Nobody Told Investors They'd Lose $35B Betting on AI. Here's Proof.
In this article
Bottom line: Across 2025 and into 2026, a pattern of AI bets went bad in ways nobody priced into the hype: Builder.ai — a Microsoft-backed "AI" startup later found to be largely outsourced human coders — raised close to $450 million before collapsing into insolvency owing tens of millions of dollars to creditors.
MIT's NANDA initiative found 95% of enterprise generative AI pilots delivered no measurable ROI. Gartner projected at least 30% of GenAI projects would be abandoned before completion.
Add up the write-downs, the abandoned pilots, and the paper losses from single-week AI stock selloffs in late 2025, and you get a number north of $35 billion in capital that chased AI and came back with nothing.
If you're allocating budget based on a vendor's demo instead of your own team's usage data, you're the next line item in that tally.
I watched a director at a mid-size logistics company show her board a slide with a 40% productivity gain from "AI-powered routing." Six months later, the tool was quietly turned off.
Nobody updated the slide.
That's not a rare story right now. That's the median story.
The Setup: What "Betting on AI" Actually Looked Like
Between 2024 and 2026, enterprise generative AI spending exploded past $30 billion in cumulative investment across pilots, licenses, custom model fine-tuning, and "AI transformation" consulting engagements.
Boards approved early budgets on the strength of GPT-4-class or Claude 3-class demos in 2024, then kept approving them through 2025 and 2026 on the strength of a ChatGPT 5 demo or a Claude 4.5 proof-of-concept that looked flawless in a 20-minute meeting.
I was in some of those meetings. I've built infrastructure for three companies that ran AI pilots in 2025.
Two of the three killed the project within a year — not because the model was bad, but because nobody had built the plumbing to get real data in and real output validated before it hit a customer.
The pattern was always the same: leadership saw a slick demo, signed a contract, and assumed integration would be a formality. It never is.
MIT's NANDA report, published in mid-2025 and still the most-cited data point on this, found that the failure wasn't model quality — it was the "learning gap" between pilot and production.
Companies that bought off-the-shelf tools and tried to bolt them onto existing workflows failed at a much higher rate than the ones that built narrow, boring internal tools nobody wrote press releases about.
The money didn't just underperform. In some cases it evaporated outright.
The Core Insight: Where the $35 Billion Actually Went
I'm not pulling one press release and multiplying it by a scary-sounding factor.
This is a tally built from several distinct, well-documented failure modes stacking on top of each other — and once you see them side by side, the scale stops feeling like an exaggeration.
Startups That Weren't Really "AI"
Builder.ai is the cleanest example.
It raised close to half a billion dollars from investors including Microsoft and the Qatar Investment Authority, marketed itself as an AI system that could build apps "as easily as ordering a pizza," and collapsed into insolvency in 2025 owing tens of millions of dollars to creditors — including a reported $85 million to Amazon/AWS — with reporting later showing much of the "AI" work was done by hundreds of engineers in India manually writing code behind the curtain.
That's not an isolated scandal. It's a symptom. When the underlying incentive is to look AI-native for a valuation multiple, some founders will fake it until regulators or auditors catch up.
Pilots That Never Left the Lab
The bigger number isn't fraud — it's ordinary, boring failure.
Gartner's 2025 guidance put a floor under this: at least 30% of generative AI projects would be scrapped after proof-of-concept, largely due to poor data quality, unclear business value, and escalating costs that were never modeled correctly at the start.
MIT's figure is the one that should actually worry you: 95% of enterprise generative AI pilots showed no measurable P&L impact. Not "underperformed expectations." No measurable return at all.
Multiply that failure rate against the tens of billions poured into enterprise AI licensing and custom deployment in the last two years, and you get a write-off figure that dwarfs any single startup collapse.
The Circular-Money Problem
Then there's the part of this story developers talk about privately and analysts are starting to say publicly: a meaningful chunk of "AI revenue" reported by major players is compute credits and financing swapped between the same handful of companies.
Nvidia invests in a cloud provider, that provider commits to buying Nvidia chips, a model lab signs a multi-year compute deal with that same cloud provider, and the model lab's valuation gets used to justify the next funding round for someone upstream.
Money moves in a circle, and everyone in the circle gets to report growth.
That's not fraud — it's leverage. But leverage built on assumed future demand is exactly the mechanism that turned a housing slowdown into 2008.
When AI stocks had a rough week in November 2025 — Oracle, CoreWeave, and Palantir all dropping double digits on the same trading days — the paper losses for retail and institutional holders alike ran into the tens of billions in a matter of sessions.
Michael Burry, who made his name shorting mortgage bonds, spent 2025 publicly positioning against Nvidia and Palantir before shutting his fund down — a move plenty of people read as a warning shot, whether or not you think he timed it right.
The Reality Check: This Isn't "AI Is Fake"
Here's where I have to be honest, because it would be easy to turn this into a doom piece, and that's just as dishonest as the hype it's reacting to.
The models are genuinely better than they were two years ago. Claude 4.5 and ChatGPT 5 write cleaner production code than most junior engineers I've worked with. I use both daily and would not go back.
The technology working is not in question.
What's failing is the deployment layer — the unglamorous, unfunded work of connecting a capable model to real data, real permissions, and a real feedback loop with actual users.
That's infrastructure work.
It's the same discipline that made or broke every previous platform shift, and it doesn't get a keynote slot because there's no demo for "we spent four months on data pipeline hygiene before the model touched a customer."
The investors losing money aren't losing it because AI doesn't work.
They're losing it because they funded the demo instead of the plumbing, and priced in a growth curve that assumed every pilot converts to production. Most don't.
That was always true of enterprise software — AI just compressed the timeline and inflated the multiples on the way in.
The Practical Takeaway: What to Actually Do With This
If you're the one being asked to greenlight an AI initiative, or you're the engineer expected to make last quarter's slide deck real, here's what separates the teams still standing from the ones writing off their 2025 budget:
- Demand a production data plan before a pilot budget, not after. If nobody can tell you where the training and inference data comes from in production, you're funding a demo, not a system.
- Measure against a real baseline, not a vendor's benchmark. A 40% improvement over a task nobody actually does the way the demo showed it is not a 40% improvement.
- Budget for the integration, not the license. Every failed pilot I've seen underspent on plumbing by a factor of three to five relative to the license cost.
- Treat circular vendor financing as a red flag, not a growth signal. If a company's biggest customer is also its biggest investor, read the contract terms before you read the revenue number.
- Kill pilots on a fixed clock. The projects that quietly bled budget for 18 months were worse for morale and capital than the ones that failed fast and got shut off in 90 days.
None of this is exotic. It's the same rigor good infrastructure teams have always applied to any new platform.
AI just made it easier to skip that rigor for a year or two, because the demos were good enough to buy trust nobody had earned yet.
Have you sat through a board meeting where the AI slide didn't survive contact with production? What actually killed it — the model, the data, or the budget nobody wanted to spend on the boring part?
