Scaling 9,000 GPUs in Six Hours: Inside Luma's Dream Machine Launch
Key points
Key takeaways from Tank Talks interview with Keegan McCallum, co-founder and CEO of uRun and former head of ML infrastructure at Luma AI:
The Dream Machine launch. In June, Keegan led Luma's inference infrastructure for the Dream Machine launch, scaling from five hundred to nine thousand H100 GPUs in about six hours when requests hit one hundred thousand, processing over half a million videos in twelve hours and reaching a million users in four days.
Scaling under whiplash. The team was severely under-provisioned,and his first lesson was to plan for success. A large queueing system kept requests flowing,and a portable inference stack with no hard Kubernetes dependency,built to run on raw VMs, Slurm, or Kubernetes,let them absorb spare GPUs without week-long rewrites.
The production-infra gap. At Luma,he saw best talent gravitate toward training the next model,leaving production infrastructure second-class; he argues even leading labs ship broken infrastructure as a result,which convinced him the market needed a dedicated,elite ML-infrastructure team serving any model lab.
uRun's bet. Keegan founded uRun as an inference provider, "batteries included" for persistent,steerable,interactive AI: continuous streams of video,audio,text,and images rather than one-shot request-response,with fast media transport,so AI feels like a presence sitting with the user,rather than an API you email back and forth.
Why video models are harder. Video generation models are compound workflows—VAE encoders and decoders,denoising stages,and pixel encoders,orchestrated like small pipelines where each stage has different GPU requirements;and unlike LLMs,which are memory-bound,video models are compute-bound,needing many GPUs working in parallel to render chunks of video at once.
Handling viral growth. AI launches have tight two-to-three month hype windows,so founders must ride the momentum,and ship continuously; he also stresses proper on-call rotations and staffing plans before the spike,because overnight heroics do not sustain week-long viral load.
Hiring for moats. Don't be scared of hard problems,because that's where the moats are: hard infrastructure problems attract elite talent,capital,and customers,and that gravity is what uRun bets on when recruiting a top ML-and cloud-infrastructure engineering team.
The interactive shift. Keegan predicts an explosion of highly interactive,real-time AI experiences,and argues what unlocks them is stateful,real-time infrastructure,rather than just better models,with persistent AI sessions rewiring agent design,content workflows,and how teams are built.
Read more: Tank Talks (YouTube) · Tank Talks Substack
