Tesla Autopilot: Building A Network Effect Around Safety

Last week Tesla released their Q3 2019 earnings report, and analysts have covered profits, vehicle margins, and Q4 delivery forecasts at length. But one of the most overlooked results from Tesla’s Q3 report was their vehicle safety numbers. And more specifically, the safety of the Tesla Autopilot system. 

I found it odd that no analysts on the Tesla Q3 conference call had any questions about Tesla’s updated safety data. Especially considering the large improvement in Tesla’s numbers and the importance of safety for anyone buying a new car.

So here, I make a case for safety being the most important metric for Tesla going forward, as they start building a valuable network effect centered around vehicle safety. Over time, I believe OEMs that don’t use a complete set of active safety hardware to constantly improve their vehicles will face serious challenges in attracting new customers.

Tesla’s Vehicle Safety Data

In Q3 2018, Tesla issued their first ever safety report, highlighting crash data from their entire vehicle fleet under a number of driving modes. Below are the initial report findings.

On Autopilot: One crash every 3.34 million miles

With active safety: One crash every 1.92 million miles

Without Autopilot or active safety: One crash every 2.02 million miles

US Average (NHTSA data): One crash every 492,000 miles

The obvious conclusion from these findings is that Tesla cars are safer than the average US vehicle, particularly when driving on Autopilot. Tesla has been advertising this long before their Q3 2018 safety report. But the Q3 2018 data isn’t conclusive in isolation. Critics have pointed out three main arguments against the significance of Tesla’s safety data.

#1: Tesla only had a fleet size of roughly 400,000 cars in Q3 2018. That is too small of a sample size for accurate data.

#2: Tesla builds luxury cars most likely driven by old, wealthy men. That demographic is less likely to get into crashes.

#3: Autopilot is only used on highways, crashes are less likely to occur there.

Luckily, Tesla has been releasing their safety data every single quarter since Q3 2018, and we now have a perfect set of YoY data to compare. Seasonal trends in road safety (snow, ice, rain, etc…) make comparisons of Tesla’s Q4, Q1, and Q2 data against their original data difficult, but each passing quarter we’ll get a clearer picture of the trends forming in Tesla vehicle safety.

Tesla Autopilot Results: Q3 2018 vs. Q3 2019

Q3 2018

Tesla entered Q3 2018 having delivered roughly 360,000 cars, and ended Q3 2018 with a fleet size of almost 450,000 cars, but almost 50,000 of the overall fleet wasn’t equipped with the hardware required for Autopilot. So the sample size for Autopilot/active safety accident data in 2018 came from about 350,000 vehicles.

If we assume that the average Tesla drove 3,369 miles in the quarter, Autopilot-enabled Tesla vehicles drove roughly 1.18 billion miles in Q3 2018. And with an estimated crash rate of 1 every 3.34 million miles on Autopilot, the number of Tesla vehicles that actually crashed on Autopilot during the quarter couldn’t have been higher than 353.

Of course, that assumes every Tesla drivers used Autopilot for 100% of their driving, which is obviously wrong. Based on Lex Fridman’s 2018 Autopilot data, roughly 11.1% of miles driven in Teslas were on Autopilot in 2018. So the actual number of Tesla crashes on Autopilot during Q3 2018 is more like 39, instead of 353.

Q3 2019

Fast forward to Q3 2019, and Tesla’s fleet size has doubled. Tesla started Q3 2019 with almost 700,000 cars in its fleet, and ended the quarter with almost 800,000. But when we remove the 50,000 non-Autopilot cars, Tesla’s Autopilot-enabled fleet size drops to an average of 700,000 cars in Q3.

Using the same mileage assumptions above, that means Tesla cars drove 2.36 billion miles in Q3 2019. However, Teslas on Autopilot only registered accidents once every 4.34 million miles during the quarter. An impressive feat considering overall fleet mileage doubled, meaning Tesla’s actual results are likely closer to the mean expected crash rate than in 2018.

Based on Lex Fridman’s latest projections, Autopilot miles accounted for 12.9% of all Tesla miles driven since Nov 2018. That means there were roughly 70 Tesla crashes on Autopilot during Q3 2019. But if no improvements had been made to Tesla’s Autopilot this year, there would have been about 91.

