This is not a thought experiment. Buyers are already writing checks for how real businesses operate.
The public internet is exhausted as training fuel. The next datasets are the ordinary records of running a business. We collect the coverage so you can see the market without living on X.
micro1 says it will spend $1 billion over the next 12 months acquiring and licensing enterprise operational data through its Company Data Partnerships program, financed with capital from Citi and Hercules Capital. The de-identified data trains AI agents in environments built from real business workflows.
A billion dollars of announced demand, not money paid out yet. Not every company will qualify, and buyers still pay for prepared, rights-cleared records, not raw exports.
Frankfurt Kurnit Klein & Selz says operating companies are now getting the calls shutting-down startups got first. Most cannot sell everything a buyer asks for: NDAs, vendor contracts, privacy notices, labor law, and Slack's API terms all limit what can move.
This is the unglamorous half of the market. Inventory, rights, exclusions, and paperwork come before any price. That is the work we do before a buyer sees anything.
The court-appointed ombudsman told the bankruptcy court that Spirit and Google have taken adequate steps to remove or de-identify customer personal information from the enterprise data Google wants for AI training. The sale is still a proposal.
The approval path ran through privacy, not price. Show your exclusions and de-identification up front and a deal can clear review.
Led by Oak HC/FT. CEO Jerry Wu says four of the top five closed-source U.S. labs pay for its training environments. Its due-diligence benchmark was built from anonymized private-equity deals; the best model averaged 51%.
Labs are paying to rebuild real professional work, and the raw material is real deal files and email threads. A firm's own history of doing the job is the input.
The Third Circuit affirmed for Thomson Reuters against ROSS Intelligence. The opinion found the market for licensing headnotes as AI training text is "rapidly developing" and that ROSS took that opportunity without paying. ROSS says it will seek Supreme Court review.
A federal appeals court treated licensing training data as a real market worth protecting. Licensing is the legitimate route, and owners get paid on it.
SpaceXAI has held internal talks about buying customer and operational information from troubled or defunct startups as a cheaper data source, mirroring Google's Spirit bid. The talks are informal and may not lead to a deal.
Buyers will take operating history wherever it is cheapest, including from companies no longer around to object. A running business can set the terms and keep customer data out.
A Sept. 3 notice offered 25% over Google's winning bid, all cash, with micro1 doing de-identification in-house. Separately, a software vendor objected that records its platform generated for Spirit belong to the vendor under its SaaS contract.
Two signals. Buyers are competing for the same records, and your vendors may claim part of them. Check SaaS contracts for data-ownership clauses before anyone prices a package.
Box CEO, quoting the Spirit Airlines sale: AI is thirsty enough that data is valuable in almost any form. How a company manages and mines its organization's intelligence will decide competitiveness.
That is the category. Everyday operations are no longer exhaust. They are a line item, if someone can turn them into a package a buyer can actually purchase.
Google won the bankruptcy auction for the airline's internal data and custom software, and said it was not buying customer or credit-card information. Mercor bid $7.5 million and called Spirit its usual licensing process applied to a bankruptcy estate.
Even in a fire sale, the valuable object is how the company ran, and the buyer still drew a line at customer PII. We run privacy checks with that line in mind, and buyers apply their own.
Ooak connects to Gmail, Slack, Notion, Jira, Drive, and SharePoint, anonymizes it into a structurally identical "digital twin," and sells it to labs to train and evaluate agents. Its partner page puts closed deals between $20K and several hundred thousand dollars.
Real numbers, smaller than the headlines. Price tracks years of history, headcount, and how well your tools connect. Clean, linked records are what move it.
Demand from OpenAI and Anthropic has data firms like Mercor approaching startups that are shutting down or being acquired. Warmly heard from Mercor about its Slack, GitHub, and meeting records days after agreeing to sell to HubSpot.
Interest is the easy part. Rights, employee privacy, and what you can actually license are the hard part. That is the work.
Bloomberg reported early talks at double Mercor's last valuation. CEO Brendan Foody said annualized revenue crossed $2 billion, doubling in four months. Mercor also announced it is acquiring Deeptune, which builds simulated work environments for training agents.
The middlemen are scaling because labs need more real work to train on. Their supply is someone's operating history. Make sure yours is priced as yours.
Its Company Data Partnership Program pays companies for real-world business workflows used to train the next generation of models. CEO Ali Ansari is running open forums on how the program works.
Those numbers are a market signal, not a quote. Most companies will not qualify. The ones that do still need the corpus prepared.
micro1's CEO calls it a structural claim, not a short-term prediction: AI will automate the tedious parts of the economy, and self-learning without human data is impossible at the frontier.
Demand-side thesis, not a quote for your business. A trillion-dollar market still needs someone to take a messy operation and make a licensable corpus. That is the job.