1
Here’s a number that should stop every conversation about AI startups: 92%. That’s the failure rate for AI startups, according to research tracking 200 companies across three continents. Not startups in general. AI startups specifically. The ones that are supposed to be the future.
For context, the failure rate for traditional tech startups is around 63%. Bad, but at least you’re flipping something closer to a coin. AI startups fail at a rate that makes the base case for any individual company essentially: this will not work.
And yet AI captured nearly half of all global venture capital in 2025. Roughly $110 billion flowed into AI companies in a single year. The money is moving faster than ever toward a category where the overwhelming majority of recipients will be dead within two years.
Something doesn’t add up. Let’s look at the numbers that explain why.
- * *
The most comprehensive autopsy of startup death I’ve found comes from CB Insights, which tracked 431 venture-backed companies that shut down since 2023. Combined, those 431 companies raised $17.5 billion in equity funding before dying. The median company raised $11 million. The average was $48 million, pulled upward by a long tail of spectacularly funded failures.
Subscribe. The failure rate of this newsletter is lower than the AI startups it covers.
The median time from last fundraise to death was 22 months. That’s less than two years between “we just closed our round” and “we’re shutting down.” The money didn’t buy survival. It bought a longer runway to the same cliff.
And the stated reasons for failure are almost comically predictable. Seventy percent cited “ran out of capital” as a cause — but that’s the final symptom, not the disease. The actual causes underneath: 43% had no product-market fit. 29% had bad timing. 19% had unsustainable unit economics. These are not exotic failure modes. These are the same reasons startups have always failed. AI didn’t change the fundamentals. It just made the fundraising easier and the reckoning more expensive.
- * *
The enterprise side of the story is even worse, and it comes from a source that’s hard to dismiss.
MIT’s NANDA initiative published a report in 2025 based on 150 executive interviews, a survey of 350 employees, and analysis of 300 public AI deployments. The headline finding: 95% of generative AI pilot programs fail to deliver any measurable return on investment. Ninety-five percent.
The remaining 5% that succeeded shared specific traits. They picked narrow, well-defined problems. They bought specialized tools from vendors rather than building internally. They empowered line managers to drive adoption instead of centralizing AI decisions in innovation labs.
The 95% that failed did the opposite. They started with the technology and went looking for a problem. They built custom solutions instead of buying proven ones. They created centralized AI teams detached from the workflows they were supposedly improving. The tools never learned from the organization’s actual processes, and the organizations never restructured to accommodate the tools. Both sides waited for the other to adapt. Neither did.
The MIT researchers identified this as a “learning gap” — not a technology gap. The models were capable enough. The integration was the failure point. Which means the problem isn’t that AI doesn’t work. It’s that most organizations don’t know how to make it work, and most AI startups are building products for an adoption curve that barely exists.
- * *
The pattern of what’s dying is as revealing as the rate.
SimpleClosure’s analysis of 2025 startup shutdowns found that AI was the largest single category of failures — though its share actually declined slightly from 17.7% to 15.9% as other boom-era sectors started hitting their own walls. The dominant subcategory within AI failures was what they called “wrappers and apps”: copilots, assistants, productivity tools, content generators, and vertical SaaS products whose core differentiation was an LLM front-end rather than proprietary data or infrastructure.
These companies found early usage but couldn’t convert it into durable economics. Margins compressed as API costs shifted. Switching costs were low. Differentiation eroded as the same capabilities became widely available. The story writes itself once you see the pattern: build a thin layer on top of someone else’s model, get initial traction from the novelty, watch retention collapse when the novelty wears off, run out of money.
Meanwhile, the infrastructure and developer-tool layer of AI — companies building the underlying systems rather than the applications on top — had a much lower failure rate. The picks-and-shovels metaphor is overused. It’s also correct.
- * *
Here’s the stat that haunts me: 42% of AI startups fail because they built a product without verifying that anyone wanted it.
This is not new information. “No market need” has been the number-one reason startups fail for over a decade. CB Insights has been publishing this data since 2014. Every founder has seen the chart. And yet 42% of AI startups — companies founded by smart people, backed by experienced investors, operating in the most hyped technology category in a generation — are making exactly the same mistake.
