Tesla ready to add FSD HW5 to 85,000 Nvidia H100 chips running AI compute for CyberCab service

Tesla will be adding 50,000 Nvidia H100 AI chips to the existing 35,000 units at its vehicle autonomy data centers by the end of the year.

It aims to improve FSD as fast as possible and present its safety record to regulators who'd have to approve both its standalone Robotaxi, and the upcoming CyberCab ride-hailing service.

If the Robotaxi and a cheaper car are what Tesla refers to as the "affordable models," they will launch in Q4 of 2024 or Q1 of 2025, as per Elon Musk, and will both ship with Hardware 4.0 FSD kits still.

There were rumors that the eventual Model 2 that was supposed to come in H2 2025 may land Hardware 5.0 with 3nm chips as an option. Elon now says that HW5 is almost ready as a kit, yet Tesla will start shipping cars with it towards the end of 2025.

In any case, Tesla is preparing to use the downtime of its autonomous driving computers like those in the FSD HW4 or HW5 kits to run AI inference calculations in addition to its Nvidia H100-equipped centers:

It's pretty hard to put together 100 gigawatts of AI compute. Even in an autonomous future where the car is, perhaps, used instead of 10 hours a week 50 hours a week, that still leaves over 100 hours a week when the car inference computer could be doing something else.

Tesla is evidently all in on vehicle autonomy, and is exploring avenues to monetize the strategy. Besides the "purpose-built" Robotaxi that it will announce on August 8, Elon Musk said that it will launch a CyberCab hailing service of sorts.

In it, Tesla will have a fleet of autonomous vehicles that it owns and operates, perhaps said Robotaxis. Any Tesla owner, however, will be able to pick whether to rent out their car as part of the CyberCab fleet. Elon compared it to Airbnb on wheels doing Uber-style runs, while Tesla even showed what a CyberCab ride-hailing app UI will look like.

It is also in talks with a major automaker to license FSD, and it will provide not only the software, but the HW4 or HW5 computer and camera kits, too. That will be an awful lot of computing power on the roads that could run inference calculations to improve FSD in a feedback loop in the vehicle's downtime.

So much so that some predict that self-driving vehicle computers may at one point produce emissions to rival all the world's data centers.

Before this can become reality, though, Tesla will have to face regulators and convince them that FSD is a safer driver than humans. This somewhat explains the investment in AI computing power and the lowering  of the price barrier to log FSD miles for millions of Tesla users.