For starters, I estimate there are about 1.4m U.S. Teslas on the road today running HW3, which currently top out at software version v12.6.4, compared to the HW4 at v13.2.9. I recently tested a Model Y running HW3 and, while I found it impressive, it lacked the smoothness and object recognition that HW4 yields.

This is consistent with what Musk mentioned about six months ago on the December 2024 earnings call. In response to the question: “Is it expected that Tesla will need to upgrade Hardware 3 vehicles?” Elon responded: “The honest answer is that we’re going to have to upgrade people’s Hardware 3 computer for those that have bought Full Self-Driving, and that is the honest answer and that’s going to be painful and difficult but we’ll get it done. Now I’m kind of glad that not that many people bought the FSD package.“

I suspect those customers who bought FSD on HW3 will end up getting an $8k-$10k FSD refund instead of a hardware upgrade. I estimate the cost of that refund would be somewhere between $40-$80m, based on my belief that around 10,000 licenses were sold. Tesla has around $37B in cash and investments, so the refund would be immaterial to the company’s financials.

I estimate there are about 1.2m vehicles on the road with HW4 today. The 35 or so vehicles (up from about 15 at launch three weeks ago) that make up the Austin robotaxi fleet are running a modified, geofenced version of the company’s most advanced consumer facing software, version 13.2.9, which was released in May 2025.

HW4 began its rollout in early 2023 and is based on a Samsung 7 nm process. Musk claims it offers 3× to 8× the computing power of HW3. In terms of raw specs, HW4 is roughly 5× faster than HW3 and uses higher-resolution cameras. The hardware stack for Model Y and Model 3 still relies entirely on vision and does not include radar or lidar.

The good news is that HW4 is holding up well in the real world, and so far the robotaxi rollout has been smooth. My concern is that HW4 eventually won’t be enough to power Level 4 or Level 5 autonomy, especially as the pace of neural network improvements continues to accelerate.

All eyes are on the pace of the Robotaxi expansion. Musk has set the bar that the company will make measurable progress this year, ending the year with more than a thousand cars in the fleet in multiple cities, which I believe are realistic milestones. Long term, to compete with Uber, Tesla will need hundreds of thousands of FSD enabled vehicles on the road, a milestone that is increasingly looking like it will require the upcoming HW5.

New Teslas today all ship with HW4, and Musk has teased HW5 will be out early next year and will be about 10× more powerful than the power of HW4. That means the gap between HW4 and HW5 will be twice as wide compared to the gap between HW3 and HW4. Needless to say, HW5 will do a better job than HW4, and in the March of Nines quest, any lead is exponentially more valuable. In other words, HW4 is a huge step up, but it may still fall short of the compute needed for complete self driving, making HW5 the probable requirement to ramp autonomy.

All of this begs the question, when will the race for better hardware end? The answer, of course, is never. I’m trying to answer the question: will HW5 be enough to get us to level 4 or 5 for the first time? My sense is yes. And if I’m wrong, it underscores that hardware improvements beyond HW4 will be needed to move us forward.

Disclaimer