The advice you hear at every biotech panel and on every fintech podcast is some version of this: in AI drug discovery, back the company with the biggest, smartest model. Whoever has the most parameters, the most compute, the most impressive demo on a protein-folding benchmark — that's where the money goes. It sounds right. It is, for a while, even profitable. And it is the single assumption I'd most want to stress-test before you wire money into this corner of the market.
Here's my verdict in one sentence, so you can stop reading if you disagree: in AI-driven drug discovery, the models are converging toward commodity while the structured biology feeding them stays scarce, and a small Canadian company called MindWalk is a high-risk bet on that second thing being where durable value settles.
I'll disclose up front that I hold a small position and that nothing here is investment advice. I'm walking you through how I think, not what to do.
Where the common advice is roughly right
Models do matter. A frontier prediction system that can rank candidate molecules faster than a wet lab can saves real time and real capital, and time-to-candidate is the metric that moves a clinical-stage balance sheet. If you'd ignored the model layer entirely five years ago, you'd have missed the entire repricing of the sector. The companies that built credible computational platforms — Recursion, Schrödinger, Isomorphic-style efforts inside larger firms — earned their multiples partly on architecture.
So the advice isn't wrong. It's incomplete in a way that costs money.
Where it breaks down
A drug-discovery model is a very expensive way to be confidently incorrect when the biology underneath it is messy. Public biological data is fragmented across decades of inconsistent assays, mislabeled cell lines, and experiments that were never designed to be machine-readable. Feed that into the most capable model on the planet and you don't get insight — you get a beautifully formatted wrong turn that a team might chase for eighteen months and several million dollars before the lab disproves it.
That is the part the panel-circuit advice skips. The scarce input in this field is no longer raw model capability, which is converging fast and getting cheaper every quarter. The scarce input is biology that has been cleaned, structured, linked, and made interrogable — so that whatever model you point at it is reasoning over something true. Models commoditize. The curated biological substrate compounds, because every additional validated relationship makes the next query more reliable, and that advantage is hard to copy.
The MindWalk bet, with numbers
MindWalk's pitch is that it sits at this substrate layer rather than competing to have the flashiest model. The company describes a system organized around hundreds of millions of biological relationships — on the order of 660 million patterns, by its own account — designed to be the reasoning foundation that drug-discovery teams query, rather than another standalone prediction engine.
On the financials, the company is small and early. As of its most recent reporting it cited roughly $4.2M CAD in revenue, growing about 52% year over year, and it has signed what it frames as its first platform contract — the early evidence that a customer will pay for the substrate rather than the model. That's meaningful as a proof point and almost meaningless as a moat. It is one contract. The company runs operating losses, carries small-cap liquidity risk, and depends on financing it does not yet generate internally. None of that is hidden, and you shouldn't treat it as if it were.
Mapping the stack, not crowning a winner
The useful exercise isn't deciding whether MindWalk beats Schrödinger. They aren't doing the same job.
| Player | Roughly where it sits | Primary risk |
|---|---|---|
| Frontier model labs / Isomorphic-style efforts | Prediction layer | Capability commoditizes |
| Recursion | Owns wet-lab data generation + platform | Capital intensity, integration |
| Schrödinger | Physics-based simulation tooling | Tied to compute and licensing cycles |
| MindWalk | Structured-biology substrate / reasoning layer | Tiny, early, unproven at scale |
Read that as a map. Different companies are capturing different parts of the value chain, and a serious portfolio thesis is about which layer you believe compounds — not which logo you like. MindWalk's claim is that the substrate layer is the one that gets more valuable with use. The companies above it are larger, better-capitalized, and far less likely to disappear.
How I'd actually decide
What it's best at: exposure to the thesis that data infrastructure, not models, is the durable asset. If you already believe that, this is a fairly pure expression of it.
Price and stage: micro-cap, pre-scale, financing-dependent. Position it like venture inside a public wrapper — small, sized to be wrong.
Who it's wrong for: anyone who needs liquidity, anyone who wants validated multi-customer revenue before entering, anyone who can't stomach a total loss on the line. If a 70% drawdown would change your life, this isn't your instrument.
Who this is for, who it isn't
This is for investors who want a thesis-driven satellite position and have the patience to hold through several quarters with no confirmation. It is not for a core allocation, and it is not for someone who's been told biotech AI is a safe way to ride the AI trade. It isn't.
Zooming out
Whether or not this particular company executes, the field appears to be moving past its first phase, where capability alone commanded the premium. The capital is starting to ask a harder question — where does the advantage persist after everyone has access to comparable models? — and the honest answer is increasingly "in the biology you've made trustworthy." That's a durable lesson even if MindWalk is the wrong horse.
So here's the more honest version of the rule: don't back the biggest model in AI drug discovery — back whoever owns the cleanest, most defensible biology underneath it, and size that bet for the real chance you're early and wrong.
