When OpenAI quietly signaled last week that it was “saying goodbye” to its video-generation app Sora, the brevity of the announcement masked a deeper reality: one of the most hyped products in generative AI had become too expensive to sustain.
Behind the decision lies a hard constraint now shaping the industry — compute. Running Sora, by multiple estimates, was costing the ChatGPT maker close to $1 million a day, a figure that appears to have tipped the internal balance against the product as the company sharpens its focus ahead of a potential public listing.
Unlike text-based systems such as ChatGPT, video generation operates at the extreme end of compute intensity. Each prompt requires the model to generate sequences of frames, maintain temporal consistency, simulate motion physics, and, in many cases, synchronize audio. The result is an exponential increase in GPU usage per query.
Register for Tekedia Mini-MBA edition 20 (June 8 – Sept 5, 2026).
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab.
At scale, that becomes unsustainable.
Sora’s early trajectory suggested a breakout success. Within weeks of its September debut, it climbed to the top of app store rankings, amassing millions of downloads and dominating online discourse. Yet the initial surge proved difficult to maintain. By early 2026, download momentum had slowed sharply, even as the cost of serving each user session remained high.
That mismatch, declining marginal growth against persistently high operating costs, appears to have been decisive.
Internally, the calculation is increasingly about compute allocation efficiency. Every GPU cycle spent rendering video is a cycle not used to train or serve higher-margin products. For a company competing at the frontier of AI, where training next-generation models can cost billions of dollars, such trade-offs are no longer theoretical. They have become existential.
OpenAI’s subsequent exclamation points in that direction. The company said it would redirect the Sora team toward world simulation research, a domain tied to robotics and embodied AI. That shift underlines reprioritization: moving away from consumer-facing novelty applications toward foundational systems with clearer long-term commercial pathways.
The decision comes at a crucial time in the ChatGPT maker’s history. As OpenAI edges closer to a possible IPO, investors are likely to scrutinize not just growth metrics but cost discipline. High-burn, low-monetization products such as Sora complicate that narrative. In contrast, enterprise-facing tools, coding assistants, workflow agents, and API services offer more predictable revenue streams and better alignment with compute spending.
The company’s evolving product decisions reinforce that shift. Features such as instant checkout and more experimental consumer-facing modes have been scaled back, while development has intensified around integrated “superapp”-style functionality designed for workplace productivity.
In that sense, Sora’s shutdown is less an isolated move than part of a broader restructuring of priorities.
There is also a competitive layer. Rivals are increasingly focusing on enterprise deployment, where reliability, latency, and cost-per-query matter more than viral appeal. In that environment, a product that consumes vast compute resources without a commensurate revenue model becomes difficult to justify.
Sora faced pressure on another front as well: governance. Video generation tools sit at the middle of ongoing concerns around deepfakes, intellectual property, and misinformation. Efforts to impose safeguards tend to increase operational complexity and, in some cases, reduce user engagement — further weakening the business case.
The convergence of these factors left Sora exposed.
The underlying technology, however, is unlikely to disappear. By shifting resources into world simulation, OpenAI is effectively repositioning video generation as an enabling layer for robotics and physical AI systems rather than a standalone consumer product. Models capable of simulating environments, motion, and object interaction are critical to training machines that can operate in the real world.
That reframing suggests Sora’s demise is not about technical failure, but about economic prioritization.
The decision highlights a defining feature of the current AI cycle: compute has become the industry’s scarcest resource. Companies are now forced to make explicit choices about where to deploy it, often at the expense of high-profile products.
In Sora’s case, the conclusion appears to be that the cost of running the platform, at scale, and without a sufficiently strong revenue engine, became too high to justify, even for one of the best-funded players in the field.