Stability AI and the curse of the open-source first mover

A note on why this is an essay and not an entry. Most of what runs on this site is an obituary, filed only when the product itself stopped operating. Stability AI did not stop operating. The company that released Stable Diffusion is still here in mid-2026, under new ownership and new management, shipping commercial models. What follows is not a death notice. It is the harder thing to write, an account of a company that invented a category, held the lead for about a year, and then handed it to everyone else. The decline is the story. The company surviving it is the twist.

On August 22, 2022, Stability AI helped release Stable Diffusion, an open-source text-to-image model built with researchers at LMU Munich and Runway. The release broke the pattern the field had set. OpenAI's DALL-E 2 had launched that April behind an invite-only waitlist, then a content filter, then a credit meter. Midjourney ran inside a Discord server. Stable Diffusion asked for none of that. The weights were public under a permissive license, the code was on GitHub, and anyone with a mid-range graphics card could download the model, run it on their own machine, and generate images in seconds with no waitlist, no filter, and no per-image charge. Within weeks it was the substrate of a sprawling community of forks, fine-tunes, and interfaces. By most measures it became, and for a long while remained, the most-used image-generation model in the world.

The strategy that produced this was Emad Mostaque's, and for a season it looked like the winning one. Stability positioned itself as the open alternative to a field of closed labs, the company that would democratize generative AI while OpenAI and Google kept their models locked behind APIs. The narrative was good enough to raise serious money. In October 2022 Stability closed $101 million led by Coatue, Lightspeed, and O'Shaughnessy Ventures at a valuation reported around $1 billion, making it a unicorn roughly two months after the launch that made it famous. The press treated Stable Diffusion as the open-source counterweight to a closing industry. The product had won mindshare on a scale most startups never touch.

Then came the problem that mindshare does not solve. Stability had given away the asset. A company whose core product is a freely downloadable, freely modifiable, freely runnable model has, by construction, no obvious place to put a price. The thing that made Stable Diffusion spread - that you could take it, run it yourself, and never send Stability a dollar or a request - was the same thing that meant the spreading produced no revenue. Every fork that proved the model's reach was also a customer the company would never bill. Hosting it through DreamStudio, Stability's own paid interface, put Stability in competition with a long tail of cheaper and free third-party tools built on the exact weights it had published. The open release was a triumph of distribution and a hole in the income statement, and the two were the same event.

The costs, meanwhile, were not open-source. Training image models burns compute, and Stability was training a lot of them across image, language, audio, and 3D. By Bloomberg's reporting the company was spending in the range of $8 million a month by late 2023 while bringing in very little against it. It tried to raise new money at a $4 billion valuation and could not. Press accounts through late 2023 and early 2024 described a cash crunch, unpaid bills, and a company hunting for a buyer or a bridge. One generative-AI vendor publicly accused Stability of being slow to pay a five-figure invoice. The figures that surfaced later were stark: by the reporting around its restructuring, the company had generated less than $5 million in revenue in the first quarter of 2024 against losses above $30 million, and owed creditors close to $100 million. The most-used image model in the world was attached to a business that could not pay for it.

The leadership came apart alongside the finances. The three researchers most associated with the original Stable Diffusion work departed. Senior executives left in a steady stream through 2023. The investor relationships that a cash-hungry company most needs were, by multiple accounts, strained. On March 22, 2024, Mostaque resigned as CEO and from the board, framing the exit as a choice rather than an ouster and noting that he remained the majority shareholder. His stated reason was philosophical, that one does not beat centralized AI with more centralized AI, which read, against the backdrop of the company's balance sheet, as a graceful caption for a harder situation. The CTO and COO stepped in as interim co-CEOs. Within weeks the company laid off around 10 percent of its staff. A first-mover with a globally adopted model had, eighteen months after its defining launch, run low on cash and lost its founder.

While Stability was learning that an open model is hard to monetize, its rivals were busy monetizing the moment it had created. Midjourney, a small and famously un-funded team, kept its model closed, charged a subscription from early on, and built a paying base in the tens of millions without taking outside venture money, the clean inverse of Stability's give-it-away thesis. OpenAI folded image generation into ChatGPT, first through DALL-E 3 and then through the GPT-4o image model, so that hundreds of millions of people generated images inside a product they were already paying for, never thinking of it as a separate purchase. Adobe shipped Firefly in 2023, trained on licensed and owned imagery, and sold it to exactly the professional buyers who needed to know their outputs would not trigger a copyright fight, then wired it directly into Photoshop. Each of these competitors had the thing Stability lacked: a surface where the image model sat inside a business that already knew how to charge.

That contrast is the heart of it. Stability's bet was that owning the open standard would translate into owning the market, the way open infrastructure has sometimes minted companies that sell the support, the hosting, or the enterprise edition around a free core. The bet was not absurd. It was just unmatched to the cost structure of frontier model training and to a competitive field where the closed players could cross-subsidize image generation with revenue from elsewhere. An open database company can run lean while the community does the distribution. An open model company has to keep paying for the GPUs whether or not anyone pays it back. Stability proved that being first and being most-used are real achievements, and that neither one is a business model.

Here the story departs from the rest of this archive, because the company did not die. In June 2024 Stability raised about $80 million from a group that included Coatue, Lightspeed, Greycroft, Sean Parker, and Eric Schmidt, and installed Prem Akkaraju, formerly of the visual-effects house Weta Digital, as CEO, with Parker as executive chairman. James Cameron joined the board that September. The new management pointed the company away from the give-everything-away posture and toward commercial, safety-filtered model endpoints sold to media, gaming, and advertising. By December 2024 Akkaraju said the company had eliminated its debt and was growing quickly. In March 2025 the marketing giant WPP took a stake and signed a partnership to use Stability's models across its agencies. As of early 2026, Stability AI is an operating company with a real enterprise pitch, which is precisely why it is not an entry here.

Whether the second act works is an open question, and worth flagging as analysis rather than fact. The pivot to paid, filtered, enterprise endpoints is a reasonable answer to the monetization hole, but it is also a tacit admission that the original open thesis did not pay, and it puts Stability into direct competition with Adobe, OpenAI, and Google on their turf rather than its own. The brand that meant uncensored and free now means licensed and safe, which is a harder thing to be distinctive at. The company that once defined the open frontier is now one enterprise vendor among several, trading on a name that earned its fame doing the opposite of what the new business does. That can work. It is not the position a first-mover wants to be defending.

The lesson is not that open source is a mistake. Open releases built Stable Diffusion's reach, seeded a research community, and arguably did more to spread image generation than any closed product. The lesson is narrower and older. A first-mover advantage is only an advantage if it can be converted into something durable before the followers arrive, and a free, downloadable, infinitely copyable model is close to the hardest asset to convert there is. Stability won the launch, won the adoption, and won the place in the history of the field. What it did not win, in the window when winning it was possible, was a way to charge for any of it. The competitors who showed up later, with worse models and better businesses, took the part that pays. Being first to the future is not the same as owning it, and the company that gives the future away should not be surprised when someone else sells it back.