The AI industry has a favourite scary statistic: “80–90% of AI startups fail.” It shows up in pitch decks, conference keynotes and LinkedIn threads — and it is almost always an estimate, borrowed from general startup folklore rather than measured from real products. A newly updated dataset takes the opposite approach: instead of guessing how many AI products die, it counts them, one by one.
The numbers come from ToolDirectory.AI, a directory that has catalogued 2,550 AI tools and maintains a hand-reviewed “AI Graveyard” of products that no longer exist. According to its AI tool mortality report, 217 of those 2,550 tools — 8.5% — had shut down or been acquired as of July 2026. Every entry is dated and verified by an editor, and the underlying dataset is published under a CC BY 4.0 licence, so the figures can be checked rather than taken on faith.
Acquisition, not collapse, is the leading cause of death
Of the 217 retired tools, 101 shut down outright, while 116 disappeared through acquisition. In other words, a tracked AI tool was slightly more likely to be bought than to fail. And an acquisition is not automatically fatal: 53.4% of acquired tools kept shipping under their own brand, while the rest were folded into the buyer\u2019s product line and sunset.
That distinction matters to anyone who depends on AI software. When a vendor announces it has been acquired, history says it is roughly a coin flip whether the product survives in a recognisable form.
2025 was the deadliest year on record
Retirements are accelerating: 21 tools wound down in 2023, 48 in 2024 and 89 in 2025 — the deadliest year on record. By mid-July 2026 the count for the year had already reached 44, putting the record within reach. The report also notes that these figures lag reality, because failures often surface months after the fact. Around 23% of recorded deaths were confirmed not by an announcement but by a lapsed domain — a reminder of how quietly this market buries its dead.
Consolidation is concentrated at the top
Eighty-five distinct companies have acquired at least one tracked AI tool, but the leaderboard is top-heavy. NVIDIA and OpenAI lead with seven acquisitions each, followed by Google, Meta and Salesforce with three apiece. Notably, Anysphere — the startup behind Cursor — has also acquired three tracked tools, a sign that the consolidation phase is no longer just incumbents swallowing startups; well-funded startups are now buying each other.
Where AI tools go to die
In absolute terms, the biggest graveyards are Developer Tools (32 retired tools), AI Infrastructure (26), and a three-way tie between AI Art & Image Creation, AI Content Writing and Marketing & SEO (17 each).
Failure rate tells a different story. Against the 8.5% baseline, Automotive AI tools have died at 16.1%, NSFW tools at 12.5% and coding assistants at 12%. Coding assistants are an instructive case: a crowded category squeezed from above as foundation-model vendors ship the same capabilities natively.
So why isn\u2019t the number 90%?
Definitions. The dataset tracks products that actually shipped and were public enough to be catalogued — tools that survived the idea and launch stages where most startup deaths happen. It also covers a young population: most tracked tools are under five years old, and mortality compounds with time. Read 8.5% as a measured floor, not a contradiction of the folklore. As the report itself puts it: “Commonly cited AI startup failure figures run 80–90% — but they\u2019re estimates. Ours is measured at the product level: 8.5%, hand-verified, and updated as the Graveyard grows.”
What buyers and builders should take from this
- Check the category, not just the vendor. A tool in a segment with double the baseline failure rate deserves an exit plan before you build a workflow on it.
- Treat acquisition as a risk event. More tools disappear through M&A than through outright failure, and nearly half of acquired products are eventually sunset. Ask what happens to your contract and your data on a change of control.
- Watch for silence. Almost a quarter of recorded deaths announced themselves only through a lapsed domain. If a vendor has gone quiet, that is data too.
The 8.5% figure will rise as more of 2026\u2019s failures surface — the dataset is updated continuously, and the full JSON is free to download and reuse with attribution. For a market producing both a genuine platform shift and an extraordinary volume of disposable products, a measured mortality rate is a better compass than folklore.
