Which Rents Survive

Frontier AI does not erode moats. It reprices every rent that was earned because thinking was expensive — all of them, at once. Here is the screen for what is left.

By VK Singh Vashisht — Datura Labs

---

Most of what a small investment firm earns is not a return on its assets. It is a return on the fact that thinking is expensive for everyone else. Diligence costs money, so a weird asset trades at a discount. Analysis takes a team, so an analytical edge pays. Drafting takes a lawyer, so a fast, clean offer closes ahead of a slow one. Every one of those spreads has a cognition-scarcity term buried in its price, and none of them were labeled that way, because for the whole history of the business the term never moved.

It is moving now. When frontier-class capability — the kind that reads a filing, prices an oddity, or drafts a counter in minutes — becomes commonplace among a firm’s competitors, counterparties, and adversaries, the scarcity term strips out of every rent simultaneously. Not the strong rents first and the weak ones later. All of them, on the same clock, because they were all priced against the same scarce input.

I ran this screen on my own firm before I ran it on anyone else’s. Most of what I thought were moats were not. That finding is the reason Datura Labs is built the way it is, and it is the reason this essay sits under a theme called Destructive Dharma: what diffusion destroys was never really yours. It was a rent on other people’s expense.

The comfortable story, and who tells it

The consensus story about AI and competitive advantage is a story about assets. Your capital survives. Your patience survives. Your relationships survive. Your proprietary data becomes more valuable. Therefore your moat is intact, or even widening, and the correct move is to adopt the tools and keep doing what you did.

That story is not wrong about the assets. It is wrong about what the assets earn. And it is told, with great confidence, by three parties who need it to be true. Incumbent firms, because their fee structures were priced in the old market and repricing them is an admission. The vendors selling AI adoption, because “you’ll do the same thing faster” is a purchase order and “the thing you do no longer commands a premium” is not. And every principal, myself included, because the alternative is admitting that a decade of edge was substantially an artifact of everyone else’s overhead.

The tell is the vocabulary. “Moat” is a word about assets. Nobody who says it is asked to name the mechanism by which the asset produces a return, or what happens to that mechanism when the input it depended on becomes free.

Grade rents, not assets

The correction is one sentence: grade the rent, never the asset.

A rent is a return earned because something is scarce for someone. Name the something and name the someone and you have the mechanism. “This earns because diligence is expensive for retail buyers.” “This earns because a distressed seller cannot price his own asset.” “This earns because incumbents take three quarters to deploy what I can deploy in a week.” Once the mechanism is on the page, the question of what diffusion does to it is usually not a matter of opinion.

Run each rent through six questions.

Is it a cognition-cost artifact? If the rent exists because analysis, diligence, or drafting is expensive, it converges. The weird-asset discount is the canonical case: the aversion was never to the asset, it was to unpriceable risk, and when a two-hundred-thousand-dollar oddity can be priced for the cost of a lunch, the discount compresses toward zero.

Does the counterparty get the tools too? Sellers, brokers, and opposing counsel are not exempt from diffusion. A spread that feeds on their inability to price shrinks from their side even where your skill is untouched.

Is there a floor effect? A structure already at near-zero overhead has nowhere to improve while every competitor converges toward it. What looked like a permanent cost advantage was a head start.

Is there an adoption asymmetry? Deploying in a week against incumbents who take quarters is a real rent — and a wasting one. That gap is the window. It closes on the incumbent’s compliance calendar, not yours.

Is it possession? Records, relationships, licenses, standing facilities, physical presence, accumulated reputation. These cannot be downloaded. They hold. And where capability parity makes private data the entire remaining asymmetry, they do more than hold — they amplify, because everyone now has the same engine and the only thing that differs is what it is pointed at.

Is it two-sided? Diffusion cuts your cost and your adversary’s. In any adversarial process — a dispute, a negotiation, an auction — the structure may hold while the rent thins, because both sides’ cost-per-round just fell together.

Six questions produce three verdicts. Converge: the rent was priced on scarce cognition and the scarcity is ending. Durable: the rent rests on possession and holds. Amplify: the rent rests on possession that capability parity makes more decisive, not less. Nothing gets “untouched” without surviving an explicit attack — the first time I ran this on my own book, the attack pass flipped four of six verdicts I had marked safe.

What the screen said about my own firm

Datura’s original offer catalog was six capability classes: things the firm could do that a family office or small investor could not easily do for itself. Every one of them graded converge. Not eventually — on the same clock as everyone else’s capability rents, because they were the same rent. Building a retrieval system, classifying deal signals, drafting a memo, reconstructing a capital history from statements: all cognition-cost artifacts, all priced against a scarcity that frontier models are ending.

What survived the attack pass was not a capability at all. It was a discipline layered on top of capability, and it survived because it answers a question that capability parity makes more urgent, not less.

The first survivor is verification. When anyone can produce the analysis, the scarce thing is proof that the analysis is right — and, harder, proof that the analysis was honestly withheld when the evidence did not support a conclusion. A receipt on every run, an exceptions ledger for what the engine refused to guess, a measured “nothing found” that can be audited rather than trusted. That rent rests on the discipline, not the model, and the discipline is possession.

The second is stewardship on a cadence. A screen that runs once is a capability. A screen that runs every day against written criteria, with every decision logged and every pass still searchable, is an operating record — and the record is what compounds.

The third is sovereignty of deployment, and it carries a finding I did not expect. When the pipeline runs on hardware the client owns, against the client’s own corpus, the amplifying rent — private data times a common engine — accrues to the client, not to Datura. The firm does not own the moat it builds. I decided to say this out loud rather than obscure it, because it is the honest architecture and it is also the sales argument: a vendor who keeps your corpus is keeping your moat.

That read is why the site is shaped the way it is. The marketplace gate — a storefront opens only after the machine has produced a verifiable sample deliverable in that vertical — is not a marketing sequence. It is the screen’s conclusion made operational: do not sell a capability, because the capability converges; sell the proof, because proof is possession.

Clocks, not a window

One global “AI timeline” is a vibe. The rents on a book close on different clocks because the counterparties adopt on different clocks. A distressed seller in a secondary market fields frontier tooling later than a bulge-bracket desk; a regional broker later than a national platform; a solo practitioner later than a firm with an innovation budget. The binding constraint on any position is the earliest clock touching its largest rent. Build that table per counterparty class before deciding which rent to harvest, which to exit, and which possession to deepen. The firm that skips this step is holding a portfolio priced on an average that describes none of its positions.

The one action that falls out of a completed screen is always singular: harvest a rent that is eroding, exit one that is gone, fortify a possession that holds, acquire an edge that diffusion has made cheaper to exploit, or spend an adoption window converting velocity into possession before it decays. Never a menu. A menu is what a firm produces when it has not finished the screen.

The bet

My bet is that within three years the phrase “our moat” in a small-firm deck is read the way “our synergies” was read after 2001 — as the sound of someone who has not named a mechanism. The firms still earning will be the ones whose rents were rebuilt on possession: records, standing, cadence, verification, sovereignty. The rest will be doing the same thing faster for a spread that no longer exists.

What would prove me wrong is diffusion being slower and lumpier than the priors say — counterparty adoption stalling on compliance and habit long enough that cognition-cost rents persist for a decade rather than a few years. That would make the screen premature, not wrong; the rents would still be repriced, just later, and the firms that rebuilt early would have paid a cost of carry for being right on the mechanism and early on the clock. I am carrying that cost on purpose. A rent I can name a mechanism for is one I can defend. A rent I can only call a moat is one I have already lost and not yet noticed.