The Data Dividend: Should You Sell Your Data?

In August 2026, Google spent ten million dollars to buy the inbox of an airline that had stopped flying. Spirit Airlines was in bankruptcy, its planes grounded since May, and the most valuable thing left in the estate was not a gate slot or an Airbus. It was the data. About 100 million emails, roughly 500 million Teams messages, some 30 million lines of code, plus the pricing models, the booking curves, the refund histories, the record of every small decision a low-cost carrier made for years. Google outbid a company called Mercor, which had offered 7.5 million. After the auction closed, a third bidder, Micro1, said it would have paid 12.5 million.

Sit with that for a second. Three sophisticated buyers looked at the exact same pile of a dead company’s messages and named three prices: 7.5, 10, and 12.5 million dollars. That is a sixty-seven percent spread on an identical asset. When three experts who buy data for a living cannot agree within two-thirds of each other, it tells you something simple and uncomfortable. There is no such thing as the price of a data set. There is only what a specific buyer thinks it can do with it.

I have sat on the other side of this table. I have had the conversation where a larger company, or lately an AI lab, floats a number for access to something my company generated just by operating. The number always feels like found money. It is revenue against an asset you were storing anyway, an asset that never showed up on the balance sheet, that you half-considered deleting to save on storage. Taking the check feels like getting paid for garbage. That feeling is the trap, and this piece is about how to see through it before you sign.

Three bidders, three prices, and the thing none of them will tell you

The Spirit auction is useful precisely because it is so clean. There is no ongoing relationship to protect, no brand to worry about, no employees to keep. The estate wanted the highest number and the buyers wanted the data. It is a pure read on how the buy side values corporate information in 2026, stripped of every soft consideration.

And the read is this: the buyers were not pricing emails. They were pricing what a low-cost airline knows. How to price a seat when a competitor drops fares on a Tuesday. How refund requests cluster after a weather event. How wifi and inflight purchases convert by route and time of day. What thirty million lines of code reveal about how a real operations team solved real problems under real cost pressure. Google did not want a museum of messages. It wanted the reasoning trace of a company that survived on thin margins in one of the most competitive businesses there is, so that its models could learn to reason that way too.

Here is the part that matters for anyone who still has a company that is alive. The seller in that auction was a bankruptcy estate, which has exactly one job: convert assets to cash today. It has no future to protect. You do. When you sell or license your data while your company is still operating, you are selling something to a buyer who has thought hard about what it can build with it, and you are usually pricing it as a line item you did not know you had. The gap between those two states of knowledge is the whole game.

I wrote a while back about whether your data even gives you a moat in the first place, the difference between data that just piles up and data that compounds into an advantage. You can read that argument in the piece on the data-moat test. This is the sequel question, and it is harder. Say you have decided your data is genuinely an edge. Now someone wants to buy it. What are you actually being offered, and what are you actually giving up?

The check prices your data as storage; the buyer prices it as a weapon

The single most important idea in this whole discussion is that every data deal has two prices, and they are almost never close.

The first price is what I call the export price. It is what the data is worth to you, sitting on your servers, as an asset you already have. And here is the strange thing: to you, that number is often near zero. You did not spend money to create most of it. It accumulated as a byproduct of running the business. The support tickets, the transaction logs, the internal chats, the code. You would keep it whether or not anyone paid you. Its marginal cost to you is a storage bill. So when a buyer offers you two million dollars for it, your brain does the math against zero and the deal looks incredible.

The second price is the capability price. It is what the data is worth to the buyer, which is a completely different quantity. The buyer is not valuing your data as inventory. It is valuing it as the thing that closes a gap. An AI lab training an agent to handle airline operations, or medical intake, or legal review, or whatever your company does, has a specific hole in its model. Your data fills that hole. The value is not the bytes. The value is the years and the access and the real-world outcomes those bytes represent, which the buyer would otherwise have to generate itself, slowly, or never.

Between the export price and the capability price sits a gap. That gap is the buyer’s margin. It is also your future. When you price against zero and they price against the capability they gain, any number they offer above your storage cost looks like a win to you, while any number below their true capability value looks like a steal to them. You can both walk away happy, and you can both be right about your own math, and you can still have made a bad trade. The picture below is the trade.