They build the technology first and look for customers second. They optimize for benchmark performance instead of user retention. They raise money on the strength of a demo and discover, months later, that nobody will pay for what they’ve built. The AI startup graveyard is not full of companies that built bad technology. It’s full of companies that built good technology for problems that didn’t exist, or for problems that people weren’t willing to pay to solve.
The gap between “this is technically impressive” and “this is something I need” has never been wider. And the funding environment is making it worse, because impressive technology is what gets funded, regardless of whether anyone needs it.
- * *
The burn rate data tells the rest of the story.
AI startups burn faster than any previous generation of technology companies. The cohort of AI startups launched in 2022 burned through $100 million collectively in three years — a cash-burn speed that doubled that of earlier generations. The median AI startup has a lifespan of roughly 18 months before shutting down or attempting a desperation pivot.
The math is simple and brutal. A seed-stage AI startup burning $75,000 per month — which is median — with $500,000 in the bank has about seven months of runway. The standard advice is to start fundraising with 12 to 18 months of runway remaining. For most AI startups, by the time the product is built and initial data is in, the runway is already short enough that fundraising happens from a position of weakness rather than strength.
And the fundraising environment has tightened. The median gap between funding rounds stretched to 696 days in mid-2025. Almost two years between checks. If your product hasn’t found product-market fit in that window — and 43% of startups haven’t, according to the autopsy data — there is no next round.
CB Insights found that two-thirds of companies in their shutdown dataset were shrinking headcount in the six months before death. Nearly a third died with ten or fewer employees. The shutdown isn’t sudden. It’s a slow bleed that everyone inside the company can feel but nobody wants to name, followed by a quiet announcement that nobody outside the company notices.
- * *
The optimistic reading of all this data is that it’s a healthy correction. The 2023-2024 cycle rewarded speed and novelty, producing a long tail of thin products that raised early capital simply by being first to market. The 2025-2026 cycle is filtering for what’s real: proprietary data advantages, actual unit economics, deep integration into workflows that create genuine switching costs.
And that reading is probably correct. This isn’t the end of AI as a category. It’s the first real selection event in a market that’s been running on hype and capital availability rather than fundamentals. The founders shutting down these companies are often the same ones who will start the next generation of AI-native startups — with clearer differentiation, tighter burn, and a more honest view of where value actually accrues.
But the less optimistic reading is also worth stating plainly: a staggering amount of money and talent is being allocated to building things that almost no one will use in two years. $17.5 billion in confirmed dead companies is just the visible layer. Below it sits a much larger base of zombie companies — technically still operating, technically still spending investor money, but with retention numbers and revenue trajectories that everyone involved knows are terminal.
The AI startup graveyard isn’t full of bad founders. It’s full of good founders who built the wrong thing, for the wrong reason, with the wrong assumptions about what the market actually wanted. The technology worked. The business didn’t.
- * *
If there’s a single lesson in all of this data, it’s one that predates AI by decades: the market doesn’t care how impressive your technology is. It cares whether you solve a problem that someone will pay for, repeatedly, for a long time.
92% of AI startups are learning this the hard way. The 8% that survive are the ones that figured it out before the runway ran out.
The numbers don’t lie. They just arrive too late for most of the people who needed them.
Sources
- AI4SP.org — “Why 90% of AI Startups Fail” — Research tracking 200 AI startups across three continents, 92% failure rate finding
- CB Insights — “Why Startups Fail: Top 9 Reasons” (2026) — Analysis of 431 VC-backed startups that shut down since 2023, $17.5B in combined funding, 22-month median time from last raise to death
- MIT NANDA Initiative — “The GenAI Divide: State of AI in Business 2025” — 150 interviews, 350-person survey, 300 deployment analysis, 95% pilot failure rate finding
- SimpleClosure — “State of Startup Shutdowns 2025” — AI as largest shutdown category, wrapper/app subcategory dominance, infrastructure vs. application failure rate divergence
- Digital Silk — “Top 35 Startup Failure Rate Statistics Worth Knowing in 2026” — 90% AI-native first-year failure rate, 18-month median lifespan, $110B total AI ecosystem investment in 2025
- Clarifai — “Why AI-Native Startups Fail: Data, Compute & Scaling Mistakes” (2026) — 90% first-year fold rate, 80% never past proof of concept, enterprise pilot failure confirmation
- Fortune / MIT — “95% of generative AI pilots at companies are failing” (2025) — Detailed breakdown of the learning gap vs. technology gap distinction
- * *