolerance, and a host of one-off factors too idiosyncratic to list here. The estimate she needs lives in the interaction of those variables, and that interaction is not sitting anywhere in the literature as a passage to be found. It has to be imagined, then computed.
This is why a larger context window does not rescue it. A model that can hold a million tokens, or a hundred million, is still searching for patterns that are present in the text in front of it. The pattern that matters here was never in the text. It is a high-dimensional relationship across structured variables, the kind of thing no amount of reading will surface, and it becomes an answer only when someone computes it.
Knowing which of those variables matter, and how they combine into a number a clinician can defend at the bedside, takes people who have done the work in clinic. That judgment belongs to the clinician, the founder, the person building the product, and it is the scarcest input in the entire stack.
This is the work I find genuinely exciting, and it is where I think the next decade of healthtech value will be made. Choosing the decisions worth solving, framing them precisely enough to compute, and proving the answer drives real clinical value is product work in the truest sense, and it is irreducibly human.
It is the kind of work we do at Kesis & Sisters. You start from the decisions that actually trip clinicians up, work backward to the evidence and the math that resolve them, and keep that evidence current instead of letting it freeze the day someone last had time to read. It is detailed, unhurried work, and it is the part the infrastructure money does not buy. Building the data platforms is one job. Knowing what to ask of them is another, and it is the one that decides whether any of this reaches a patient.
What this means if you are deciding where to invest
Canada now ranks fifth in the world for AI venture capital, with $3.1 billion invested in 2025. A great deal of attention in that flow is on infrastructure, which is understandable, because it is concrete and the government is standing behind it. The differentiated, defensible value is one layer up, in the companies that turn public capacity into outcomes a customer will pay for.
In healthcare specifically, the moat is not really compute or data access, both of which are becoming public goods, by design, with public money. It sits above the engineering, in the clinical judgment to know which problems are worth solving, the deterministic tooling that can be trusted with a number once you have chosen the right question, and the confidence of the clinicians who have to use the result. The data engineering can be hired. The taste for which decision matters, and the credibility to make clinicians believe the answer, cannot. That is what is slow to build and hard to copy, and it is precisely what makes it worth backing.
Canada is funding the foundation, and it is the right thing to fund. The returns, and the better care, will belong to the people who know what to point it at, and to a country willing to let many of them try rather than betting on one. If you are building in that layer, or weighing where it fits in this buildout, I’d like to compare notes. Discuss a collaboration.