The first deaths of the AI boom

On May 20, 2025, the chief executive of Builder.ai told employees the company would file for insolvency. Builder.ai had raised more than $445 million from Microsoft, the Qatar Investment Authority, and SoftBank affiliates, reached a $1.5 billion valuation, and sold software it said an AI assistant named Natasha assembled to order. Natasha, reporters and former staff found, was substantially around 700 human engineers in India. It was, by most accounts, the first major bust of the generative-AI funding boom, the wave that began when ChatGPT arrived in late 2022 and money poured into anything that could plausibly attach itself to a model. The boom is still mostly alive. But it has started filing obituaries, and the early ones rhyme.

The archive now holds a cluster of deaths from 2024, 2025, and the first months of 2026 that, read together, sketch the failure modes the boom produces. Builder.ai and the news reader Artifact. The video model Sora and the AI music tool Splash. The developer search engine Phind and the LLM-tooling platform Humanloop. The assistants theGist, Astra, and Hiro. The recruiting startup Joonko, the design-to-code tool CodeParrot, the model-comparison playground Yupp, the primary-care company Forward. They sold different things. They died in a handful of recurring ways.

The first pattern is the wrapper undone by the platform. Phind built an answer engine for developers, raised about $10 million in December 2025, and shut down roughly six weeks later when ChatGPT, Perplexity, Google, and Claude had all added the web search that was its only real differentiator. A product whose moat is a capability the model labs can add in a release does not have a moat. This is the most common shape of the early deaths, because the boom funded a great many thin, useful layers over models the labs kept making thicker.

The second pattern is the model that costs more to run than the audience will pay. OpenAI's Sora, the most talked-about AI product of early 2024, was reportedly burning close to a million dollars a day in inference for an audience that peaked near a million users and then fell as the novelty wore off, and OpenAI shut the standalone app in 2026 to point its chips at work that paid. Splash, an AI music generator backed by a $27 million round, decided after eight months that its Roblox game was the better business and discontinued the music tool. Generative audio and video are expensive in a way text is not, every second of output a small render farm, and the boom funded plenty of products on the bet that the cost curve would bend faster than the audience thinned. It did not always.

The third pattern is raised big, acqui-hired fast. Humanloop took a $36.2 million Series A in May 2025 and was gone three months later, its team absorbed by Anthropic and its platform switched off. Hiro, an AI personal-finance app, was live for about five months before OpenAI hired its team in 2026 and the app went dark. Adept, from authors of the transformer paper, raised about $415 million and was effectively wound down into Amazon. In this boom the buyers are the model labs and the platforms, and the deal that pays the founders and turns off the product has become the modal soft landing for a startup that built real technology and no defensible business.

The fourth pattern is the appearance of AI standing in for the substance. Builder.ai's Natasha was mostly people. Joonko, a recruiting startup that claimed its AI sourced diverse candidates, turned out to be built on fabricated customers and revenue; its founder was charged by the SEC and the Department of Justice in 2024 with defrauding investors of more than $27 million. The boom rewarded the look of artificial intelligence, sometimes more than the thing itself, and a market that pays a premium for the look will reliably attract people selling the look. These are not failures of technology. They are failures the technology's hype made possible.

The fifth pattern is the frontier moving and stranding the products built for the last position. theGist and Astra, two assistive agents that each did one task well, both blamed the same thing on the way out: the industry's loud turn toward fully autonomous agents drained demand for the helpful, in-between tools they had built. The boom does not only kill products by competition or economics. It kills them by fashion, by promising a bigger thing that makes the shipping smaller thing look obsolete before it has found its market.

Not every early death was a disaster, and it is worth recording the graceful ones, because they are rare. theGist returned its remaining capital to investors rather than spend it down. Yupp, a model-comparison playground that signed 1.3 million users on a $33 million seed, concluded it had not found product-market fit and wound down in 2026. These are the endings of founders who decided the bet had not worked and said so, in a year crowded with messier exits. The boom has produced fraud and flameouts. It has also produced a few people who gave the money back.

The most distinctive feature of the boom's first deaths is their speed. Phind died six weeks after a raise. Hiro lasted about five months from launch to acqui-hire. Astra folded four months after taking angel money. The combination that produces this velocity is specific to the moment: products are cheap to start, because the model is an API call away, and therefore cheap to kill, and the same API dependency means a competitor's new feature or a lab's new release can end you overnight. Earlier waves in this archive died over three to twelve years. The generative-AI wave is capable of doing it in weeks.

The honest framing is the one the publication keeps: most AI products are not dead, the boom is mostly still running, and the obituaries here are a minority report. But the first deaths are the most useful data the field has, because they show which bets the boom cannot fund: the thin wrapper, the unaffordable model, the company that is mostly people, the helpful tool the next frontier orphans. Builder.ai will not be the last $445 million entry. It is the first, which makes it the one worth reading closely, the way the first obituary in any wave is the one that tells you what the rest will say.