The VC case for ‘full stack deeptech’

Start with the pass that matters most, selling technology to incumbents, which is what most deep tech startups in industrial categories do because the incumbents own the infrastructure and the distribution. Rountree says it looks like the easy path precisely because it scopes the company down to productizing one piece of technology for one big buyer. What kills it is not that you lose the argument; it is the clock. Founders raise serially, typically 18 to 24 months of runway, and have to start fundraising about six months before the money runs out, which leaves roughly 18 months to produce whatever the next round requires. If the customer’s sales cycle is longer than that, the company is dead by arithmetic rather than by merit. His example is a 2016 investment with strong machine learning and interoperability software selling into carmakers, which died with good investors and good people because Ford and Toyota purchasing outlasted the round. Kann adds pilot purgatory: big companies happily pilot, and the gap between pilot and commercial rollout is never a knowable number. Rountree’s practical advice is to work out how the buyer is personally compensated, because incentives explain most otherwise baffling behavior and tell you whether the sales cycle is compatible with your company at all.

The second reason is price rather than time. Even where an incumbent agrees the technology is ten times better and wants it, the offer comes in far below what the seller thinks it is worth, and pushing back does not move it; if you are not in the pole position you do not get paid for the value you create. That leads into the second pass, commercializing science. Rountree’s view is that a business exists to earn margins above its cost of capital and technology is only a means to that, so being mind-blowing and widely cited is not the same as being economically valuable. The founder failure mode he describes is a hammer looking for nails, common among people commercializing their own doctoral work, when the discipline required is obsession with a problem and indifference about the solution. There is also a timing objection: he would rather fund a novel application of an existing technology than wait for something to climb the maturity ladder, and says he has lost money waiting. Fusion is the clean test case, and he is explicit that he does not invest in it and cannot wait for it, while hoping for humanity’s sake that it works. The exception he grants is biology and pharmaceuticals, where the steps are standardized enough that an entire capital market has learned to underwrite technical maturity, and large pharmaceutical acquirers buy on a discounted cash flow rather than on technical milestones. Kann asks why that cannot be copied elsewhere; the answer is that no other field is scientifically and regulatorily defined enough for a capital market to form around it.

The positive archetype follows from all of that. Full stack is borrowed from a 2015 Chris Dixon post at Andreessen Horowitz about companies like Uber, Lyft and Airbnb that used software to reshape an industry instead of selling software to it, extended here to companies that build physical things as well. The operative half is selling the end product: work out who buys the thing your technology improves, and sell them that thing. Mining is the worked example. Startups offering satellite imagery and AI to help miners find deposits found that miners would not pay much, so Earth AI, one of Cantos’s largest investments, bought mineral rights, hired geologists and drilled its own hypotheses, and now holds deposits rather than a software licence. That costs more capital and more operational risk. Commodities, usually a venture pejorative, are for the same reason his favourite case. He frames a startup as carrying technical, execution and market risk, and argues that if you are taking extra technical risk in hard tech you must offset it with less market risk or you are simply making worse investments. A product molecularly identical to its competitors’ but cheaper to make has essentially no market risk, because the spot price tells you exactly what you have to beat. Kann qualifies this, noting commodities are less fungible than they look, since where and when an electron is delivered matters enormously, and points to Crusoe as a company whose differentiated input, flare gas, gave it a structural advantage in a commodity business.

The second archetype, the weird company with no comparables, gets the thinnest support in the conversation. The rationale is competitive rather than technical: in a crowded area you fight other companies for talent and capital, whereas a genuinely singular company pulls both toward itself. The evidence offered is two anecdotes, and Rountree concedes that weird usually means controversial, so the strategy requires disagreeing with people. Both archetypes raise the bar on the founder rather than lowering it, since vertical integration means raising more money, so he looks for people who are both real technologists and fluent enough to pull capital toward them. It is worth being precise about what this frame is and is not. It is a rule for what produces venture-scale returns on a pre-seed fund’s clock, not a claim about which technologies matter, which is why fusion can be both a pass and something he hopes succeeds.