The easy market that wasn't

The conventional wisdom held that developer tools would be the easy AI market. Developers are a known buyer. The workflow already had a wallet. You did not have to convince a coder that software was worth paying for, or teach a hospital procurement committee what your product did, or wait on an FDA clearance. You shipped a useful tool, charged for it, and watched the engineers expense it. Kite tried exactly this, building an AI code-completion tool years before GitHub Copilot existed, and could never figure out how to charge for it. It shut down in 2022, having proved the use case and captured none of the value. The developer market was not the easy one. It was the one most exposed to the platforms, and the archive shows it dying in two distinct ways.

The first death is obsolescence by the platform you sit on. Kite's autocomplete became a free feature of Copilot. Phind, an answer engine built for developers that resolved coding questions with cited sources, raised about $10 million in December 2025 and shut down roughly six weeks later, in January 2026, once ChatGPT, Perplexity, Google, and Claude had all added capable web search and left a dedicated coding search engine with nothing to defend. CodeParrot turned Figma designs into front-end code until Copilot and Replit folded the same trick into tools developers already had. Snazzy AI, one of the first GPT-3 copywriting tools, was bought by Unbounce, run for years as Smart Copy, and finally switched off in January 2026 as the same capability became table stakes inside every builder. The pattern is constant: a useful thin layer over a model, and a platform one cycle behind that can absorb the layer into its base price.

Microsoft ran the same play on its own tools. LUIS, the Language Understanding service, and QnA Maker, its question-answering counterpart, were the workhorses thousands of enterprise chatbots were built on through the late 2010s. Both were retired in October 2025, made obsolete by the large language models Microsoft itself was now selling, which did in one general-purpose call what LUIS had needed hand-labeled intent classifiers and training utterances to approximate. When the platform owner ships the thing your tool used to do, the deprecation notice is only a matter of scheduling.

GitHub Copilot Voice, the spoken-dictation interface for the coding assistant, died for the plainest reason of all. It ran from late 2022 to April 2024 and was shut not because it failed in the market but because the team was folded into the broader Copilot organization and voice was not a priority anyone there was willing to defend. A feature inside a platform survives exactly as long as someone with a budget believes in it.

The second death is quieter and, for the infrastructure layer, more total: absorption by the company whose stack you optimize. OctoAI, founded by the University of Washington researchers behind the Apache TVM compiler, built a platform for running open models cheaply across any hardware and raised about $132 million at a roughly $900 million valuation. CentML, out of the University of Toronto, built an optimization layer between models and chips and raised about $31 million. Both were bought by Nvidia, in 2024 and 2025 respectively, and both had their services shut down within weeks of the deal. The layer that promised to make any hardware run AI efficiently turned out to be a feature of the hardware, and the hardware company agreed by buying it and turning it off.

The labs and the application platforms did the same to the tooling above them. Humanloop built a development platform for LLM applications, prompt management, evaluation, and observability, used by Duolingo and Gusto and Vanta, and raised a $36.2 million Series A in May 2025. Three months later Anthropic hired its team without buying its assets, and the platform shut down in September with customer data deleted. Dashworks built an enterprise assistant that searched across a company's apps, raised about $9.5 million, and was absorbed into HubSpot's Breeze suite in 2025, the standalone product wound down at the announcement. The tools that helped companies build on top of the models were bought by the companies that make the models, and by the platforms that wanted the feature.

The structural problem is specific to selling to developers and to the platforms developers live on. In most markets, your customer cannot become your competitor overnight. In dev tools, the platform you build on can build what you built, and frequently will, because the feature you sell is often the feature that makes the platform stickier. The distance between a useful wrapper and a native platform capability is one product cycle, and the platform sets the clock. Worse, your best customers are the same engineers who could rebuild your tool in a sprint once the base model is good enough, and the base model gets good enough on a schedule you do not control.

Builder.ai is the cautionary extreme of the developer-tool dream, the no-code app builder that promised software as easy as ordering pizza and turned out to be running roughly 700 human engineers behind an AI assistant named Natasha. It raised more than $445 million and reached a $1.5 billion valuation before collapsing into insolvency in 2025. It is filed under dev tools because that is what it sold, and it is a warning about the whole category's central temptation: the gap between what an AI dev tool promises and what the models can actually deliver is wide enough to hide a small army in, and the market eventually finds the army.

The picks-and-shovels story, sell tools to the gold miners rather than dig yourself, has a flaw the AI version makes vivid. It works right up until the owner of the mine starts making shovels, and in this boom the owners of the mine, the model labs and the cloud and developer platforms, make excellent shovels and give many of them away to keep you mining on their land. The merchant tooling layer gets squeezed from above by the platforms and from below by open-source releases, and the squeeze is fastest precisely where the tool is most useful, because useful is what the platform most wants to own.

None of this means AI developer tools cannot be a business. It means the easy framing was wrong. The survivors are the ones selling something the platform cannot trivially copy or does not want to: deep workflow integration, proprietary data, a surface the model labs have no interest in owning. The tools in this archive mostly sold the layer the platform wanted most. They proved the use case, charged too little or too late, and were obsolesced or absorbed by the company one level down the stack. Selling to developers was never the easy market. It was the market where your supplier, your customer, and your competitor kept turning out to be the same company.