HWNEWS
Self driving cars are posting their best safety numbers yet, but no one has measured whether the sensors and AI inside them can hold up for a full decade of always on service.
A self-driving car on the road today is expected to remain safety-critical into the mid-2030s. Its perception stack (the AI software that turns raw sensor data into a model of the road) runs AI inference continuously. And the engineers building it admit, on the record, that they don't yet know whether the hardware will hold up over that decade.
The freshly released Insurance Institute for Highway Safety data on Waymo is the strongest safety case the industry has ever published. Waymo's L4 vehicles (L4 means fully driverless in a defined operating area) experienced 68% fewer reportable crashes than human drivers and a 91% lower rate of rear-ending another vehicle, per the IIHS report cited in Semiconductor Engineering's analysis of the aging problem. Wire coverage will read those numbers as a win. The harder question is what they look like after 100,000 hours of always-on perception on the same camera, radar, and lidar (laser-based distance sensing) modules.
Traditional automotive electronics were designed for intermittent duty. A power-window motor runs when you press the switch. A backup camera wakes up in reverse. Modern software-defined vehicles (SDVs, vehicles whose driving functions are controlled primarily by updatable software rather than fixed hardware) keep their perception stack awake every moment the car is on, fusing camera, radar, and sometimes lidar feeds to track everything around the vehicle. That workload is closer to a data center than to a power window. It is sustained, thermally stressful, and novel for parts originally rated for short bursts.
"The AI is always on. That is the real problem," a Rambus executive told Semiconductor Engineering, framing the issue as continuous inference aging hardware that wasn't designed for it. The mechanism is straightforward: components that run hot for long stretches drift in spec. Cameras shift color balance. Radars drift in power output. Lidar lasers age in pulse characteristics. The drift is gradual, which is exactly why it's hard to catch in a six-month test cycle.
The safety horizon makes the gap more uncomfortable. A car sold in 2026 is expected to remain in service through at least 2036. Regulators, fleet buyers, and insurers are being asked to take the decade-scale safety case on faith, because the decade-scale data doesn't yet exist. Workloads vary by driving frequency, aggressiveness, region, and corner-case exposure, which means two identical sensor suites can age very differently. The IIHS figure is a snapshot of year-one behavior. The industry's safety claim is a ten-year promise.
On top of that, the AI itself is changing. Software-defined vehicles receive regular over-the-air updates, sometimes monthly, that retune the perception models on the road. That means the AI's expectations of the sensors are shifting while the sensors themselves are aging. A model trained on a fresh camera is being asked to interpret data from a camera that has spent three winters in a Minneapolis fleet. The industry has not published a study on how those two curves interact.
There is a legitimate counterargument the engineers raised: traditional automotive electronics are famously over-spec'd, and suppliers may have unpublished decade-scale reliability data from related programs. If that data exists and gets shared, the story shifts. If it doesn't, the question stays open and the question is the story.
What makes the gap actionable is who has the data and who doesn't. Automakers, the OEMs, are pouring tens of billions of dollars into autonomous deployment. Tier-1 suppliers (chip and sensor vendors that supply directly to automakers) build the parts. Regulators set the rules. Fleet buyers write the purchase orders. None of them have published a multi-year aging study for an always-on perception stack, and the industry hasn't yet settled when sensors degrade, how AI models should manage that drift across software updates, or how to design whole systems that meet rapidly scaling AI.
The test for a fleet buyer or regulator is concrete: ask the vendor for year-seven and year-ten failure-mode data on the specific sensor suite that will ship in the cars you are buying or regulating. If the answer is "we are still gathering that," treat the IIHS numbers as year-one evidence and price the year-ten uncertainty accordingly. The 68% figure is real. The 91% rear-end reduction is real. What neither measures, yet, is whether the same vehicles will still be earning those numbers when the warranty runs out.