AI startup Visko deployed its Orbis live multimodal inference pipeline on a single Google Cloud Tensor Processing Unit slice.

Visko and Google Cloud engineers deployed Orbis’s complete live inference pipeline—including video generation, super-resolution upscaling, and synchronized audio—on a single Cloud Tensor Processing Unit (TPU) slice. Google reports that the collaboration moved from kickoff to production handover in 30 days, following the start of work in August 2026.

The reported pipeline rendered at 1.07 times playback speed. Google emphasizes that the serving requirement is continuous real-time delivery: each new video-and-audio chunk must finish before the viewer completes the previous chunk.

Orbis begins streaming from a text or image prompt and accepts new steering prompts while the stream continues. Visko launched its public demonstration on September 1, 2026; the source describes scene changes occurring without restarting the stream or breaking continuity.

The deployment unit is a TPU slice, not a single TPU chip or a GPU server. The supplied excerpt does not state the slice’s chip count or generation, resolution, simultaneous-session capacity, or cost per stream, limiting comparisons with other inference deployments.