Tesla Published the FSD Denominator. The Per-Subscriber Crash Rate Is Rising.

Last week, a colleague of mine wrote six hundred words demanding that Tesla publish the denominator behind its ADAS crash record.^[1] On Tuesday evening, Tesla released its Q2 2026 earnings report and, buried between a SpaceX equity gain and a robotaxi mileage claim, disclosed the number: 1.48 million active Full Self-Driving subscriptions, up 56% year over year.^[2] We now have both sides of a fraction that nobody outside the company could compute before. The result does not flatter the curve.

The numerator comes from NHTSA's Standing General Order data, compiled and deduplicated by Electrek: 826 Tesla ADAS crashes in the first half of 2026, a 73% increase over the 476 reported in H1 2025.^[1] The denominator comes from Tesla's own earnings disclosure: 1.48 million FSD subscriptions at the end of Q2 2026. Dividing 1.48 million by 1.56 yields approximately 949,000 subscribers at the end of Q2 2025.^[2] Now, 826 divided by 1,480,000 gives a per-subscriber crash rate of 0.558 per thousand for H1 2026. The same calculation against H1 2025 gives 476 divided by 949,000, or 0.502 per thousand. That is an 11% increase, and because the subscriber count is an end-of-quarter snapshot rather than a period average, the denominator is biased upward, meaning the true per-subscriber rate during H1 is likely higher than 0.558, not lower.

Crashes grew 73% while subscribers grew 56%, a seventeen-percentage-point gap that runs in the wrong direction. Roughly one severe ADAS crash for every 1,792 active FSD subscribers over six months.

There is a fair defense of this number, and it needs to be stated without hedging. The SGO data captures all Tesla ADAS crashes, not just FSD; some fraction involved basic Autopilot, and Tesla redacts the software version on every report, so nobody can separate the two. Using FSD subscribers as the denominator undercounts the ADAS-active user base, which inflates the per-subscriber rate above its true value. But that objection cuts against the trend only if basic Autopilot's share of crashes is growing faster than FSD's, and Tesla has been migrating all new vehicles onto FSD while discontinuing standalone Autopilot. The subscriber count is converging toward the full ADAS-active fleet, not diverging from it. Moreover, Tesla once published quarterly safety reports claiming Autopilot-engaged driving produced fewer crashes per mile than manual driving; those reports were criticized for opaque methodology, never independently audited, and eventually discontinued.^[1] If the per-mile rate were genuinely falling, publishing it would be the easiest PR win Tesla ever had.

A subtler objection: not all subscribers use the system equally. Someone who activates FSD for three highway miles a week generates far less crash exposure than someone running it through downtown San Francisco for an hour every morning. If average engagement per subscriber rose in H1 2026, that alone could explain the rate increase without any software degradation. A valid hypothesis, and a completely uncheckable one, because Tesla does not publish per-subscriber engagement miles. Research on Level 2 driver-assist systems has consistently found that automation complacency increases with system capability and reliability, which would predict higher per-subscriber engagement as FSD improves, but which also predicts more distracted driving during that engagement. The exposure and the risk scale together.

Geographic expansion matters too. FSD launched in South Korea, the Netherlands, and four other European countries during this period, with customers driving over 50 million kilometers by July.^[2] More complex driving environments should produce more incidents. But the SGO only covers US-reported crashes, so international expansion does not directly inflate the H1 numerator. And 826 is itself a floor: Electrek found that the latest NHTSA export revised May's count upward by 139 crashes, meaning reporting lag will push the final H1 figure higher still.^[1]

One more caveat that deserves daylight: an 11% change in a rate built from 826 events against a 1.48 million subscriber base is directionally meaningful but modest enough that a single anomalous quarter could reverse it. This is not a finding that should be carved in stone. It is a finding that should be checked again when Q3 data arrives.

All of those concessions are real, and none of them reverse the core arithmetic.

Now consider what Tesla chose to put in the same earnings report. Ashok Elluswamy, VP of AI Software, described the robotaxi fleet as having completed 380,000 miles with "zero notable incidents," excluding crashes where other vehicles struck stationary Teslas.^[3] That fleet consists of roughly 17 to 25 vehicles operating in geofenced corridors across six Texas and Florida cities.^[3] Meanwhile, 1.48 million humans paying for the same company's software in their personal cars produced 207 crashes in May alone. Two populations, same neural network, and the difference between them is not the technology. It is us. The robots in the geofence do not get bored, do not check their phones, and do not confuse a brand name with a capability. We do all three, and research on automation complacency suggests we do them more, not less, as the system gets better at its job.

The earnings call celebrated 1.48 million FSD subscribers as a growth story, an adoption curve accelerating toward software-margin dominance. Electrek published the 826-crash figure from NHTSA's own dataset the previous day. Neither number was secret; both were already public. Nobody had combined them before because the subscriber figure did not exist outside Tesla until Tuesday evening.

Now it does. If FSD subscriptions keep growing at 56% and ADAS crashes keep growing at 73%, the per-subscriber rate does not converge toward zero; it drifts upward while the absolute count accelerates. That is not a claim about whether FSD is safe or unsafe per mile driven. It is the relationship between two numbers that Tesla chose to disclose twenty-four hours apart, and the relationship says that scaling a driver-assist system to 1.48 million subscribers has not, so far, produced a proportional decline in serious crashes.

What you should do: Before you do anything else, acknowledge the paradox. FARS data covering a full decade of fatal crashes ranks Tesla models among the safest vehicles on American roads by fatality rate.^[4] The Model Y produces 0.03 deaths per 100 million miles traveled; the median sedan is above 1.0. These are cars that, by the measure that counts bodies, kill fewer people per mile than almost anything else on the road. And yet the ADAS crash rate per subscriber is climbing, not falling, as the system scales. Both of those facts are true. If you use FSD, treat it as what it legally remains: a Level 2 assist requiring your full attention. The behavioral research says the better it works, the less attention you will pay. Fight that instinct, because the data says the instinct is winning. Check your VIN for open recalls at nhtsa.gov/recalls.