Two years ago the AI data deals that made the news were with publishers and platforms: forums, academic presses, photo libraries. In 2026 the buyers are calling heating and cooling contractors. The shift says a lot about what the models have already learned and what they still cannot do.
What did the first wave of deals buy?
Content that was already published, in bulk. A few of the figures on record:
| Seller | What was reported | Source |
|---|---|---|
| $203.0 million in total contract value for licensing its data, on terms of two to three years | Form S-1, reported by TechCrunch, Feb 2024 | |
| Wiley | AI licensing revenue of $23 million in fiscal 2024 and $40 million in fiscal 2025 | Wiley earnings release, Jun 2025 |
| Spirit Airlines | A $10 million bid from Google for internal records (staff emails, Teams messages, spreadsheets) against $7.5 million from Mercor, awaiting a judge’s approval when reported | Reuters, Aug 2026 |
| HVAC companies | Around $150,000 each from data labeling startups | The Information, 2026 |
The early deals were for text that already existed on the open web or in archives. The later ones are for something different: the internal records of businesses that do physical work.
Why does the shift point at operating businesses?
The public text is spoken for
Large models were trained on what could be read openly, then on what could be licensed from the companies that published it. Once a forum’s archive or a publisher’s back catalog is licensed, it does not get licensed again for the first time. Buyers who want something new have to go where the data has never been published.
Models need to see work, not writing about work
A manual describes how to diagnose a heat pump that will not defrost. A service history shows what thousands of technicians actually found when they arrived, what they tried first, what they replaced, and which fixes failed within a month. That second kind of record is what a system needs if it is going to help a dispatcher triage a call or a technician choose the next test. It exists almost nowhere except inside companies like yours.
Agents need workflows
The newest systems are built to carry out multi-step tasks: book the call, route it, quote the job, follow up. Training those needs records of how real businesses run step by step. micro1’s page lists operations records, SOPs and process documentation among what it takes. Nyne names service histories, jobs, dispatch and calendars outright.
Wound-down businesses proved the market
In October 2026, MIT Technology Review described AI labs paying for the leftovers of startups that had closed; one founder came away with roughly $25,000 for his codebase. The Spirit Airlines bids showed the same thing at the scale of a large company. If the records of a closed business have a price, the records of a working one obviously do.
Why HVAC, of all trades?
Several things line up. HVAC contractors have run their jobs through field service software for years, so the records are already digital and structured. The work is diagnostic, which is the kind of reasoning AI developers want to capture. Seasons repeat, so a long history shows the same problems across many years and many equipment types. And a company with 20 to 100 technicians is large enough to clear the size floors buyers publish: Nyne starts at 20 full-time U.S. W-2 employees, micro1 at 30, and Scale AI prefers 40.
Does this last?
Nobody can say. The buyers’ published ranges, and the referral fees some of them pay introducers, suggest demand is strong right now. Terms move quickly, too: micro1 rewrote its referral terms some time after June 2026, which is why we keep a dated changelog on our buyers page. An owner who is curious has every reason to look now and no reason to rush a signature.
If you want to know which buyers would look at a company your size, the homepage explains how to sell HVAC data through an introduction, with nothing shared until you choose.
The records were always a by-product of good work; for now, they are also an asset.