The offer you receive will land somewhere in that gap, close enough to your storage-cost anchor that it feels generous, and comfortably below the buyer’s ceiling so the buyer keeps the surplus. This is not because buyers are villains. It is because prices form where information sits, and in a data deal the information sits almost entirely on their side.

The only person in the room who does not know the price is you

There is a body of research on what happens to prices when one side of a transaction knows more than the other. The short version is that the party with less information consistently undervalues what it is trading. In studies where buyers can see the true quality of a product, they pay close to its real worth. Where they cannot, they discount heavily to protect against the risk of overpaying. Data deals run this dynamic in reverse and at full strength, because the seller is the one flying blind.

Think about what the buyer knows that you do not. The buyer knows which specific capability your data unlocks, because it has the model and it can measure the gain. The buyer knows what comparable data has sold for across dozens of deals, because it does these constantly. The buyer knows how much it would cost to generate this data synthetically or to buy it from someone else, which sets a real ceiling on what your data is worth to it. You know none of these things. You know your storage bill and a vague sense that the number on the table is more than zero.

One survey of tech companies found that most had never inventoried their data assets at all, and that a large majority were sitting on data worth several times what they assumed. That is not a story about greedy buyers. It is a story about sellers who show up to a negotiation without having done the one thing that would let them negotiate. If you have not measured what your data can do, you have delegated that measurement to the person writing the check, and that person has every reason to measure it quietly.

This is the same asymmetry I keep running into in other founder decisions. When I wrote about taking money off the table through a secondary, the danger was selling equity from a place of not knowing what your own shares were worth. Data is worse, because at least with equity there is a funding round and a 409A and a market of comparable companies. With data there is no ticker, no obvious comp, and the only expert appraiser in the building works for the buyer.

You can un-sign a customer; you cannot un-train a model

Most assets a company sells are reversible in some rough way. If you sell a building and regret it, you can buy another. If you take on a bad customer, you can fire them at renewal. If you license your logo for a campaign, the campaign ends. Data trained into a model is not like this, and the difference should change how you price it.

Once a data set is used to train a model, deleting your copy does nothing. The model has already learned the patterns. There is a whole field trying to solve this, called machine unlearning, and the honest summary is that it does not work well. Removing what a model learned from a specific data set is either prohibitively expensive, because it means retraining from scratch, or it is approximate and cannot be verified to have actually worked. So the buyers know this, and they write it into the contracts. Nearly every serious data license now includes a clause stating that any model already trained on your data survives the end of the license. The contract can expire. The capability the buyer built with your data does not.

That single fact breaks the mental model most founders bring to these deals. You think you are renting. You picture a two-year term, a renewal conversation, a chance to walk away later if it goes badly. What you are actually doing is selling, permanently, with a payment schedule that happens to look like rent. The term of the contract controls how long the buyer keeps sending you money. It does not control how long the buyer keeps the advantage. Those are two different clocks, and only one of them ever stops.

When I price a permanent sale in my head, the number I need goes up a lot, because I am not being paid for two years of access. I am being paid, once, for the rest of the asset’s useful life to someone else. If a buyer wants perpetual, worldwide, all-models rights, the correct response is not gratitude that the term is so simple. It is a much larger number, or a much narrower grant.

The person who buys your data rarely stays your only competitor

There is a second reason data sales are not like other sales, and it is about who ends up holding what you sold. When you license data to an AI lab, you are almost never the only source it is buying from, and the lab is almost never building something that will only ever help you.

Two things tend to be true in these contracts. The rights are non-exclusive, so the lab can buy the same category of data from your competitors too and blend it. And the output, the model, gets sold or served to whoever wants it, including your competitors and, increasingly, including a version of your own product. So the sequence goes like this. You sell your operational data because it feels like free money. The lab trains a capability on it. The lab offers that capability, as an API or an agent or a feature, to the entire market you compete in. Your data helped build the thing that lets a competitor do what you do without having spent the years you spent learning how.

This is the part founders miss because it happens off-screen and on a delay. The check clears in Q1. The competitive erosion shows up eighteen months later as a general capability that everyone in your category can now buy cheaply, and you never trace it back to the deal you signed, because by then it feels like just the way the market went. I have watched a version of this in the shift I described in the piece on service as software, where the thing you sold as a service becomes a feature in someone else’s software. Selling your data is one of the mechanisms that makes that happen faster, and you get paid a modest amount to accelerate it.

