Growing numbers of Tesla owners running Hardware 3 (HW3) vehicles are reporting that the Autopilot computer is overheating and, in some cases, failing outright since the automaker widened its release of FSD v14 Lite in July. The reports arrive at an unusually bad time for Tesla, which is already facing pressure either to replace or compensate the owners of roughly four million HW3 vehicles after admitting the hardware cannot support the unsupervised self-driving it was sold with the promise of eventually running.

The reports were bubbling up on social media for days before eventually catching the attention of a number of industry outlets, including Electrek. Owners have posted service-mode screenshots showing the Autopilot ECU triggering an APP_w141_ECU_Thermal_Issue fault once the board exceeds 90 degrees Celsius, with some measuring temperatures as high as 96-99 degrees. Very few chips are built to handle temperatures in this range; HW3 is categorically not one of them. 

On the road these errors show up as a forced “red hands” disengagement and a “Self Driving Unavailable” message until the computer cools. In more serious cases, however, the computer has reportedly failed outright and required replacement. Vehicles sold between April 2019 through early-to-mid 2023 contain HW3; many older Teslas dating back to around 2016 were also retrofitted to accommodate the system.

FSD v14 Lite was pushed to wide release on 21 July and functions as a pared-back version of the full v14 model built for HW4 vehicles, albeit heavily compressed to run on HW3’s far more limited compute. HW3 carries just 8 GB of RAM and roughly one-eighth of HW4’s memory bandwidth, constraints that kept Tesla running the older v12.6 branch on these cars for more than a year before Lite arrived.

It should be noted that the specific thermal fault isn’t new and has historically been traced to cooling-system issues like low coolant or a clogged radiator rather than software. Tesla has not confirmed FSD v14 Lite is responsible for the reported failures. But the clustering of reports around the rollout, alongside other v14 Lite regressions like phantom braking and hesitation that point to the same compute ceiling, makes the timing difficult to dismiss as coincidence.

The development, if true, feeds into a growing narrative about the limited capabilities of HW3. In a rare moment of contrition, Chief Executive Elon Musk confirmed back in January 2025 that HW3 cannot achieve unsupervised FSD at all, citing memory bandwidth as the primary constraint. This came after years of promises, including those contained in customer contracts, that purchases of the FSD system would eventually benefit from fully-autonomous driving in the vehicle for which it had been purchased. Many of those vehicles ran HW3, and some of their owners have subsequently taken legal action against the automaker. 

Tesla has floated multiple band-aids for this problem, including discounted trade-ins toward AI4 vehicles and hardware retrofits, but has not committed to a clear resolution for existing owners. If a software update is now measurably shortening the lifespan of the very computers for which Tesla already owes a fix, many now out of warranty, then a software rollout turns into liability problem at a seven-figure scale.

While FSD v14 Lite may be pushing older hardware to, and perhaps beyond, its limits, there is clearly a trade-off at play. Independent testing has found that the update cuts obstacle reaction time by roughly 37% versus the older v12.6.4 branch released two years prior. This denotes a genuine safety improvement for HW3 owners, even though it remains fully one-third slower than the full v14 model running on HW4. Of course, that improvement appears to come from asking more of hardware already operating close to its thermal and memory limits.

Ultimately this territory is not entirely new for Tesla, but rather a situation its compute strategy keeps generating. HW2, HW2.5 and HW3 have each been retired in turn as the company’s growing neural network models outpaced what the hardware could run. With its next-generation AI5 chip already taped out for a mid-2027 rollout at several times HW4’s compute budget, the same obsolescence cycle now straining HW3 owners is one HW4 owners should expect to face eventually too.