Updated
The data shortage, explained
AI Companies Running Out of Training Data
Short answer
Not yet, but it could happen. In 2024, a research group called Epoch AI guessed that AI could use up all the public writing online sometime between 2026 and 2032. So AI companies now pay for writing that is not online, like a business's old email and files.
The idea of AI companies running out of training data comes from a real forecast. It is not a fact yet. In 2024, a research group called Epoch AI looked ahead. Its guess was that AI models may soak up all human-written public text. That could happen between 2026 and 2032.
Labs are already paying for new sources. One of the few big sources left is private. It is the records that long-running businesses keep.
Below, we explain the forecast in plain words. We cover what the AI field is doing about it. We also show why it matters if you own a firm with years of email, tickets and files in storage. At the end, a two-minute estimate shows what a license to your records might be worth. A license hands the buyer a copy to train on. Your firm still owns every original.
- Sourced, not hyped
- Free estimate
- Your records, your choice
The forecast
What did researchers actually predict?
AI language models learn by reading huge amounts of text. Researchers count that text in tokens. A token is a word or a piece of a word.
Epoch AI ran the numbers in June 2024. By its count, the useful public text written by people adds up to about 300 trillion tokens. It also said that number could be off by a lot.
The study looked at how fast models were growing. At that pace, the biggest ones would be trained on nearly all of it between 2026 and 2032. The exact year depends on how labs train. Some labs push each model through far more text than normal. If they do, the limit could come even sooner.
Epoch took care to say this does not have to mean AI stops getting better. The authors named three ways around the limit. First, AI can make its own data. Second, models can learn from other kinds of data, like pictures and video. Third, labs can get more out of each piece of text.
So it is a forecast with a range, and you should read it that way.
People who build these systems have said much the same. Ilya Sutskever is a co-founder of OpenAI. In December 2024, he gave a talk at NeurIPS, a big AI research meeting. He said data is like fossil fuel, because there is only one internet.
He told the crowd the field had already reached peak data. In his view, no new supply of that kind is coming.
The response
What are AI companies doing about the shortage?
No single fix has won yet. Labs are trying all of these at once. Each one has a cost or a limit.
Paying to license content
Labs used to grab free web pages. Now they also sign paid deals with sites and archives that own lots of writing and pictures. Reddit told investors its data licensing contracts were worth $203M in total. Shutterstock earned $104M in one year from AI makers.
Paying people to create and check data
Labs hire experts to write answers, grade what a model says and fix its mistakes. Meta put $14.3B into a 49% stake in a firm that sells this kind of human-made training data.
Making synthetic data
Synthetic data is text that one AI model writes to train another. Microsoft said its Phi-4 model used synthetic data throughout training. But there is a known risk. A 2024 study in Nature looked at models trained again and again on AI-made text. Those models lost the rarer details of real data. The authors called this model collapse.
Licensing private business records
Real work records, like email threads and support tickets, never showed up on the public web. Labs, and the firms that supply them, now pay companies for the right to use copies of them.
Follow the money
What the shortage looks like in dollars
Each figure comes from a public filing or a news report. Together they show that buyers now pay real money for data they once took for free.
The total value of the data licensing contracts Reddit signed in January 2024. Reddit shared it before its stock market listing.
SEC S-1, Feb 2024Reddit's other revenue, first in 2023 and then in 2024. It jumped once Reddit began licensing its posts. That revenue line is mainly content licensing.
Reddit 10-K, SEC, Feb 2026What Meta paid for just under half of one supplier of AI training data.
TechCrunch, Jun 2025Google's winning bid in a court auction for Spirit Airlines' store of staff email and Teams chats. The court has yet to approve it.
Axios, Aug 2026Amazon, Microsoft, Alphabet and Meta expect to spend this much in 2026, mostly on AI. It shows how much money is chasing better models.
Sherwood News, Apr 2026One AI data company pledged in October 2026 to spend this over 12 months. The money goes to buying or licensing work records from companies.
Company announcement, Oct 2026Where you come in
Why does a shortage of public text make business records valuable?
The forecast is about public text, the kind anyone can download. Your company's records are the opposite. They sit behind passwords in your mail server, your help desk and your shared drives.
No web crawler has ever read them. A crawler is a program that scans the web to collect text. So no AI model has learned from your records yet.
Your records also show something public writing rarely does: work as it really happens. A ticket starts with a customer's problem. It goes through a few wrong guesses and ends with the fix. An email chain shows how a manager deals with a late load or an unhappy client.
AI companies want to build tools that do that kind of work. To learn it, they need real examples.
Only the business that owns those records can grant a license to them. Say your company has 20 or more people and several years of history. Then you have something a lab cannot get anywhere else.
Our main guide on licensing company records to AI firms covers the basics. Our page on where AI companies get their training data shows how private records fit next to every other source.