None of this means the buyer is acting in bad faith. A lab buying broadly and serving broadly is doing exactly its job. The point is that you have to price the deal as what it is: not a one-to-one transaction where you hand over bytes and get money, but a one-to-many transaction where you hand over a capability that will be distributed to a market that includes the people trying to beat you.

How data differs from every asset already on your balance sheet

It helps to line data up next to the other things a company sells, because the contrast is what makes the pricing mistake obvious. When you sell equity, there is a round, a valuation, comparable companies, and a market that will re-price your shares next year. When you sell a customer contract or a product line, the buyer gets what you had, and you no longer have it, which is at least an honest trade. When you sell real estate, the asset is still there afterward, just owned by someone else. Data breaks all three patterns at once. There is no clean comp, you keep a copy so it feels like you lost nothing, and yet the buyer walks away with something you can never fully get back.

That combination is unique and it is dangerous. Because you keep your copy, the sale registers as pure addition, money in with nothing out. Because there is no comp, you have no anchor except your storage cost. And because the capability the buyer trains is permanent, the one thing that actually left your possession, your exclusivity, is invisible on the day of the deal. You feel richer and you are, in cash, this quarter. What you cannot see is the exclusivity you spent, which only shows up later as competitors doing what you do.

I have made versions of this comparison before in other contexts. The shift from fixed salaries to metered agent costs that I described in the piece on agent costs and the margin trap is another case where a number that looks like one kind of thing, a simple bill, is actually another, a variable claim on your margin. And the way headcount stopped being a badge of scale, which I covered in the piece on revenue per employee, is the same lesson from a different angle: the obvious reading of a number is often the opposite of what it means now. Data sales sit squarely in that family. The table below is the comparison that keeps me honest.

Read that bottom row on its own. It is the only asset where you keep the thing and still lose it, and the only one with no market price to protect you. That is why the same instinct that serves you well selling a building will lead you astray selling data. The instinct says: the asset is still here, so I have lost nothing, so any price is a bonus. For data, the asset being still here is exactly the illusion that costs you.

The exhaust you would throw away is the record of how you win

The reason all of this sneaks past smart operators is a framing problem. We have been trained to think of operational data as exhaust. It is the smoke coming off the engine, the logs nobody reads, the tickets that get closed, the chat history that scrolls away. Exhaust is worthless by definition, so being offered money for it feels like pure upside. Who would not sell their smoke?

But the labs are not buying smoke. They are buying the engine’s behavior. The most valuable data in 2026 is not internet text, which has mostly been scraped already. It is the reasoning data that comes from professionals doing real work with real consequences. Doctors making calls, lawyers structuring deals, engineers debugging under deadline, an operations team pricing seats against a rival at midnight. That is exactly what your logs are. Your support tickets are a record of every way your product confuses people and how your best agents resolve it. Your transaction history is a map of how demand actually moves in your category. Your internal decisions are a trace of judgment that took you years to build.

Look at what got the highest bids in the data market. Mercor, a company whose entire business is sourcing exactly this kind of expert reasoning data for AI labs, reached around two billion dollars in annualized revenue and a valuation near twenty billion, paying its experts millions of dollars a day. The whole enterprise data licensing category exists because labs figured out that the trace of real work is the scarce input now. Your exhaust is their fuel. Calling it exhaust is how you talk yourself into selling the engine’s blueprint and feeling clever about it.

I am not saying operational data is priceless and you must clutch it forever. Plenty of data really is inert, and I will get to when selling is the right move. I am saying that the word exhaust is doing dishonest work in your head, and the first job is to replace it. Before you can price your data, you have to stop describing it as garbage.

Who should sell, and who is quietly selling their engine

So when is selling right? The decision comes down to two questions, and almost nothing else. First: is this data part of how you actually win, or is it genuinely inert? Second: does the buyer, or the market the buyer serves, compete with you? Put those on two axes and you get a map that tells you what to do.

The safe corner is the bottom left. Data that is genuinely inert, sold to a buyer who does not touch your market, is clean money. Take it, but even here do the work of inventorying what you have and getting a real market price rather than the first number offered. The danger corner is the top right, and it is where most of the regret lives. Data that is your actual edge, sold to a buyer whose model will serve your competitors, is you selling the engine that makes you special to the people trying to catch you. There is no price that fixes this, because the thing you are giving up is not measured in this year’s dollars. It is measured in how long you stay ahead.

