OpenAI’s Sora serves an endless feed of AI‑made 10‑second videos — hyperreal cameos, trademarked characters, and surreal micro‑memes streamed at scale. The app is a technical stunt, a legal lightning rod, and a potential emissions problem all at once — and it’s forcing a reckoning about what unrestricted generative video might cost.

Sora burst into public view in early October 2025, climbing to the top of Apple’s US App Store after a soft launch OpenAI rolled out by invite. The app supremely simplifies a hard technical task: on‑demand video synthesis that stitches appearance, voice, motion, and music into short clips that feel real enough to make viewers suspend disbelief. That convenience is the point — but it also concentrates three emerging tensions in generative AI: compute scale and carbon footprint, fuzzy intellectual‑property boundaries, and the social harms of photorealistic deepfakes.

How Sora makes a ten‑second fantasy

Those tensions matter now because Sora is not a research demo or a niche lab toy. OpenAI is treating it like a platform. CEO Sam Altman told users on October 3 that “we are going to have to somehow make money for video generation,” signaling productization and monetization. When a single company begins offering unlimited, cheap video synthesis to tens of millions of users, the technical tradeoffs become civic problems: who pays for the electricity, who sues when a celebrity is simulated, and how do models trained on copyrighted media get governed?

How Sora makes a ten‑second fantasy

At its core, Sora combines advances from multiple subfields: large language models for dialogue and control, diffusion and transformer‑based video generators for motion, and voice‑cloning pipelines to recreate timbre and prosody. OpenAI has engineered an interface that hides these moving parts; a user selects or creates a “cameo” (a photoreal avatar and voice), then composes a prompt and gets a 1–10 second clip back.

The climate tab nobody’s totaled (yet)

That simplicity masks enormous compute. Generating even short, high‑framerate video requires many more floating‑point operations than a text token. The models involved must produce coherent spatial frames, consistent identity across time, and synchronized audio — which typically means combining an image or frame generator with a temporal model that captures motion dynamics. OpenAI has not published the specific model sizes or flop counts for Sora; the company’s blog announcing the product framed the problem as both engineering and product design and noted monetization will be necessary to sustain video workloads.

For researchers, the technical leap is impressive. For practitioners, it’s a cost problem writ large. Video synthesis uses more memory, more parallel compute on GPUs or specialized accelerators, and longer inference pipelines than text. That’s why OpenAI’s decision to enable unlimited generation is as much an operations challenge as a modeling headline: more users producing more clips will rapidly multiply backend energy demand and hardware utilization.

The climate tab nobody’s totaled (yet)

Copyright, cameos, and the lawsuit horizon

Conservative engineering estimates and prior literature suggest the energy and emissions burden could be substantial. Strubell, Ganesh, and McCallum’s 2019 analysis of large NLP training runs showed that training state‑of‑the‑art models can emit carbon on the order of hundreds of tons. Inference at massive scale — billions of short video generations per month — would compound that baseline, especially if OpenAI keeps quality high and latency low.

OpenAI’s Sora team has acknowledged the challenge indirectly: Altman and others have pointed to capacity investments and new data centers. But the company has not released per‑generation energy figures, nor has it published a lifecycle estimate for the app. That opacity matters. If Sora’s economics depend on cheap, carbon‑intensive grid power, the result could be a rapid spike in emissions tied directly to entertainment consumption. The math is straightforward: even a tenth of a kilowatt‑hour more per request, multiplied by millions of requests, reaches megawatt‑hours and metric tons of CO2 quickly.

There are engineering levers to reduce the load — quantized models, sparsity, on‑device inference for cheap use cases, and scheduling workloads to low‑carbon hours — but those require tradeoffs in latency, fidelity, and cost. The industry has moved before on efficiency when regulatory or market pressure demanded it. If governments or markets start conditioning platform access on measurable carbon intensity, OpenAI will face the same calculus as cloud providers and hyperscalers: invest in cleaner power, buy offsets, or throttle demand.

Fairness, hallucinations, and the social seamstress

Copyright, cameos, and the lawsuit horizon

Sora’s permissive defaults — the app synthesizes trademarked characters, copyrighted music, and realistic depictions of public figures — have already drawn pushback. MIT Technology Review reported that OpenAI sent notices to rightsholders asking them to opt out if they didn’t want their IP included in Sora’s training and output; the company said it would build “granular control” tools. The Wall Street Journal first reported on the opt‑out letters, intensifying scrutiny.

Legally, Sora sits at an unsettled junction. U.S. copyright law has yet to be fully tested for model training and generative outputs; whether fair use applies to large‑scale ingestion and synthesis is an open question that courts will refine over the next several years. Beyond copyright, the cameo feature — where a user can create a hyperreal avatar of a living person — raises personality‑rights and defamation risks. OpenAI has added controls letting people restrict political uses and certain phrases, but those are policy filters, not legal immunity.

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)

• OpenAI — Sora announcement (2025-10-03)

• arXiv (Strubell, Ganesh, McCallum) — Energy and Policy Considerations for Modern Deep Learning (2019-06-04)