Nothing But Datacenters
In progress
Single-channel video with sound — 90-second test edit; final work will run 30 minutes as a continuous loop
An image feedback loop in which a diffusion model reinterprets its own output frame by frame. The rectangle persists throughout; everything else is renegotiated. Readings slide across scale rather than shape — a silicon die becomes a server rack becomes a windowless building becomes a campus seen from orbit — so the form holds while its meaning drifts by five orders of magnitude.
The piece runs on the infrastructure it depicts. Producing it consumes GPU-hours in a facility indistinguishable from the ones on screen, which makes the work less a representation of datacenters than a small instance of one.
No people appear. The negative prompt excludes technicians, and uninhabited infrastructure is the argument: the apparatus does not need anyone in frame to be about people. Logos are excluded as well, keeping the subject a category rather than a company.
Threaded through the sequence is a register of consequence — cooling towers, evaporation ponds, transmission corridors, diesel backups, heat exhaust — alongside a register of failure, dead racks and e-waste pallets. The model renders server corridors beautifully; that tendency is used rather than corrected. The critique stays inside the imagery.
What’s posted here is a short test edit — enough to see how the form churns and how the scale-slides read, not yet enough to feel the duration, which is where the piece is meant to do its work. The full render is 43,200 frames.
Colophon
Generated with SDXL via a local img2img pipeline on an RTX 4090.
Hot Bit Soup, 2026.