8,000 images a month: the unit economics of an on-prem AI pipeline

The first version of our pipeline ran fully in the cloud. Quality was excellent. The economics weren't: every image carried GPU rental cost, and throughput was hostage to cloud queue times. At catalog volume — thousands of images a month — per-image cloud billing compounds into a permanent tax.

The two changes that fixed it

First, a custom caching method that cut inference time roughly in half with no quality drop — the same image that took two and a half minutes dropped to under ninety seconds. Second, moving production on-prem: the pipeline now runs on the studio's own RTX-class workstations, where a processed image costs electricity, not rent.

Cloud didn't disappear — it became the overflow valve. When volume spikes past local capacity, we burst to rented GPUs, and a fresh cloud instance is installable in about 10 minutes. On-prem for the baseline, cloud for the peaks.

Who runs it day to day

Not us — and that's deliberate. Each deployment creates an AI operator role on the studio's own team: the person who runs batches, watches the review gate, and flags edge cases. We install, train the operators, and stay on for support, debugging, and model upgrades. The studio owns its throughput; we keep the system fast.

Why this matters if you're evaluating vendors

Ask any AI imagery vendor two questions: where does inference run, and who owns the pipeline when the contract ends? If the answer is "our cloud, per-image pricing," you're renting a capability. Our answer is: your hardware, your operators, our system — licensed, supported, and improving.

See how the deployment model works. → Explore the deployment model