Insights · 6 min read

How Long Does It Take to Sell Data to an AI Lab? Real Timelines

Nobody publishes average deal times. The stages are knowable, and most of the clock is in the ones you control.

Bearish AI · October 21, 2026

Owners ask how long a sale takes, and the honest answer starts with what is not known: we found no public benchmark of how long data sales to AI labs take, and the reporting we reviewed does not give one. Anyone quoting a fixed number without looking at your data is guessing. What can be described is the sequence of stages, which of them you control, and what reported cases show.

The stages and who controls the clock

The first three stages are the ones you own, and they tend to decide how the rest goes. A prepared sample with a data card gives a buyer’s team something to evaluate immediately. Missing documentation gives them a reason to wait.

What the reported cases show

The best-documented public example is not typical because it ran through a bankruptcy court. Reuters reported on 17 August 2026 that Google had won the auction for Spirit Airlines’ business data with a $10 million bid; the airline had stopped flying in May. A consumer privacy ombudsman recommended approval in a report dated 5 October, almost seven weeks later, after the parties agreed to remove booking and transaction records and to certify de-identification before each transfer. The court still had the final say. Private sales skip the court, but the same kinds of review exist: privacy, scope, and certification.

A second data point is smaller. MIT Technology Review reported a founder who sold a codebase for about $25,000 while payment for workspace data was still pending when the story ran. We cannot say from that how long the second sale took, only that one sale in a package can close before another.

Contracts also run on their own calendars once signed. Reddit’s prospectus described data licensing contracts with terms of two to three years, so a sale can be the start of a renewal cycle and not a single event.

What stretches a timeline

• Unclear rights. Contracts to read, third parties to ask, and exclusions to agree. The FTC has said it has required deletion of models built on unlawfully obtained data, so buyers do not skip this step.
• Many sources. Each source, such as chat, tickets or recordings, needs its own cleaning and its own removal counts.
• Personal data. More personal data means more removal, more sampling, and more legal review. EU data protection rules add their own conditions.
• Building tasks. Turning bug history into verified fail-to-pass tasks involves environment builds, repeated test runs and human review of tests. See /insights/bug-history-to-fail-to-pass-coding-tasks.
• Exclusivity. Interviews put exclusive deals at about four to five times the price of non-exclusive ones, and both sides then have more to define: field, term and derived products.
• A single buyer. One buyer’s queue is out of your hands. In our view, talking to more than one gives you more information and some protection against delay.

What “done” looks like at each stage

Define an exit test for every stage so you know when to move on. Rights are done when counsel has signed off on a list of what you can and cannot license. Preparation is done when a random sample of the output shows no secrets and no unreplaced identifiers, and you have counted the misses. The data card is done when each claim in it can be backed by a log or a document. The sample is done when it was built by the same pipeline as everything else. Delivery is done when the buyer’s acceptance test passes and the files and hashes you sent are recorded.

What shortens it

• A rights memo reviewed by a lawyer before the first conversation.
• A data card with sources, processing and removal counts. See /insights/data-cards-and-provenance-for-ai-training-data.
• A sample built exactly like the full set.
• A short list of terms you will and will not accept. See /insights/ai-data-licensing-agreement-terms.
• A named person on your side who can answer buyer questions the same day.

Questions to ask a buyer about their process

• Reviewers. Which teams must approve: research, legal, security or privacy?
• First look. What do you need to see first: a sample, a data card, a rights summary?
• Deal-breakers. What would make you stop? Knowing early saves everyone time.
• Personal data. Do you scrub again, and do you require certification?
• After acceptance. What are the delivery format, the payment trigger and the file custody?

Asking these before you start lets you plan around the buyer’s calendar, not guess at it.

A caution about deadlines

Companies winding down can feel pressed to sell fast. MIT Technology Review notes that founders have little visibility into who buys their data or how it is used, and that proceeds typically go to creditors and investors first. In our view, a rushed sale leaves little room to check rights, read the grant or ask for a sample review. If time is short, prepare in the order of /insights/prepare-company-data-for-ai-labs, and put rights first.

How to plan

Work backward from the stages above. Put dates on the ones you control and leave visible gaps for the ones you do not. Before you commit time and money, the free estimate at /estimate gives a conservative range so you can judge whether a longer, more careful process is worth it. For the full selling sequence, read /insights/how-to-sell-data-to-ai-companies.