One tracker counts 364 dead AI startups. If you're a solo builder shipping a wrapper, here's the actual lesson in the wreckage

One startup-failure tracker, IdeaProof, currently counts more than 364 AI companies that have shut down since 2023, with a combined $226 billion in capital raised before the doors closed. The same source, echoed by a cluster of blog posts, puts a number on the future too: roughly 80 percent of AI startups are projected to fail by the end of 2026. I spent an afternoon trying to trace that 80 percent figure back to an actual CB Insights or Gartner report and could not find one that says it. That does not make the underlying pattern fake. It means the confident-sounding stat is doing more work than the evidence behind it can carry, and if you're a solo builder shipping a thin AI wrapper right now, the real lesson has nothing to do with whether the true number is 80 percent or 40 percent.

what these trackers actually show when you look closely

IdeaProof's own page is not internally consistent. The headline says "364+ documented," a few paragraphs down it references a "corpus" of 562 documented failures analyzed, and elsewhere it mentions a database of 1,000 "verified true-failure events." The example cards illustrating the AI slice include WeWork, Silicon Valley Bank, BitMEX, and Terraform Labs, none of which are AI startups by any reasonable definition. That's a tell: the dataset looks scraped and loosely tagged rather than painstakingly verified, and the 80 percent projection is footnoted as an "industry-wide estimate from external sources" with a different denominator than the corpus itself, not something the tracker measured.

None of that means the pattern is invented. Builder.ai is the cleanest verified case I found, documented by TechCrunch, The Register, and Rest of World. It filed for insolvency in May 2025 after its "Natasha" AI assistant turned out to lean heavily on outsourced human engineers, a story The Wall Street Journal first raised back in 2019. Revenue forecasts got slashed 25 percent by October 2024, debt passed $100 million, an emergency $75 million raise fell through in February 2025, and a creditor seized $37 million from its accounts that May, which ended the company. It's instructive precisely because the failure mode wasn't "no product-market fit." The AI claim was the whole pitch, and the gap between the story and the reality couldn't survive due diligence once the money got tight.

the pattern behind most of the failures: wrapper, no moat, burn

Set the exact numbers aside and the shape of the typical failure is easy to describe: a thin layer of prompts and UI wrapped around a single foundation-model API, no proprietary data or workflow lock-in underneath it, and a runway spent mostly on inference bills rather than product or distribution. Inference cost per million tokens reportedly dropped around 80 percent between 2023 and 2025, which is great for users and brutal for anyone whose entire business model was the margin between what OpenAI charges and what they charge their customers. Add to that the concentration of AI funding, with a large share reportedly going to a handful of foundation labs, and you get thousands of application-layer companies competing for what's left while those same labs quietly ship native versions of the exact feature the startups were selling.

why this is a warning for solo builders, not just VC-backed startups

Here's the part that made me uncomfortable writing this. The exact "thin wrapper, no moat" product is the one a solo builder can ship fastest. One API call, a weekend of UI work, no data pipeline, no enterprise sales motion. That's what makes it tempting, and it's exactly what makes it risky. A solo operator doesn't have a burn rate in the VC sense: no runway clock, no board asking about the next raise. But you do have a burn rate. It's your evenings, your weekends, the eighteen months you spend iterating on a ChatGPT wrapper instead of building something with actual defensibility underneath it.

Funded by your own time instead of venture cash just means the failure is slower and quieter. Nobody writes a teardown post about the solo SaaS that never crossed $500 in monthly revenue before you quietly let the domain lapse. It still failed. It just failed privately, on your own schedule, with no term sheet to blame and no press release to explain it.

what's actually surviving as a solo or small-team product

The narrower pattern worth paying attention to is what hasn't been failing: AI products built around proprietary data, a specific workflow embedded deep in a customer's process, or a direct customer relationship a model provider has no reason to replicate. A scheduling tool wired into a physical therapy clinic's billing and patient records survives a ChatGPT feature update because OpenAI has neither the access nor the interest in that specific workflow. A tool that has spent two years accumulating a dataset nobody else has, even a small one, has something a general-purpose chatbot can't copy overnight. Contrast that with a generic "AI writing assistant" or "chat with your PDF" tool, where the entire value proposition could plausibly become a checkbox in someone else's settings menu.

The honest take

Here's the self-test I'd actually run before writing another line of code on an AI product: if OpenAI, Anthropic, or Google added your core feature as a checkbox in their own product tomorrow, would yours still have a reason to exist? If the honest answer is no, that's the actual roadmap problem, and it's worth solving before you spend another month polishing the UI.

I want to be upfront about where this argument could be wrong. Some genuinely thin wrappers have survived for years purely on distribution and marketing execution, an app that owns a category in the app stores, or a creator with an audience who converts attention into subscriptions regardless of how replicable the underlying tech is. Moat isn't the only variable, and "wrapper" alone isn't a death sentence if you're better than everyone else at getting in front of customers. I'd also flag, again, that the two headline numbers anchoring pieces like this one, the 364 figure and the 80 percent projection, come from a tracker whose own math doesn't fully add up and from a stat that multiple blogs repeat without a traceable primary source. Treat them as a rough signal of a real trend, not a precise measurement. The self-test above doesn't depend on the exact count being right, and that's the point: whether it's 364 dead companies or 3,000, the question of what happens to your product the day a foundation model absorbs your feature is worth answering now, while it still costs you an afternoon of thinking instead of a year of building.

Author

Lukas

@lukcombinator

Sources

• 364+ AI Startups That Failed (2023-2026): Data, Patterns & Lessons | IdeaProof
• Once worth over $1B, Microsoft-backed Builder.ai is running out of money
• Builder.ai coded itself into a corner, now it's bankrupt
• What was Builder.ai and why did it shut down?
• Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027