Every Tesla critic loves to focus on cars that do crash, but none of them consider the 21 Teslas that were saved from crashes by Autopilot this quarter.

Answering The Tesla Bears

Now that we understand the facts of Tesla’s Autopilot crash data, we can make a rebuttal against the three main criticisms of Tesla’s original data.

Criticism #1: “Tesla only had a fleet size of roughly 400,000 cars in Q3 2018. That is too small of a sample size for accurate data.”

Rebuttal #1: Tesla’s Autopilot-enabled fleet size has now doubled, and safety has meaningfully improved for both Autopilot and active safety driving modes.

Criticism #2: Tesla builds luxury cars most likely driven by old, wealthy men. That demographic is less likely to get into crashes.

Rebuttal #2: 82% of all Tesla sales in Q3 2019 were for the Model 3. The mass-market Model 3 attracts a more diverse and representative set of customers, and safety meaningfully improved across Autopilot and active safety driving modes.

Note: The above criticism looks like it was probably correct for non-Autopilot/active safety driving last year. Without any active safety or Autopilot features, crash rates went up by 10% this quarter. This probably reflects the more diverse/risky set of drivers that bought a Model 3 in Q4 2018 and 2019.

I also expect the non-active safety figures will likely normalize around their current levels. Tesla’s feed of fleet data is far less helpful here. The factors that could lead to future improvements are things like improved braking power, wheel traction, and overall car agility. I expect these improvements to be far smaller moving forward compared to active safety/Autopilot improvements.

Criticism #3: Autopilot is only used on highways, crashes are less likely to occur there.

Rebuttal #3: Between Oct 2016 and Nov 2018, Autopilot miles accounted for only 11.1% of all Tesla miles. But based on Lex Fridman’s latest projections, Autopilot miles account for 12.9% of all Tesla miles driven since Nov 2018. Tesla cars are using Autopilot more often, and safety still meaningfully improved across Autopilot/active safety driving modes.

So to summarize: There are more Teslas with a more diverse set of drivers, using Autopilot on a more diverse set of roads, and Tesla still blew away all of last year’s (already impressive) safety benchmarks. So what does this all mean for the company?

Safety: Tesla’s Ultimate Network Effect

Tesla Autopilot Safety

A Tesla is 30% to 41% less likely to crash in Q3 2019  vs. Q3 2018 while using Autopilot or active safety features. The average car in America? Only 1% safer this year. And the rich data being collected by every Tesla on the road (and not being collected by every other OEM) explains the massive gap in improvement.  This data captures all sorts of crashes, near-misses, and driving anomalies that Autopilot and Tesla’s active safety features can learn from to avoid crashes.

All the data comes from the cameras, radar, and ultrasonic sensors installed on every Tesla today. A suite of sensors that no other automaker has matched to date. And it’s a problem for all other automakers because of how quickly Tesla’s safety is improving.

Because once Tesla starts stacking 30-41% improvements on-top of each other, why would anyone not buy a Tesla? Especially if every other automaker is improving safety by only 1% each year.

A Nasty Positive Feedback Loop

And what happens when Tesla starts pulling new car buyers away from other OEMs because of their superior safety? Those new customers feed even more data into Tesla’s Autopilot system, further widening the gap between Tesla’s safety and all other cars.

Just like Google’s superior search algorithms swayed people to use it over competitors, giving Google even better data to further optimize their algorithms and pull even more users away from competitors, Tesla seems poised to give other automakers the Google treatment. It’s a nasty positive feedback loop.

While search quality may be more important for search engines than safety is for automakers, safety is still damn important. Time will tell exactly how important safety is, but other automakers would be foolish to wait and find out.

Now I don’t want to count out other automakers just yet. Many of them still have larger fleet sizes than Tesla today. But if they don’t make an urgent push to start collecting data to meaningfully improve vehicle safety, they’re giving away the brand and product leadership they’ve worked for decades to earn.

Because when it comes to innovation, a company’s rate of improvement is more important than their starting position.

And right now Tesla is improving faster than everyone else.