Grok Shipped Seven Minutes of AI Video. A Finished Minute Now Costs More Than 175x What It Sells For

Three months after promising a full-length Odyssey feature, xAI published seven minutes of one on X instead. Grok Imagine posted Episode 1 of its Odyssey series on Oct 23, 2026 — 7 minutes 12 seconds, with English dialogue, made in collaboration with Wonder Studios and director Dev Walker — and it drew millions of views within days. Generative videocan now hold a scene, a face and a voice for seven minutes straight. What it cannot yet do is find a buyer for the finished minute.

In the only sale that has cleared, a finished minute of AI video fetches about seven cents from the buyer. Fountain 0's Odysseus: The Fall — 135 minutes, every scene rendered by models — was budgeted in the mid-five figures and is rented direct to viewers at $9.99, which works out at $0.07 a finished minute. Generating that same minute costs $13 to $220 at the model layer depending on tier and resolution. No distributor has closed that gap, and the gap is the whole story.

The distribution layer is where generative video's unit economics invert. A finished minute is a fixed production cost that can be streamed to unlimited viewers at near-zero marginal cost, so a feed can amortize one generation bill across unlimited impressions; a subscription bundle cannot, because every minute of AI content competes against subscriber revenue already committed to licensed titles. That is why the first frontier-lab feature shipped as a seven-minute episode on X and the first released AI feature sold at $9.99 on its own site.

The platform response confirms the mechanism. YouTube's July 16, 2026 clarification disqualifies three classes of inauthentic content from Partner Program monetization — generic, repetitive or template-based video including cookie-cutter AI output, off-putting content, and AI personas discussing sensitive topics. The one channel that can absorb AI volume at scale is starting to price quality instead of quantity.

Generative video has not solved rights, and the music catalog is where it is being tested first. On June 5, 2026 the American Federation of Musicians sued Warner Music Group and Universal Music Group in Manhattan federal court, arguing that the labels licensed members' recordings to Suno and Udio without permission and that the resulting settlements and prospective licences allow continued use without compensating the artists. Warner settled with both Suno and Udio in 2025; Universal's case against Suno is still open.

And the licensed archive went bankrupt. Getty Images, which owns one of the few fully licensed visual libraries and licenses music and manages digital assets on top of it, filed for Chapter 11 protection in Delaware on Oct 21, 2026, five days ago. Its June-quarter balance sheet carried $2.06 billion of total debt against $51.6 million of cash, and second-quarter revenue of $229.1 million was already down 2.5% year on year. The restructuring brings up to $533 million of new sponsor equity from the Getty family and Koch Equity Development. Strategic value and equity value have come apart: the archive is worth more to AI models than to its own shareholders, because the models do not pay for it.

If a finished minute fetches seven cents and costs $13 to $740 to make, the only layer that can earn a normal margin is the one selling the tools and the inputs. NVIDIA earned a 75.0% gross margin on $96.22 billion of quarterly revenue in the quarter ended July 26, 2026, with data-center revenue of $89.0 billion, up 117% year on year. Micron earned an 86.8% gross margin on $54.23 billion of revenue in its September quarter. And xAI, which ended 2025 with more than one million H100-equivalent GPUs, raised $20 billion in a Series E round in January 2026 with NVIDIA and Cisco Investments among the strategic investors — so part of the cost of every generated minute already returns to the chip maker as equity.

The model layer is where AI video revenue is finally visible, and it is smaller than the headlines suggest. Kling AI, Kuaishou's video model, booked RMB 650 million in the March 2026 quarter, more than 300% above a year earlier, on an annualised run rate of roughly $500 million — after ending December 2025 at a $240 million run rate with more than 60 million creators and 600 million videos generated. Kling rendered every scene of Odysseus: The Fall alongside image and research models from Alphabet and language editing from Anthropic's Claude, which is the clearest evidence yet that a first AI feature is a multi-vendor stack rather than a single model.

The next four quarters will be decided by cadence and contracting, not by spectacle. Watch three things: whether Episode 1 becomes a release schedule, because each episode repeats the generation bill without a distributor contract attached; whether any premium streamer or studio licenses AI-native content, given that the first released AI feature found no buyer; and the AFM case in Manhattan, which tests whether the 2025 label settlements survive as prospective licences.

Over one to three years, the rights layer captures what the studio layer gives up. If inference costs fall toward the $3 to $6 fast-tier range and continuity tooling matures, the distance between a $13 cost and a $0.07 price narrows — but only if somebody pays per minute. The structural risk to the whole chain is that nobody ever does, and the first business model that clears is advertising and branded content, where the asset is a fixed cost and the impression is the unit.

What would change the thesis: a distributor paying more than $13 for a finished minute, or an ad-supported AI series whose impressions cover the generation bill many times over. Until one of those shows up in reported revenue, the correct reading is that this is a compute-and-rights trade, not an entertainment trade.