Tesla HW3 vs HW4 FSD debate resurfaces
- Tesla’s HW3-versus-HW4 fight flared back up after Elon Musk said on April 22 that HW3 cannot reach unsupervised FSD, reversing years of messaging. - The key detail is the bottleneck: Musk said HW3 has just one-eighth of HW4’s memory bandwidth, and Tesla says cameras must change too. - That turns FSD into a hardware-class issue, with resale, retrofit logistics, and robotaxi eligibility now splitting Tesla’s installed base.
Tesla’s self-driving fight is no longer just about software. It’s about which computer is inside the car. That’s why the HW3 versus HW4 argument came roaring back after Tesla’s April 22 earnings call, when Elon Musk finally said the quiet part out loud: Hardware 3 cannot achieve unsupervised FSD. That matters because Tesla spent years selling the idea that cars built since 2019 already had what they needed for full autonomy. ### What are HW3 and HW4, exactly? They’re Tesla’s onboard self-driving computers — plus the sensor stack wrapped around them. HW3 is the older “FSD computer” that replaced HW2.5 in many cars. HW4, also called AI4, is the newer generation shipping in newer vehicles. Tesla’s support pages still frame FSD as a supervised system today, not an autonomous one, and they note that available features can vary by hardware and vehicle configuration. (news.alphastreet.com) ### What changed this time? The big change is that Tesla stopped implying HW3 would eventually get there and started saying it won’t. On the Q1 2026 call, Musk said HW3 “simply does not have the capability” for unsupervised FSD, and tied the limit to memory bandwidth. That’s the formal break with the old all-cars-are-ready story. ### Why does memory bandwidth matter so much? (tesla.com) Because self-driving is not just “run the model.” The car has to ingest camera feeds, move data around fast, keep multiple neural nets alive, and do all of it with low latency. Musk’s shorthand was that HW3 has only one-eighth the memory bandwidth of HW4. Basically, this is like having a decent brain attached to a tiny desk — the problem is not only compute, but how fast the system can feed that compute with fresh information. (news.alphastreet.com) ### So is HW3 dead? Not for supervised driving. Tesla still sells and supports FSD (Supervised), and Ashok Elluswamy said HW3 would get a “distilled” v14 build intended to preserve most of the major feature set from the HW4 branch. TeslaNorth’s recap says that version was targeted for the end of June. So the split is not “HW3 works” versus “HW3 is useless.” It’s supervised usefulness versus unsupervised ceiling. (news.alphastreet.com) ### Why are owners so worked up? Because this hits three things at once — promises, money, and timing. If you paid for FSD years ago, you were not buying a lane-keeping assist package. You were buying into a future capability story. Once Tesla says one hardware generation can’t cross the final line, owners immediately start asking whether their car just became the wrong platform, whether resale will take a hit, and whether Tesla owes them a real path forward. (tesla.com) ### Can Tesla just swap the computer? Turns out, no — not cleanly. Musk said moving a HW3 car to HW4 means replacing the computer and the cameras. That makes this a vehicle retrofit problem, not a laptop-style chip upgrade. Tesla has floated “micro-factories” in major metros because normal service centers would be too slow for the scale of the job. (teslanorth.com) ### Does that mean HW4 owners are safe? Safer, yes. Future-proof, maybe not. Tesla already introduced HW4+ with doubled RAM on April 23, one day after the HW3 admission. That doesn’t mean HW4 suddenly becomes obsolete, but it does remind owners that Tesla’s autonomy stack is moving like consumer electronics, not like traditional car platforms. ### What’s the real takeaway? (teslanorth.com) The debate resurfaced because Tesla finally turned a fuzzy suspicion into a hard boundary. HW3 still matters if you want usable supervised FSD. But HW4 — and now HW4+ — is where Tesla is drawing the line for the unsupervised future. For owners, that changes the question from “Will my car improve?” to “Which class of Tesla do I actually own?” (news.alphastreet.com) (electrek.co)