Tesla FSD Ran Offline
Tesla's Full Self‑Driving reportedly navigated San Francisco correctly during a power outage that disrupted other fleets—running all inference locally on the vehicle rather than relying on cloud connectivity. The episode highlights tradeoffs between edge inference and cloud‑assisted autonomy. ( )
San Francisco’s December 20–21, 2025 blackout forced Waymo to suspend its robotaxi service after multiple vehicles stalled at darkened intersections, while social‑media clips and company responses showed Tesla‑equipped cars continuing to drive using onboard systems and drew a public comment from Elon Musk. (techcrunch.com) Tesla’s Full Self‑Driving stack is architected to perform neural‑network inference on an in‑vehicle AI computer (Tesla’s HW3/HW4/AI‑computer family) rather than routing real‑time perception decisions to remote servers, and Tesla’s support documentation confirms the AI computer is used to deliver FSD features via over‑the‑air updates. (en.wikipedia.org) Hardware performance figures discussed publicly place HW3 in the low‑hundreds of TOPS and HW4/AI4 in the mid‑hundreds (papers and analyst reports cite roughly ~144 TOPS for early HW3 estimates and ~500 TOPS class capability for HW4), illustrating why Tesla can run large vision models locally but must accept per‑vehicle compute, power and thermal budgets. (en.wikipedia.org) Waymo’s fleet behavior during the outage highlighted a different design point: reliance on HD maps, remote coordination and teleoperation fallbacks led many robotaxis to stop or wait for remote confirmation when traffic signals failed, producing hundreds to more than 1,500 recorded “stoppage events” that prompted city hearings. (cnbc.com) Edge vs. cloud tradeoffs made visible by the incident: on‑device inference reduces dependence on low‑latency network availability during infrastructure failures but shifts costs into per‑vehicle silicon, energy and thermal design, while cloud‑assisted approaches can centralize compute and map updates at the expense of single‑point failure risk when connectivity or power goes down. (blog.roboflow.com) San Francisco officials and Waymo have since documented software and operational changes and held oversight hearings about the December blackout’s impact on robotaxis, marking a near‑term regulatory and engineering focus on resilient fallback behaviors, emergency‑mode protocols, and fleet‑level communications during citywide infrastructure failures. (cnbc.com)
