Hold one question through everything that follows, because the answer is not where the week’s reaction has been looking for it. What, exactly, is the scarce thing in video?
Not the answer people assume. Keep it open for now.
OpenAI showed a model called Sora on Thursday that generates video up to a minute long from a line of typed text, holding a consistent scene with multiple characters and specified motion through the length of the clip. It is not available to anyone outside a group of testers examining it for risks. The demonstrations are the entire public record so far.
What follows sets two activities beside each other. Producing a minute of video, which is what was demonstrated. And using a minute of video, which is what anybody paying for video is actually buying.
The first activity: producing the minute
Take the demonstration at face value, which is the generous reading and also the useful one.
A minute is a genuinely hard target for this class of model, and the difficulty is not resolution. It is consistency over time. A character has to keep the same face across six hundred frames. An object that leaves the frame has to be in the right place when the camera comes back. Light has to fall from the same direction throughout. None of that is guaranteed by a system generating frames, and holding it for a minute is a real result.
So on the narrow question of whether a machine can produce a coherent minute from a sentence, the demonstration is the answer, subject to the usual caution that a published reel is a selection.
The choice of a minute as the headline number is itself informative. A minute is longer than any single shot in most finished video and shorter than almost any finished piece, which places it precisely at the boundary where a clip stops being a fragment and starts having to hold together. It is the right thing to measure, and it is a measure of the generator rather than of the job.
The second activity: using the minute
Now put a minute of video into an actual job and watch where the money goes.
A thirty second advertisement is not thirty seconds of footage. It is a brief, several rounds of concept, a script that legal reads, a storyboard the client signs, a shoot, an edit, a client review, a second edit, a music license, a voice session, a compliance review in regulated categories, versioning for six placements and four aspect ratios, and a delivery deadline that all of the above has to fit inside.
Shooting is a line in that. It is rarely the largest line, and it is almost never the line that determines whether the project is delivered on time.
The same holds outside advertising. A training video, a product explainer, a documentary sequence: in every case the footage is an input to a process whose cost is dominated by specification, revision and approval. Anybody who has produced video for a living recognizes this immediately, and anybody who has not tends to assume the camera is the expensive part.
What the difference teaches
Set the two side by side and the answer to the held question starts to show.
Generation attacks the acquisition of footage. That is a real cost and it is the most visible one, because it is the part with equipment and people standing around in a car park at six in the morning.
What it does not attack is revision. The client who says make it warmer, and shorter, and can the woman be holding the product in her left hand, is asking for a change to a specific thing that already exists. A system that produces a superb minute from a sentence, and a differently superb minute from a slightly altered sentence, has not answered that request. It has offered a replacement.
Controllability, not generation, is the thing production actually needs, and it is a harder problem, because it requires the system to hold everything constant except the one element being changed.
Rights are the second unattacked cost. Somebody has to be able to say, in writing, that the client owns what was delivered and that nothing in it infringes. That sentence is not a technical artifact. It is an indemnity, and it is why stock libraries can charge for images that are freely visible.
Why the release pattern is the informative part
The most useful fact this week is not in the clips. It is that the thing was shown and not shipped.
The company said the model is with testers assessing risks before any general release. Announcing a capability months ahead of availability is a choice with known effects: it sets the terms of the conversation, it puts competitors on a clock, and it lets the reaction to the idea be observed before anything is in anyone’s hands.
It also means that every claim currently in circulation rests on material the company selected. That is not an accusation; it is the ordinary condition of a demonstration. But it is worth holding, because a model that produces a minute reliably and a model that produced these minutes are different claims, and only the second is currently evidenced.
The answer
The scarce thing in video was never footage.
It is the ability to get a specific thing, agreed in advance, delivered on a date, that somebody will indemnify. Footage is the raw material. Specification, revision and clearance are the product, and they are where a production budget actually goes.
Which is why the interesting question about this technology is not whether it makes video cheaper. It is whether the ability to produce an infinite supply of nearly-right footage makes the specification problem better or worse, and there is a case for worse: a supply of plausible alternatives is exactly what makes an approval process longer.
What to watch for
Over the next year the thing to watch is not sample quality, which will improve and will be reported as though improvement were the question.
Watch for editing. Whether a released product lets somebody take a generated clip and change one element of it while holding the rest still. That capability, if it arrives, is the one that turns a demonstration into a tool, and its absence is what will keep this in the category of things that produce impressive material nobody can use for the job in hand.
And watch the indemnity. The day a vendor offers written cover on the output is the day this becomes a line item in a production budget rather than a topic in a meeting. For anybody doing this work, that clause will change more about the week than any number of frames.
