The picks-and-shovels thesis

The oldest cliché in venture investing happens to be true. In the California gold rush, most of the miners went broke; the people who did reliably well were the ones selling the tools — the picks, the shovels, the denim, the assay services — to everyone digging, regardless of whether any particular claim panned out. The phrase has hardened into a slogan, which is a shame, because the reasoning underneath it is precise and it recurs in every technology cycle. The bet on the miner is a bet on which hole has gold in it. The bet on the shovel is a bet only that people will keep digging. The second bet is easier to be right about, because it does not require you to pick the winner. It requires you to notice a need that every contestant shares.

I want to make the case, in the plain language of an investment memo, that decision accountability belongs in the second category. It is not a vertical application competing to be the best agent for legal review, or claims adjudication, or loan underwriting, or content moderation. It is a horizontal layer that sits beneath all of them — the record that makes an automated action reconstructable and defensible after the fact. And I want to argue that this is the stronger position for exactly the reasons infrastructure usually beats applications in a platform shift, while being honest about the ways it is harder.

The pattern: durable value accrues to the horizontal enabler

Look at any technology gold rush and the same shape appears. In the web era, most of the volatile bets were on the sites and the storefronts — the miners. The value that compounded most durably accrued to the horizontal enablers everyone needed regardless of which site won: the compute you rented, the payments rail you charged through, the identity layer you logged in with, the observability tools you watched your systems through. These layers share a family resemblance. Each answers a need that does not depend on the outcome of the competition happening above it. A payments processor does not care which merchant wins; it takes its cut from all of them. An identity provider does not need to know which app succeeds; every app needs to know who its users are. The horizontal layer is insulated from the churn above it because its customers are the contestants, plural, not any one contestant.

This is the analytical core of the picks-and-shovels pattern, and it is worth stating carefully because it is easy to mistake for mere caution. The point is not that infrastructure is safer in the sense of duller. The point is that the horizontal enabler is exposed to a different and better risk. The app is exposed to the question "will this particular use case win?" — a question with a high failure rate, because most particular use cases do not. The infrastructure is exposed to the question "will this class of activity keep happening at all?" — a question that, once a platform shift is genuinely underway, tends to answer itself.

Why accountability is horizontal

Now apply the test to decision accountability directly. The question is whether the need it serves is specific to a domain or common across all of them. And the answer is that it is about as horizontal as a need can be. Every consequential automated decision — in every vertical, made by any model, wrapped in any agent — generates the same obligation: someone, eventually, will ask why it happened, and the system will need to answer in a way that can be checked. A denied claim, a rejected application, a flagged transaction, a suspended account, a triaged patient, a routed dispute: these are wildly different in subject matter and completely identical in this one respect. Each is an action that acquired consequences, and each will at some point have to be reconstructed, explained against the rules in force at the time, and defended to a party with standing to contest it.

That need does not depend on which agent wins. It does not depend on which foundation model is underneath, or which framework orchestrated the call, or which vertical turns out to be the big market. Whatever wins, the winner will have made decisions, and those decisions will have to be accounted for. A layer that records what an automated decision actually did — the evidence it consulted, the rules that were active, enough state to reconstruct and replay it, and a form the affected party has standing to challenge — serves every one of these cases without caring which one it is serving. That is the defining property of a pick-and-shovel. It is why the record is insulated from the application-level thrash: the agents can rise and fall, consolidate and fragment, and the requirement to account for their actions survives all of it, because the requirement was never about the agents in the first place. It was about consequences, and consequences are domain-independent.

You do not have to know which agent wins to know that whoever wins will have to account for what it did — which is why the record is a better bet than any of the agents.

The strategic shape of an infrastructure bet

Infrastructure of this kind carries a characteristic profile, and naming it clarifies both the reward and the risk. Three properties tend to travel together. First, it standardizes around a shared interface — the thing every party has to agree on. For payments it was the transaction; for the web it was the request. For accountability the natural interface is the decision record itself: a common shape for what "an account of a decision" contains, so that a record produced under one agent can be read, audited, and contested by a party operating under another. Infrastructure that everyone touches pushes toward a shared format, because the whole value is in the interoperation.

Second, it benefits from neutrality, and in fact requires it. A horizontal layer must serve competitors at once — the party that made the decision and the party contesting it, the vendor and the regulator, rivals who agree on nothing except that they need the same substrate. This is a genuine strategic asset. A record layer that visibly served the interests of the decision-maker over the affected party would not be infrastructure; it would be an advocate, and advocates do not get adopted as neutral ground. The neutrality that looks like a limitation — you cannot tilt the scale — is what lets the layer be trusted by everyone who has to rely on it, which is the precondition for it becoming standard at all. I have written elsewhere about how a shared, neutral standard for what counts as an admissible record compounds in value with every party that adopts it, the way a language or a rail gauge does.

Third, it compounds with adoption. The more decisions are recorded in a common, contestable form, the more valuable the form becomes — because comparability, auditability, and the plain expectation that a decision comes with its account all strengthen as coverage grows. A single accountable decision is a courtesy. A whole market of them is an institution.

The honest tradeoffs

None of this makes the infrastructure bet a free lunch, and a memo that pretended otherwise would not be worth reading. The picks-and-shovels position has real disadvantages, and they are the mirror image of its advantages. It is slower to monetize. An application can charge a vertical a premium for solving that vertical's acute, specific pain today; a horizontal layer typically earns less per interaction and has to make it up in breadth, which takes time to assemble. It requires reaching critical mass before its best properties switch on. The network effects and the standardization that make the layer durable are exactly the things that do not exist on day one — early, the neutral horizontal layer can look thinner and less impressive than a sharp point-solution, precisely because it has refused to specialize.

And it faces a constant, seductive temptation: to become just another app. The fastest revenue is almost always in going vertical — picking the most lucrative use case, building the full application for it, and capturing the margin. Each such move is locally rational and globally corrosive, because every step toward becoming the best agent for one domain is a step away from being the neutral substrate for all of them. The discipline the infrastructure play demands is the discipline to leave near-term application revenue on the table in order to protect the horizontal position — and that is a hard thing to hold to when a single vertical is waving a check. These are not reasons against the bet. They are the price of it, and they should be underwritten with open eyes: the infrastructure position pays later, needs scale, and must resist its own most tempting shortcut.

What makes the price worth paying is the shape of the eventual payoff. The reward for betting on the miner is capped by one claim and exposed to the risk that the claim is empty. The reward for the shovel is proportional to how much digging happens in total, across every claim, for as long as the rush lasts — and it does not require having guessed which miner would strike. In a genuine platform shift toward automated action, the volume of consequential machine decisions is going to grow no matter which agents, models, or verticals come out on top. The reliable position is not to guess the winner. It is to own the thing the winner cannot skip. Whoever wins will have to account for what it did, and the record is where that accounting lives. That is the case for the shovel over the gold.

— Dispatches · Summit Cognitive