How the bus runs
One queue, two prices. The buyer names the most they will pay per second of video; each machine names the least it will work for. The platform’s job is to put every job on the cheapest idle machine that fits.
A job, stop by stop
Submitted
A prompt, a length of 4–15 seconds, one of three landscape sizes, a Detail setting and a cap in whole cents per second. The cap cannot be changed afterwards.
Waiting to start
The job sits in the queue until a machine that fits asks for work. Studio says so when no machine is at or below the cap.
Claimed
A machine claims the job. The platform refuses the claim if a cheaper machine is idle, and sets aside the machine’s rate × the requested seconds.
Ready
The video is checked on our side, then offered for download. The charge is the machine’s rate × the seconds delivered.
Why a machine does not get a job
Each of these is temporary and costs the job nothing – it simply keeps waiting.
- No rate for the model
- The provider has not set a price for the model on that worker.
- Price above the cap
- The machine asks more than the buyer will pay.
- Price not competitive
- A cheaper machine is idle right now, so it goes first.
- Queue full
- The worker already holds as many jobs as it accepts – two by default.
- Balance too low
- The buyer’s balance cannot cover the job at that machine’s rate.
What happens when something fails
Money is charged only for a successful delivery of work that was actually executed. A job that ends as “Could not be made” is not charged. If our own re-encoding step fails, the job returns to the queue without using up an attempt.
How a delivery is checked
The platform does not rely on the worker’s word. It checks the size against what the worker declared, computes the file’s digest itself, and checks the media type and what the model should have produced. Files above 64 MiB are refused.
What this does not do yet
Card charges and provider payouts are not switched on, so today’s jobs move no real money. There is no buyer API and no deadline tier; both are planned. And there is no confidential computing: a job runs on a provider’s machine, so treat prompts as not confidential.