Running inference at 8,000m with no onboard GPU
How we helped deploy an autonomous robot on a high-altitude expedition by removing the onboard GPU entirely and streaming inference through our cloud stack.
Netflix × DeepMind × Geologic DomeThe expedition
A project with Netflix involving an expedition led by Geologic Dome and 14 Peaks. We worked alongside DeepMind to adapt the robot for operation in extreme high-altitude environments — some of the most hostile conditions on earth.
The constraint
A major constraint came from expedition logistics: Sherpas cannot safely carry loads exceeding 18 kg per person. This created a serious challenge because the robot's standard configuration exceeded that limit. Every gram mattered, and the heaviest single component was the onboard GPU.
The redesign
To address this, we redesigned the system architecture so the robot could be disassembled into two modules for transport. We also removed the onboard GPU (Jetson) entirely and instead ran inference through our cloud streaming stack. This allowed us to significantly reduce weight while still running large vision-language-action policies remotely.
The result
By offloading inference to Reflex's cloud infrastructure, the team eliminated the heaviest compute component, brought the robot within the 18 kg carry limit, and maintained full model capability — no quantization, no distillation, no compromise. The robot operated autonomously at altitude with datacenter-scale models running thousands of kilometers away.
Have a similar challenge?
If your robot is constrained by onboard compute, we can help you offload inference to the cloud.
Get started