Scroll
Innovation thesis
Creative futures
What we believe about the next era of creator tools — and where we are putting the work. Five beliefs. The evidence for each. What we are building because of them.
What we believe about the next era of creator tools — and where we are putting the work. Five beliefs. The evidence for each. What we are building because of them.
So the scarce thing stops being production and starts being judgment. Judgment is taught by people. People need footage they own, consent that holds, and a record of what they chose. Nobody is building that record.
We are. We clear creator video for AI, licence it more than once, and keep what it teaches.
Everyone else is building the printer. We are building the editor — and we own the paper.
Two curves moved in opposite directions. They already crossed.
Years of schooling between the question asked and the answer given. Anthropic Economic Index, June 2026.
Which means the winnable ground is design, and small teams win design.
You can scrape a video. You cannot scrape a signature.
copyright suits open and unresolved
now demands proof of where training data came from
paid to individual creators so far — every cheque went to an institution
Ask a frontier model for something real and you get a costume.
Stock. Cinema. Crews, lighting rigs, colourists.
A phone. A kitchen. Someone who is not acting.
Nobody owns the layer between what a machine makes and what a human will stand behind.
Consented, vernacular, performance-labelled creative work — and a model trained on what humans chose. We are building both. One feeds the other.
Creator economy in 2026. Headed past $500B.
People making content. Eight million make a living.
What brokers take from rights holders today.
Our split — in the creator's favour, published.
Clear feeds the corpus we licence. Verify feeds Judge. Judge makes Verify better.
A brand needs forty cutdowns by Friday. Here is what happens.
Creator drops the brief. The roster reads it and proposes a shot plan.
Creator picks the autonomy level. Draft only. Draft and cut. Or full pass.
Agents pull cleared footage, cut, grade, caption. Every asset carries its rights.
Forty cuts land side by side. Visual diff shows what changed and why.
Creator approves eleven, kills the rest. Each call is captured with its reason.
Step 05 is the product. Twenty-nine rejections and eleven approvals is not a chore. It is a labelled dataset nobody else is collecting.
Raw footage is pixels. This is taste, written down.
what was asked for
every variant produced
which one shipped
why the others died
how it performed
Priced by the pound.
Almost nobody has collected one at scale.
You cannot train a model to judge without watching humans judge first.
Generators cost a billion dollars and we would lose. Judges cost a fraction and nobody has one.
Fine-tuning open weights is now a managed service priced on compute. A specialised model is a project, not a moonshot.
Creators bring consented work.
They approve, they reject.
Every call is captured.
The model learns taste — and feeds it back to the community.
Every step is something we do to footage we already hold. Gross, per hour, over a 24-month window.
Twelve times the return, without shooting a single new frame.
| # | Line | Type | Live | We keep |
|---|---|---|---|---|
| 01 | Benchmark accessscore your model against ours | Recurring | Q1 2027 | 30% |
| 02 | Corpus licencenon-exclusive, 24 months, field-limited | Deal | Q2 2027 | 30% |
| 03 | Managed collectionwe shoot it, cleared, to your spec | Services | Q2 2027 | 40% |
| 04 | Corpus subscriptionthe flow, committed annually | Recurring | Q4 2027 | 30% |
| 05 | Judgment APIpriced per evaluation | Recurring | 2028 | 85% |
Lines 01 to 04 pay for the model. Line 05 is the only one that does not spend the asset it sells.
Labelled footage prices at four times raw. Same pixels. Different product.
| Product | Unit | Price |
|---|---|---|
| Benchmark access | Per seat, annual | $30,000 |
| Corpus licence — standard footage | Per minute, non-exclusive | $3.00 |
| Corpus licence — preference chains | Per minute, non-exclusive | $12.00 |
| Managed collection | Per finished hour, cleared | $9,000 |
| Corpus subscription | Annual commit, 200 hrs minimum | from $200,000 |
| Judgment API | Per 1,000 evaluations | $40 |
Anchored to observed market rates of $1–$4 per minute. Re-verify before any negotiation.
Over a 24-month licence window. Every number below is per single hour.
A labelled hour returns 2.7× a raw hour. Same footage. We just wrote down what the human decided.
Modelled on 1,200 cleared hours. The most consequential line in any contract we sign.
of 2029 net is recurring — subscription and API
cumulative to creators by 2029, published and audited
our share against theirs, on every licensing dollar
Brokers in this market take 15 to 30 percent. We take less and we publish it.
A fund-backed rival carrying margin expectations cannot publish 70/30. We can. The supply is the strategy.
Take the standard cut and the community works it out in eighteen months. Then there is no second act.
A creator who sees what their work earned tells another creator. That is our sales team.
Creators licence through us. They never sell to us. Likeness is signed apart from footage. Opt-out reaches backwards. The community decides who we are allowed to sell to.
You should hear this from us first.
Old releases never mentioned AI. Assume today the archive cannot be licensed.
Remediation is week one. Cleared percentage reported monthly until it passes 70.
Maybe a dozen firms can write a real video data cheque, and they move slowly.
Collection and the API exist to cut our dependence on lab procurement.
Adobe and the agent frameworks will ship evaluation natively.
Inside their own walls. A judge that scores rivals has to be neutral. They cannot be.
If the visual gap narrows toward zero across the next four Economic Index reports, we are wrong and we will say so.
Clear the footage. Licence it more than once. Keep what it teaches.
Start in San Antonio, because consent is a relationship and relationships are local.