The two off-diagonal corners are where judgment earns its keep. Top left is data that is your edge but goes to a buyer who does not compete with you. Here you can license, but narrowly: exclusive if you can get it, time-boxed, scoped to a use that will not boomerang into your market, and priced against the capability rather than your storage cost. Bottom right is inert data going to a competitor-adjacent buyer, where the cost to you is small but non-zero, so you negotiate hard and price in the fact that you are handing a rival a small gift. The table below is the same logic as a checklist, because a map is easier to nod at than to apply.

Refusing every offer is its own way to lose

I have spent most of this piece warning you off, so let me argue the other side honestly, because the reflexive hoarder makes a different mistake that is just as expensive.

Most data is not your edge. Founders wildly overestimate how special their logs are, in the same way they overestimate how novel their idea is. If your data is genuinely inert, or if the capability it would give a buyer is not one that will ever be pointed at your market, then refusing to sell is not discipline. It is leaving money on the table to protect a moat you do not have. The bankruptcy estate got it right for its situation. There are live companies whose data really is closer to smoke than to an engine, and for them a data deal is simply good treasury management.

There is also a timing argument that cuts against hoarding. Data has a shelf life. The record of how you priced seats in 2024 is worth more to a buyer training a model this year than it will be in 2028, when the patterns have shifted and newer data exists. If your data is the kind that decays, sitting on it in the hope of a better offer later can mean watching the asset melt. The question is never simply sell or do not sell. It is which slices, to whom, on what terms, and when, and the honest answer for a lot of your data is that a narrow, well-priced sale beats indefinite storage.

The mistake I want you to avoid is not selling. It is selling without knowing which of the four corners you are standing in, and pricing against your storage bill instead of their capability. A founder who sells the right data, hard, for a real number is playing the game well. A founder who refuses everything on principle is just choosing a slower version of the same undervaluation, where the number they miss is the whole check instead of the gap.

You cannot price what you have never counted

Underneath every bad data deal is a failure that happens long before the offer arrives: the company never took inventory of what it holds. Most firms cannot tell you which of their data sets are unique, which are derivative, which are decaying, or which encode something a competitor would struggle to reproduce. So when a buyer shows up with a number, there is nothing to check it against, and the number wins by default.

Taking inventory is not a data-science project. It is a founder-level exercise in knowing your own business, and it is part of the same operating discipline I keep returning to in the founder playbook for building AI-native. Walk your data and sort it into three piles. The first is data that is a byproduct of things anyone in your category also generates; that pile is close to inert and you can sell it with a light heart. The second is data that is unique to you but does not encode your edge, the operational trivia; you can license that carefully. The third is the small, precious pile that is the actual record of how you win, the pricing logic, the decision traces, the hard-won judgment. That pile is not for sale at any number a buyer will casually offer, because its value to you compounds and its value to them, pointed at your market, is the end of your lead.

The reason to do this before an offer arrives, not after, is that an offer changes how you think. Once a real number is on the table, the export-price anchor takes over and every pile starts to look sellable. Sorting your data in calm times, when no one is dangling a check, is the only way to make the danger pile feel as untouchable as it should. Founders who have done this walk into the negotiation already knowing which corner of the map they are in. Founders who have not are pricing in real time against a counterparty who did the homework months ago.

How to price a data offer before you answer it

Here is what I actually do when a number lands on the table, in the order I do it, before I say anything back.

First, I inventory what is being asked for, not what I think I am selling. Buyers ask for broad grants because it costs them nothing to ask. The request will often say all data, all fields, perpetual, worldwide, sublicensable. I map that against what I would sell if I were being careful, and the difference between those two is my first round of negotiation. Most of the value I protect, I protect by narrowing the grant, not by raising the price.

Second, I try to estimate the capability price, not just my export price. I cannot see the buyer’s model, but I can ask the questions that reveal it. What are you training? Who will the output serve? Is this exclusive? Would you pay more for exclusivity? The answers tell me whether I am in the danger corner or the safe one, and roughly how much the buyer would pay if pushed. If the buyer will not answer any of it, that opacity is itself information: they are protecting a large gap.

Third, I price the deal as permanent, because it is. I do not let a two-year term trick me into pricing two years. I ask what this capability is worth to the buyer over the life of the model, and I anchor my number closer to that, then let the term structure the payments however they like. If they want the capability forever, they can pay for forever.

