The AI video generation market did not recover because Sora finally worked. It recovered despite Sora largely failing to, and the clearest evidence for that is commercial rather than technical, according to The Decoder. Video startup Higgsfield, founded by former Snap executive Alex Mashrabov, says its annualized revenue climbed from $20 million to $700 million over twelve months, with its valuation rising from $1.3 billion to $5.4 billion in eight months, per the Financial Times as cited by The Decoder.
That is one company’s own disclosure, not an audited measure of the category’s size, and it says more about Higgsfield’s execution than about video AI broadly. Still, the composition of its revenue points the same direction. In January, business customers supplied under a quarter of Higgsfield’s revenue. They now supply most of it. Brands such as Dollar Shave Club moved from producing a single marketing campaign a year to several videos a day, per Mashrabov.
Other numbers in the report point the same way without adding up to one clean industry total. Netflix says AI will have touched 300 of its 1,000 titles by 2026. Promise, an AI production studio backed by Google, Disney, and Silicon Valley investors, claims hybrid productions cost 20 to 50 percent less than conventional filmmaking, per The Guardian as cited by The Decoder. Obsidian, an AI studio working with director Ron Howard on an animated documentary, estimates comparable savings of 30 to 40 percent. Those are studio estimates about cost, not third-party measurements of a market, and the article treats them that way rather than as proof of scale.
Sora’s stumble was not one event but three. OpenAI’s original demos in early 2024 were technically ahead of any competitor and rattled Hollywood, but the version that eventually shipped to the public later that year fell short of that early bar. A standalone Sora 2 app followed in the fall of 2025, generated attention, and then failed to hold onto users. OpenAI shut it down within months. A reported billion-dollar Disney partnership tied to the app collapsed alongside it. What actually disappointed, in other words, was retention after launch, not the ability to generate interest in the first place.
That distinction matters for anyone reading Sora’s underperformance as a verdict on the category. A first launch missing its mark usually means the price, the distribution, or the product itself was wrong, not that demand for the underlying capability was absent. Rivals, including Chinese developers building models such as Seedance 2.5, built tools around production workflows instead of demo spectacle while OpenAI hesitated between iterations. Studios and marketers adopted what worked for ordinary, unglamorous reasons: lower cost per finished shot and faster turnaround, not novelty.
The report also documents a second business layer built on volume rather than prestige. Inception Point AI, a startup producing synthetic podcast hosts, runs more than 100 AI personas across over 5,000 active shows. CEO Jeanine Wright says an episode needs only 20 listeners to be profitable, since each show costs next to nothing to produce. Parallel businesses have emerged in AI fashion modeling, where agency Seraphinne Vallora builds AI models for brands like Guess, and in AI-generated music, where Spotify now plans to label synthetic artist profiles.
None of this means Sora is finished, but it does mean the category’s recovery happened independent of whether Sora itself worked. Operators evaluating AI video vendors this quarter should weigh production economics (cost per shot, turnaround time, customer retention) over raw model demos, since that is what turned Higgsfield’s early adopters into repeat, high-frequency customers rather than one-time trial users.
The Decoder (Maximilian Schreiner) reported this analysis of the AI video generation market on August 17, 2026.
