When your AI tool gets acquired, you get about a week to export.

Acquisitions are 150 of the 285 entries in the biggest AI tool graveyard. Clockwise and Hiro users each got seven days. Read where the risk sits.

The Editors · 8 min read · Aug 30, 2026

A room filled with lots of boxes and plants

The likeliest way an AI tool you earn money on disappears is that someone buys the team behind it. In the largest hand-reviewed registry of dead AI products, 150 of the 285 entries ended in an acquisition and 135 in an outright shutdown. And when the buyer is an AI lab or a large software company shopping for engineers, the product tends to go dark about a week after you hear about it.

Two cases from this year set that clock. Salesforce said on 20 March 2026 that Clockwise's team was joining Agentforce, and Clockwise closed its services on 27 March. OpenAI said on 13 April 2026 that it had bought Hiro, and Hiro stopped working on 20 April. Seven days each.

So the useful number for anyone running income through a third-party AI tool is the notice window, and seven days is a working answer for it. That is shorter than most people's export routine.

The registry reads like a sales record

ToolDirectory.AI keeps a hand-reviewed list of AI products that shut down or got bought. As of its 30 August 2026 update it holds 285 entries, split three ways: 135 shut down outright, 66 acquired and sunset, 84 acquired and still operating under their own brand.

How 285 dead or acquired AI tools ended

Source: ToolDirectory.AI AI Graveyard, 30 August 2026

Read that split from the operator's side and it reorders the risk. A shutdown is loud and final. An acquisition forks into two different futures: 66 products that got folded into the buyer and stopped existing, and 84 that still run while their pricing and roadmap answer to somebody else.

The list is also growing faster than tools are dying. The same registry held 193 entries on 8 July and 219 on 2 August before reaching 285 on 28 August. Ninety-two entries in seven weeks measures how fast researchers are finding old deaths as much as how fast new ones happen. Every recent-year count in the registry is a floor.

Seven days, twice

Clockwise was ten years old. Salesforce took the team, and CEO Matt Martin told customers the Clockwise services were going away. They went away on 27 March. Data was deleted after that date, prepaid subscriptions were refunded, and users were pointed at a competitor that offered to match their price.

Hiro was five months old. Founder Ethan Bloch said his roughly ten-person team was going to OpenAI with him. The app stopped working on 20 April, and users had until 13 May to export before the data was deleted from the servers.

Different ages, different buyers, different products. Same seven days.

The export window is the part that matters most and gets published last. Hiro gave 30 days after shutdown to pull your records. Clockwise deleted them. Neither policy was knowable in advance, which is the whole problem. The terms of your exit get written by someone who bought the company after you started depending on it.

The quieter version: bought and still shipping

Those 84 entries in the "acquired and still operating" bucket are the ones that reach your margin without ever showing an error page. Nothing 404s. The logo stays. Then the price moves, or the free tier goes, or the feature your workflow sits on gets folded into the buyer's own product.

Stripe bought OpenRouter in August in a deal reported above $7 billion, and OpenRouter still routes requests to the same models. We covered what it charges to load credits the week that closed. Nothing there broke. Whether the take rate stays where it is now belongs to Stripe.

The acquihire structure sharpens this. The buyer takes the people and leaves the product standing. An analysis published in April and updated on 20 August counted more than $20 billion spent by Google, Microsoft, Amazon and Meta between March 2024 and January 2026 on deals that hired away founding teams without formally acquiring the companies. Microsoft's $650 million Inflection deal kept 12 employees. Amazon's Adept deal was around $25 million for four. A product with no engineers left is running, not maintained.

Sometimes nobody tells you

NextSlide's founder joined OpenAI in March 2026 and made it public on 8 August. Five months. Anyone building decks in it spent that stretch on a product whose team had already left, and whatever the export window turned out to be, it opened and closed before the deal was news.

That is the extreme case. Silence is closer to the norm at the small end: in the 208-tool dataset behind the registry's June report, 51% of the outright deaths announced nothing at all and simply let the domain lapse.

First-party products get the same treatment. OpenAI wound down ChatGPT Atlas, its own browser, in early August; it stopped working on 9 August 2026, less than a year after launch, and users were told to export bookmarks as an HTML file and import them elsewhere because nothing carried over automatically.

Where the deaths cluster

The June report sorted its 208-tool dataset by category. Developer tools led with 30 and AI infrastructure had 25. Then come three categories that describe where a lot of AI side income actually lives: art and image creation at 16, marketing and SEO at 16, content writing at 15.

Those three are the thin end, and here the read goes past the count. They are cheap to build, mostly thin layers over someone else's model, and each time a lab ships the same capability natively a whole category loses its reason to charge. That is analysis, not data. What the data says is that the categories closest to content and marketing work are already well represented among the dead, and the survival rate for income built this way has never matched the pitch.

What these numbers cannot tell you

Three limits, and they are load-bearing.

The registries are not censuses. ToolDirectory hand-reviews entries, which is why its 285 is precise and small. Dang.ai runs a different method across its own directory and reports 1,757 dead listings out of 6,783 published, or 25.9%. Those two figures measure different populations by different rules. Use them for shape, not for rates.

The seven-day figure rests on two cases. Clockwise and Hiro are both dated, verifiable, and they happen to match. Two matching cases make a pattern worth planning against, not a distribution.

And acquisition is not death. Eighty-four of the 150 acquired products are still shipping, and some of them will be better resourced than they were alone. The thing to plan for is a change of terms arriving on the buyer's schedule.

What to watch

The registry stands at 84 entries for 2026 through 30 August against 106 for all of 2025, and that 2025 figure grew by 20 after the June report was compiled. Watch whether 2026 clears 2025 once the backfill catches up. On current behaviour it will.

For anyone earning through one of these tools, the operational read is short. Export on a schedule rather than on a plan, because the plan assumes notice you may not get. Keep the account of record, the client list and the files somewhere you own. And when a tool you depend on gets bought, read the announcement for what happens to the team, because the engineers leaving is what comes before the product going quiet.