Wing, Alphabet’s delivery-drone operator, already flies on Nvidia’s Jetson Orin Nano Super. It now plans to evaluate Jetson Orin Nano 2, a compact edge computer Nvidia says doubles inference performance in the same form factor and uses less power at equivalent output. Wing’s head of perception, Dinuka Abeywardena, put the operational case plainly, delivery depends on machines that understand the world in real time. More onboard compute is meant to make perception, obstacle reasoning and energy use good enough that a yard drop does not wait on a radio link.
That is not 'AI flies a drone' in the slogan sense. It is a shift in where the decision lives. A drone that classifies a tree, a child, a washing line and a safe patch of grass on the aircraft itself can keep flying if the cellular link stutters. It can abort a landing without asking a cloud model first. For operators chasing volume, Wing, and the Indian and Southeast Asian delivery bets that want the same economics, the attraction is obvious. Less backhaul, faster cycles, longer endurance.
The regulator’s problem arrives with the software update. A flight-control computer that is frozen at certification is a known object. A perception stack that improves every quarter is not. Who certifies the model that will be on the aircraft next month? How does an authority test the rare case, the black dog on wet asphalt, the kite in rotor wash, when the training set will keep changing? If the drone chooses the wrong patch of grass, is the fault the operator, the module maker, the model vendor or the airline-style air operator that filed the waiver? Audit logs help only if the decision is explainable enough to reconstruct.
BEYOND Phase 2 and proposed BVLOS rules still assume performance standards a human can write down. Onboard reasoning strains that habit. A detect-and-avoid box can be measured against a miss distance. A network that reasons about a backyard cannot be frozen without killing the product that justified the chip. India’s Digital Sky and similar Asian frameworks are even less ready. They were built to register airframes and fence airports, not to supervise a learning system over a city.
The hardware will ship. The bottleneck is whether civil aviation will accept a brain that keeps changing after the type certificate is signed. Until that question has an answer, the safest drone on paper may be the one whose software is not allowed to get smarter.