OpenAI’s Sora—an invite-only app that serves an endless feed of exclusively AI-generated, up-to-10-second videos—has rocketed to the top of the App Store. It looks like a neat trick, but under the hood it’s a stress test for compute infrastructure, copyright law, and the social systems that will have to police machine-made reality.

Sora matters because it moves deep-generation models from occasional novelty into a platform-scale product that can attract millions of minutes of human attention. Unlike text or still images, synthetic video increases compute and energy costs by an order of magnitude and broadens legal exposure, since moving images and voices map cleanly onto familiar intellectual-property and likeness rules.

How Sora generates scrollable fantasy — and why that’s expensive

How Sora generates scrollable fantasy — and why that’s expensive

Sora stitches together short, TikTok-like clips using generative video models that combine diffusion, image-to-video conditioning, and neural audio synthesis. Each 5–10 second clip typically requires multiple model passes: a text-to-image diffusion backbone, a temporal-consistency step to animate frames, and a separate vocoder to synthesize speech. That pipeline is far heavier than a single ChatGPT text completion.

OpenAI has acknowledged the cost problem. In a blog post on October 3, 2025, Sam Altman wrote that “we are going to have to somehow make money for video generation.” The company has already joined major investments in new data-center capacity and power projects—moves that imply capital budgets in the hundreds of millions to billions of dollars if video generation scales.

Who foots the bill: economics, monetization, and the industrial loop

The multiplicative effect is simple arithmetic: a text response often costs a few joules at inference; an image costs more; a video multiplies by the number of frames (say, 24–30 fps). Researchers who study model carbon footprints estimate that generating a single second of high-fidelity AI video can consume tens to hundreds of times more energy than generating a single page of text. That gulf will drive product decisions—limits, throttles, or monetization—before long.

Who foots the bill: economics, monetization, and the industrial loop

OpenAI is not profitable. The strategic choice to offer free, unlimited Sora generation—while the product is invite-only—resembles a loss-leading growth push. But it exposes the company to two pressures: runaway infrastructure spend and market optics. Reporting in MIT Technology Review shows industry concerns that circular chip deals and opaque financing could distort cost signals in the AI ecosystem.

Copyright, cameos, and the coming legal fights

There are three viable monetization levers: subscription tiers, ad insertion, and API charges to third-party creators. Each has trade-offs. Ads can amplify behavioral targeting and require even more inference to personalize content—raising compute and privacy demands. Subscription fees limit scale but stabilize revenue. OpenAI’s choice will determine whether Sora ends up subsidized by venture capital, advertisers, or by the pockets of power-hungry users.

Beyond OpenAI, the supply chain matters. Large video workloads concentrate demand for GPUs and specialized accelerators. If Sora drives sustained demand for A100s or H100s, it will reshape short-term hardware markets—tightening availability, inflating prices, and encouraging vertical deals that critics argue concentrate power among a few cloud and chip vendors.

Copyright, cameos, and the coming legal fights

Bias, personalization, and the ethics of an infinite synthetic feed

Sora’s feed is a legal minefield. OpenAI allows users to create “cameos” that mimic a real person’s appearance and voice; its initial settings let creators insert other people's cameos depending on permissions. The Wall Street Journal reported that OpenAI sent letters asking rights holders to opt out rather than opt in—a reversal of customary practice—and Sam Altman conceded there will be “edge cases of generations that get through that shouldn’t.”

Those edge cases are the trigger for litigation. Copyright owners, music publishers, and estates of deceased celebrities have clear causes of action if copyrighted melodies or recognizably trademarked characters appear without license. Likeness and deepfake harms raise separate tort questions: states and countries are updating statutes, but litigation, laden with jurisdictional fights, will determine practical limits faster than statutes.

OpenAI has started to add granular cameo controls; Bill Peebles, head of Sora, posted on October 5 that users can restrict political or abusive uses of their cameos. Controls reduce risk, but they’re post-hoc defenses. The core question—who bears the burden of policing millions of generative outputs in real time—remains unanswered.

Bias, personalization, and the ethics of an infinite synthetic feed

Sources

• MIT Technology Review — The three big unanswered questions about Sora (2025-10-07)

• MIT Technology Review — The Download: carbon removal factories’ funding cuts, and AI toys (2025-10-08)

• MIT Technology Review — AI toys are all the rage in China—and now they’re appearing on shelves in the US too (2025-10-07)