Fourth, I check the second-buyer question explicitly. I ask myself who else can access what this deal creates, and I assume the answer is everyone, including my competitors, unless the contract says otherwise in writing. If the deal helps arm my market against me, the price has to include that cost, or the grant has to be scoped so it cannot.

Fifth, I decide with the map, not the mood. The offer will feel like found money. That feeling is the export-price anchor talking. I make myself name the corner I am in before I let the number move me, because the number is designed to move me and the corner is not.

None of this requires a data-valuation consultant or a six-month project. It requires refusing to price against zero, treating the sale as permanent, and asking who ends up holding what you sold. Do that much and you will already negotiate better than the large majority of sellers, who show up having never inventoried what they have and take the first number because it beats a storage bill.

The Spirit estate had no future to protect, so its only job was to find the highest bidder, and it did. You have a future, which means your job is different and harder. It is to know, before you answer, whether the thing you are being offered money for is smoke or the engine, and to price it accordingly. I keep coming back to a related question about what parts of a business the frontier will eventually absorb on its own, which I worked through in the absorption test, and it applies here too. Some of your data the market will get to anyway. Some of it is the reason you are still ahead. Selling those two things at the same price is the error, and the error is easy to make because both checks are green.

Questions founders ask about selling data to AI companies

How much is my company’s data actually worth?

There is no single number, which is exactly the problem. The same data drew bids of 7.5, 10, and 12.5 million dollars in the Spirit auction, because value depends entirely on what a specific buyer can build with it. The training-data market runs from a few million to a quarter of a billion per deal. What matters is not a market rate but your own answer to two questions: is this data part of how you win, and does the buyer serve your competitors. Until you have inventoried what you hold and estimated what capability it gives the buyer, any number you name will anchor on your storage cost, which is the wrong number.

Is selling my data the same as selling any other asset?

No, and the difference should change your price. Most assets are reversible in some rough way. Data trained into a model is not, because you cannot untrain it. The contract term controls how long you get paid, not how long the buyer keeps the advantage. Price a data sale as permanent even when the contract looks like a lease.

What is the data dividend, and why do you call it a trap?

The data dividend is the feeling that a data check is free money, income against an asset you were storing anyway and might have deleted. It is a trap because it makes you price the data against zero, your storage cost, when the buyer is pricing it against the capability it gains. The gap between those two numbers is real value moving from you to the buyer, and the dividend framing is what hides the gap.

Should I ever refuse to sell my data?

Yes, in one specific case: when the data is genuinely part of how you win and the buyer’s model will serve the market you compete in. That is the danger corner, and no price fixes it, because what you give up is measured in how long you stay ahead rather than in this year’s dollars. Outside that corner, refusing everything is its own mistake, since most data is closer to inert than founders think and some of it decays if you hoard it.

What contract terms matter most in a data licensing deal?

Three. Exclusivity, because non-exclusive means the buyer can blend your data with your competitors’ and serve the result to everyone. Scope, because broad grants like perpetual, worldwide, all-models mean you license everything once and can never license it again. And survival, because nearly every deal says models already trained survive the license ending, which is what makes the sale permanent. Narrow the grant before you raise the price.

Why do AI labs want boring operational data at all?

Because the scarce input now is not internet text, which has been scraped, but the trace of real work with real outcomes. Your support tickets, transaction logs, and internal decisions are a record of how a real team reasons under real constraints, which is exactly what labs need to train agents. That is why an entire expert-data market has grown to billions in revenue paying professionals for their reasoning. Your exhaust is their fuel.

How do I estimate what my data is worth to the buyer?

You cannot see their model, but you can ask the questions that reveal the shape of it. What are you training, who will the output serve, is this exclusive, would you pay more for exclusivity. The answers place you in the decision map and hint at the buyer’s ceiling. If the buyer refuses to answer any of it, treat that opacity as evidence of a large gap you are on the wrong side of.

Is a one-time data sale better than building a data product?

Usually the more durable move is to sell answers rather than raw data, embedding your data into a product that stays yours while the insight is what the customer pays for. A one-time sale hands over the asset; a data product keeps the asset and rents the value. Sell the raw data outright mainly when it is inert, when it decays quickly, or when a narrow high-priced license clearly beats the slow build of a product you were never going to ship.

I am a four-time founder building AI-native companies. I write about the founder mindset, AI opportunities, and the mechanics of building with AI at vikasmalpani.com.