Autonomy compute

Who owns the choke. Strip the GPS and the datalink out of a small drone — which is exactly what Ukraine-style electronic warfare does — and the aircraft must perceive, map and decide entirely onboard, inside a 10–50 W power budget. That brain is the chokepoint, and NVIDIA owns its top tier. Skydio's X10D, the airframe that headlined DIU's Blue UAS Challenge, runs its Autonomy Engine on a Jetson Orin SoC for GPS-denied ISR, and the roadmap is Jetson Thor — 2,070 FP4 TFLOPS at 40–130 W. The moat is not TOPS; it is software gravity. Western autonomy stacks were built CUDA-native — Skydio has shipped Jetson since the 2019 Skydio 2, and trade press reports Shield AI's Hivemind running on modular Nvidia edge compute aboard V-BAT — so swapping silicon means re-validating the entire perception stack at flight-test prices, on DoD qualification timelines.

The second tier is the interesting trade. Qualcomm's QRB5165 (15 TOPS at single-digit watts) is the brain of ModalAI's VOXL 2, the de facto Blue UAS Framework companion computer; it rides alongside Orin in the X10D, and, packaged on Lantronix Open-Q SOMs, it is the NDAA-compliant compute inside Red Cat's Black Widow for the Army's Short-Range Reconnaissance program. Ambarella stakes the perception-per-watt niche with its 4 nm CV7-class CVflow SoCs and early drone design wins (consumer-led so far) — and the tell is that Hanwha just locked up Ambarella silicon in an $800M+, ten-year agreement for its Vision unit's video-security, robotics and industrial lines. That deal is commercial edge-AI, not a drone program — but locking up a decade of supply is how strategic buyers behave when they believe edge-AI silicon is a choke.

What breaks it. A challenger needs three things simultaneously: performance-per-watt parity (Hailo's NPUs already deliver 26 TOPS at a few watts), an NDAA-clean supply chain, and — the hard one — portability of CUDA-native autonomy software. Low-power FPGAs from Lattice nibble at deterministic sensor-fusion and pre-processing workloads, but they complement rather than displace the NPU. The genuine bear case is doctrinal, not competitive: if attritable FPV munitions win the budget fight, autonomy compute gets value-engineered toward commodity NPUs and the premium choke deflates.

The repricing clock. Three forcing functions inside twelve months. First, the FAA's Part 108 BVLOS final rule — the reopened comment window closed February 11, 2026, and after the slip past the executive order's spring target our read is an H2 rule — effectively makes onboard detect-and-avoid and autonomy compute mandatory equipment for scaled commercial operations. Second, the FY27 NDAA — marked up by HASC on June 4, Senate next — keeps hard-coding compliant supply chains, which is precisely the moat for Qualcomm- and Lantronix-based modules. Third, the August prints: NVDA (Aug 26E) and AMBA are the cleanest quarterly reads on whether autonomy-silicon demand is inflecting (QCOM's IoT segment and LTRX's SRR shipments carry signal too). Own the choke (NVDA, AMBA), rent the second tier (QCOM, LSCC), and treat LTRX as the leveraged small-cap expression.