Recorded this chat July 2 with Ryan Henry, who runs Sand Hill North, the venture arm of a Montréal family office, and he showed up fresh off of YC’s demo day with something I didn’t expect. Check out our conversation where we focused our time on the latest and greatest batch from YCombinator as well as how to think of AI sovereignty, the topic du jour.
The batch he skipped. Sand Hill North invested in zero YC companies this batch, which, for them, is abnormal. The batch was built for one number, revenue, ARR as a headline (or is it “R” when it’s a new company?). Ryan’s read: a million in ARR is cheap to hit now. Spin up AI middleware that helps companies stand up agents, ride a problem everyone has this quarter, and the revenue arrives fast. Fast and spiky is not the same as durable, and paying today’s valuation for it means underwriting a trajectory nobody can see yet.
Strongest batch in history, asterisk. Every model called it the best batch ever. Carta’s Q1 numbers say the median seed valuation is roughly flat while the 95th and 99th percentiles go vertical. More money, fewer deals, all of it concentrating into the best of the best. Two companies landing at a $175M valuation isn’t a miracle, it’s the 99th percentile showing up right where you’d expect in a pool of two hundred. The funds that look like geniuses win on deal flow, not a better sausage machine. Ryan’s line: you pick the people who were going to succeed on the way in, and they succeed on the way out.
The spreadsheet died. Five years ago you could put a SaaS company on a spreadsheet and know if it was venture scale. That formula is gone and nothing has replaced it. What Ryan’s hunting now are models that couldn’t have existed before AI, leaning on the idea of Sarah Tavel’s “sell the work, not the software.” Charging two grand to replace an outcome instead of fifty bucks to rent a tool. One example is Teranox, which Ryan invested in, a company using AI to predict where the uranium is, raise money to buy the mining rights, drill, and if you’re right, sell the mine. Which leads to my favorite quote of the episode:
“Sure, they have zero revenue between now and when they find uranium. This is more like the very old venture, the king of Spain sending a fleet of ships across the ocean to find spices or gold.”
You’re renting the foundation. The whole Fable / Mythos thing also got me thinking. The internet was public infrastructure by birth. These AI companies’ models are not. AI is private infrastructure, and everyone building on it is a renter whose landlord has a cost floor well above zero and can change the terms whenever the IPO calendar says so. Mythos is the cautionary tale: it shipped, the U.S. government decided non-Americans couldn’t touch it, Anthropic killed it rather than fire its own team, and it came back the day before we recorded. Build your product on it and the floor drops out overnight. So the model was never the moat. It’s rented, and one agent swaps for the next without much friction. The defensible thing is being the team a customer trusts enough not to rip out and replace.
Sovereignty isn’t a hard drive. Every country got spooked and started asking whether it needs its own Mythos. Ryan’s answer: you can’t out-spend the frontier. A gigawatt data center runs $50–100B, few governments are signing that cheque, and the Chinese open-weight models are six to nine months behind anyway. The moment they ship, anyone can retrain and repackage one until, in his words, “I have my own Ryan’s Fantastic model.” He stresses he’s not in government, but the observation is simple: don’t run a race you can’t fund. Figure out what your country is genuinely good at and point that at the AI stack instead of trying to clone the frontier. Mining was his example for Canada, but the principle is to play to your strengths.
Finally, we get into how “sovereignty” is slipperier than “it’s on our soil.” My example (and I tracked down the reference now that the episode’s in the can): a French company, French data, a Microsoft stack sitting in France, and it can still be reached through the front door under U.S. law, because Microsoft is an American company.
When nobody can tell you what the business model will be in two years, you’re not underwriting a company, you’re underwriting a person or team, which honestly is what it’s always been about. Ryan signs off that you can email him at ryan@shn.vc and an AI will read it first, “depending on the version we’re allowed to use that day.”
Hot takes welcome, especially if you’re building a uranium AI company.