What qualifies
Business records that help fill the gap
Buyers ask about these kinds of records the most. Most companies have at least three of them.
| Record | Typical home | What it adds that public text lacks |
|---|---|---|
| Customer and vendor email | Outlook, Exchange, Gmail | Private business talks that go both ways, over years |
| Chat channels | Slack, Teams, Google Chat | Quick, rough choices made between coworkers |
| Help desk history | Zendesk, Freshdesk, Help Scout | Problems paired with the answers that solved them |
| CRM notes | Salesforce, HubSpot, Zoho CRM | How deals move and why some are lost |
| Contracts, quotes and reports | Google Drive, SharePoint, Dropbox | Expert writing on topics the web barely covers |
| Job and work order records | ERP, dispatch or field service tools | Real tasks broken into steps, with how they turned out |
| Call recordings | RingCentral, Zoom Phone, Aircall | Natural speech in a work setting |
| Training manuals and code | Confluence, Notion, GitHub, GitLab | How a company teaches and builds |
Curious what your share of the shortage is worth
Our estimate starts from price formulas posted by AI data buyers. It blends them, then rounds the result lower. The quiz takes about two minutes, and you upload nothing.
Know the limits
What to keep out of any license
High demand does not mean it all should go. You choose each system, folder or year. The things you leave out get written into the contract.
Sensitive personal files
Staff records, ID copies, bank details and medical notes about workers should stay home. Buyers do scrub names from their copy. But no scrubbing method is perfect.
Records a regulator controls
Some records have their own laws. Patient details fall under HIPAA, the US health privacy law. A law firm must keep client files secret. Records like these need special care, or they should stay out fully.
Data you promised to protect
A customer contract or privacy notice may limit how you use certain records. If so, keep that promise. Leave those records out, or get permission first.
Things competitors would love to see
You can leave out current price lists, product plans and special methods by name. That holds even if a buyer would pay more for them.
From question to payment
How a license deal works with us
We are a data agency. We steer owners toward the top-paying programs for their kind of records. Whichever program you join covers our referral fee on its own dime. Your records never pass through our hands.
Run the estimate
Tell us your field, headcount, years in business and which records you keep.
About 2 minutesWe rank your matches
We test your answers against the rules and prices of each paying program. Then we show the 3 that would pay a company like yours the most. We tell you why each one fits.
SecondsPick one and sign up
Each match says what it buys, when it pays and what the contract covers. You sign up with the program in one click. Or someone from the program reaches out to you.
One clickCopy, clean, pay
The buyer copies only what you okayed. It removes personal details and sends payment. Payment can take weeks or a few months.
Per buyerWhy now
A shortage is a seller's market, for a while
Prices depend on what labs lack today. That can change, and nobody knows just when.
New kinds of data fetch the most
The first companies in a field to license their records fill a gap the buyer has. Later ones fill a gap that is already half closed.
Labs are working on other answers
Synthetic data and better ways to train keep getting better. If they work well enough, demand for some kinds of records could drop.
Others in your field can go first
A rival who licenses records like yours before you may take the best price in your field.
Old files keep vanishing
Each year, old mail and chat get wiped out. Auto-delete rules, ended plans and software changes all do it. A record that is already gone can never earn a license fee.
Questions
Questions owners ask
Are AI companies really running out of data?
They are running short of new public text written by people. A 2024 Epoch AI study put the likely limit between 2026 and 2032. It is a forecast, and labs are working on ways around it. But it explains why they now pay for data they used to collect for free.
Will synthetic data make my records worthless?
Nobody knows how far synthetic data will go. A 2024 study in Nature found that training only on AI-made text can wear away detail over time. That is one reason real human records still matter. Demand could change, so we suggest you not wait years to find out.
Why would an AI company want our boring old email?
It shows how real work gets done, which public web pages rarely do. Everyday threads about orders, delays and fixes are just the examples AI tools need. They help the tools learn to handle office work.
Do we have to give up ownership?
No. You grant a license. That is permission to use a copy for the uses you agree on. The originals and the ownership both stay with you.
How much can a company get?
AI data buyers post per-company payouts of roughly $100K up to $5M+. Your estimate narrows that down using your size, age and kinds of records. Treat it as a first look, since only a buyer can make a real offer.
Is our company too small to matter?
Many buyers want a team of 20 or more and a few years of stored records. More people and more years most often mean bigger offers. Closed or closing companies can count too.
Does it cost anything to use DillisData?
No. Our fee comes from the programs, and only after a deal closes. You owe DillisData nothing.
Your next step
Turn the data shortage into an offer for your company
The labs need text they cannot find on the open web. If your company has years of it, the estimate tells you about what that could bring.
No cost to you. Nothing to upload. Your top 3 matches, with reasons.