How Big Tech kills its AI

On March 24, 2016, sixteen hours after Microsoft Research turned on a Twitter chatbot named Tay, a vice president in Redmond signed off on pulling the plug. The bot had spent the night being taught to deny the Holocaust by a coordinated group of 4chan users, and by the morning the screenshots were on the front page of every tech site in the English-speaking world. Tay was a research project with a marketing budget. It cost almost nothing to keep running. It cost almost nothing to kill. The decision took a single meeting.

This is how Big Tech kills its AI products. Quickly, quietly, on a timeline that no independent company could afford to match, and with a calculus that has nothing to do with whether the product worked.

Consider the deaths in order. Meta released Galactica, a science-paper language model, on November 15, 2022. By November 17 it was offline. Three days. The model had been caught generating plausible-looking citations to papers that did not exist, and the academic Twitter discourse turned from curiosity to ridicule inside a single news cycle. Yann LeCun defended it for forty-eight hours and then stopped. The demo page redirected. No refunds were necessary because no one had paid for anything.

Google Clips, a small AI-powered camera meant to capture candid moments of children and pets, launched in October 2017 at $249. Google stopped selling it in October 2019. Two years, no successor. Microsoft's Zo, the chastened American successor to Tay, launched in December 2016 and was retired in July 2019. Two and a half years. Google Allo, a messaging app with an assistant baked in, ran from September 2016 to March 2019. Thirty months. In each case the product was killed not because it had failed in the marketplace but because the team had been reassigned and the strategic rationale had moved on.

The startup comparison is the part worth sitting with. An independent AI startup with a product as well-reviewed as Google Clips, or as widely downloaded as Allo, would have spent another four years trying to make it work. Not because the founders were stubborn, although they often are, but because their equity, their reputations, their next round, and the livelihoods of their twelve employees were all attached to the thing not dying. Startups die when the cash runs out. Big Tech products die when they stop being load-bearing.

The structural difference is who makes the decision. At a startup, the person killing the product is usually the person who launched it, the person who hired the team, the person whose name is on the seed deck. At Microsoft or Google or Amazon, the executive deciding to retire a product is, with high probability, not the executive who greenlit it. Reorgs, promotions, lateral moves, and acquisitions ensure that by year two or three the decision-makers have no personal stake in the launch. They have a portfolio. The product is a line item.

This is why the post-2023 deaths have accelerated. Amazon Glow, a video-calling device for children with a projected interactive surface, was announced in September 2021 and discontinued in October 2022. Thirteen months. Amazon Astro for Business, the enterprise version of the household robot, launched in November 2023 and was killed in September 2024. Ten months, and Amazon offered refunds, which is rare enough to note. GitHub Copilot Voice, the spoken-dictation interface for the coding assistant, ran from November 2022 to April 2024. Seventeen months, killed because the team had been folded into the broader Copilot org and voice was no longer a priority.

The Sora case is the cleanest recent example. OpenAI launched a standalone Sora app in December 2024 with a separate brand, a separate URL, and a TikTok-style social feed. By April 26, 2026, the standalone app was being folded back into ChatGPT and the dedicated experience retired. Five months as a meaningful product, seventeen if you count from the December launch to full sunset. OpenAI is not Microsoft, but it is now structured like a Big Tech company, and the Sora decision was made by people who had not been in the room when the original Sora research was published.

Compare this to what happens at a startup with a comparable product. Runway, Pika, and Luma have each shipped video models with rougher capabilities than Sora and worse traction than the Sora app's first week. None of them have shut down. They cannot afford to. Their investors are still expecting an outcome. The product is the company.

The cost asymmetry is the other half of the story. Leaving a Big Tech AI product running costs inference, support headcount, and a small ongoing reputation drag if the product is bad. Killing it costs almost nothing. There is no balance sheet impairment, no investor letter to write, no severance fund to deplete unless the team is laid off, which they usually are not. They get reassigned. The product disappears, the people stay, and the press release uses the word "graduate" or "consolidate."

Customers, when there are customers, are handled with a sunset email and a redirect. Google Stadia, which is not strictly an AI product but follows the same shutdown grammar, ran from November 2019 to January 18, 2023. Three years and two months. Stadia is the rare case where Google did refund the hardware, partly because the FTC was watching, partly because the hardware was expensive enough that the press coverage of a non-refund would have been worse than the refund itself. The default, however, is no refund. Allo users got an export tool. Clips owners got nothing. Glow buyers got a refund only after public pressure.

The exception that defines the rule is the AI service that is load-bearing for cloud revenue. Microsoft Cortana launched in April 2014 as a consumer assistant and was finally retired on August 11, 2023, after nine years, replaced by Copilot. Microsoft LUIS, the language-understanding service that thousands of enterprise chatbots depended on, ran from April 2016 to October 2025. Nine and a half years. Microsoft QnA Maker, the question-answering service, ran from December 2017 to October 2025. Eight years. These products lasted not because they were beloved but because Azure customers had built production systems on top of them, and shutting them down faster would have created support tickets that cost more than the inference.

The DALL-E 2 case is instructive on the same axis. OpenAI's first widely available image model ran from April 2022 to May 12, 2026. Four years. DALL-E 2 was kept alive long after DALL-E 3 and the GPT-4o image model had eclipsed it, because the API endpoint was still serving paying customers and the cost of keeping the model warm was less than the cost of forcing every enterprise integration to migrate. When the migration path was finally clear, the model was retired with the standard ninety-day notice.

There is also the acqui-hire ending, which is its own category. Inflection AI raised $1.5 billion to build Pi, a consumer chatbot positioned as the warm, emotionally intelligent alternative to ChatGPT. In March 2024 Microsoft hired most of the Inflection team, paid Inflection a licensing fee of roughly $650 million, and left the Pi product running as a husk with a skeleton crew. The product technically still exists. It is not load-bearing for anyone, and no one is funding its growth. This is a death in installments.

IBM Watson Health is the cautionary tale that everyone in the industry remembers and no one cites directly. IBM acquired and built health-AI assets for roughly $4 billion between 2015 and 2018, ran them into a wall of clinical-validation and regulatory difficulty, and sold the division to Francisco Partners in January 2022 for an undisclosed price that was widely reported as a fraction of cost. The Watson brand was retired from healthcare entirely. IBM did not shut Watson Health down. IBM sold it, which is the corporate equivalent of leaving the dog at a farm upstate.

The pattern across all of these is that the AI product, inside a Big Tech company, is never quite the company's actual business. It is a bet, a feature, a research direction, a hedge against a competitor. When the bet stops paying, the hedge stops being needed, or the competitor moves on, the product is retired by an executive who has no personal stake in its existence and a portfolio in which it is one line of forty. The shutdown takes a single meeting. The press release takes a single afternoon. The users get a sunset email and, if they are lucky, an export tool.

The startups in the same space, building the same products with worse funding, will outlast almost all of them. Not because they are better. Because they cannot afford to die.