Tesla revealed a lot about all of their existing and planned artificial intelligence related products and services in the Q1 2024 earnings call.
Elon talked about offering FSD/robotaxi as a hybrid Airbnb and Uber model. Tesla will operating the fleet and the network. The enduser will be able to control their participation in the robotaxi network. There had been some people talking about Tesla only operating its own fleet of robotaxi but this is not the case.
Tesla will create an Amazon Web services (AWS) like distributed inference service. If Tesla has 100 million cars eventually out in the world, then this would be a 100 Gigawatt of distributed compute. If Tesla could make $300/year could made per car then this would be $3 billion per year.
Tesla will have its Optimus Teslabots working in their own factories by the end of this year (2024). Teslabot could be sold externally by the end of 2025.
Tesla has been testing FSD (full self driving) 12.4 and 12.5 for 3-6 months and know how good the FSD is. Tesla can forecast fairly accurately the future performance of FSD based upon the scaling of data and compute.
Tesla 12.4 and 12.5 are each total retraining of the neural network.
FSD version 12.X is going to China.
There has been over 50% usage of FSD of potential Tesla FSD vehicles (Over 900k of 1.8 million). IF half of those who try it (and not including those who purchased) then this could be an increase of 200,000 to 300,000 subscriptions in the USA and would be 300,000 to 500,000 in China if China adopts at comparables levels to the USA.
Tesla can improve the FSD performance by
1. increasing model size
2. training compute times
3. more data and
4. better architectures
Tesla is in discussions with one other major automaker about licensing FSD. Elon thinks FSD will have at least one license this year. It will take three years for any new automaker to deploy Tesla licensed hardware and software in a new vehicle.
Tesla has 35,000 Nvidia H100s at the end of March and will have over 85,000 by the end of 2024. This is 140 Exaflops at the end of March and would be 340 Exaflops at the end of the year.
The FSD milestones mentioned align with what Nextbigfuture has been predicting for over a month.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
100M cars X $300 per year each for distributed inference compute is $30B per year in revenue not $3B.
Tesla could provide some internal incentives to opt in like discounts on Superchargers or Tesla Network fees or other perks that would not add much to Tesla’s costs. That revenue would be mostly profit.
100M FSD vehicles on the road is not unrealistic cumulative if they hit the 20M per year target.
Why does Elon talk about 1 kW of compute for each car? Their fsd computer has a max of 50W today.
Also, by the time Tesla has 100 million cars on the street, the oldest of these cars will be more than 10 years old. So do these old computers have a competitive number of flops per W? If not, running them would be much more expensive than running the same workload on a modern server…
Come on younshould be talking about the elephant in the room. Tesla continued loss of revenue. The car has substandard quality. The EV industry has lost its direction.
question. What do you mean by substandard quality.
I purchased a 2024 Model Y Long Range in January. We’ve used it heavily for trips between Seattle and Portland to help family with medical issues, and I completed a 5600 mile/10 day road trip from Seattle to San Antonio/Houston/Starbase and back.
I’ve not encountered any substandard quality.
Sure, the car has less physical buttons and levers on the drives console, but the software was super responsive.
I used Full Self Driving for a lot of the trip, and it performed outstandingly. Especially navigating cities that I have never visited.
Tesla does have a direction. It’s laid out in their three ‘Master Plans’ they have posted for anyone to read.
https://www.tesla.com/blog/secret-tesla-motors-master-plan-just-between-you-and-me
https://www.tesla.com/blog/master-plan-part-deux
https://www.tesla.com/ns_videos/Tesla-Master-Plan-Part-3.pdf
The rest of the industry are the ones losing their way with EVs as they are chickening out over the cost to retool and redevelop from their ICE vehicle centric structure. Tesla took a decade to get to profitibility, having to put every dollar earned, and many rounds of investment into manufacturing and development.
If you are in the Seattle area, let me know and we will do a few hours of driving around and you can then make an informed decision about how awful Tesla EVs are.
In your list I don’t see level 5 in 2024/2025. Scared ?
SAE levels are meaningless to Tesla. They’re designed around the consensus LiDAR and Hidef maps approach that failed and that Tesla doesn’t use. FSD NN can like human drivers drive anywhere with varying quality based on training with local rules and practices, it isn’t inherently geofenced which is what SAE L4 to L5 is about.
Can the Tesla robot roughneck on a drill rig. The task is as simple and repetitive.
When did regulators approve driverless cars? Did we all miss something? This app is simply a publicity stunt and not a real product that will add anything to Tesla’s profits. The app would have to out compete the well known Uber and Lyft apps. Surely Elon knows that the moment, many years from now, that Federal, State and Local governments sporadically approve driverless vehicles to autonomously roam their jurisdiction looking for fares, Uber and Lyft will add a button to their app titled “Rent a self-driving Tesla”. Of course none of this will happen for several years. The estimate of 100 million working Teslas on the road with a 100% buy in for some distributed form of AWS is very difficult to understand for a few reasons.
1. Toyota sells 10 million cars a year including the nearly indestructible Hilux pick-up truck. They are estimated to have 100 million working cars on the road. Tesla sells 1.6 million cars a year, and not a single vehicle they have ever sold will last, even with maintenance, as long as a mildly maintained Toyota pick up truck. Tesla would have to 5x sales for a decade to match Toyota’s 100 million operating road worthy vehicles.
2. How is the data transferred to and from the vehicle? Wifi when its garaged somewhere? Don’t think Starlink, because even the new generation sats don’t have the bandwidth for a distributed inference service.
3. If it is possible to create a distributed inference service using GPU based computers, why hasn’t or won’t NVIDIA do the same thing with GPU’s sitting in 100s of millions of gaming desktops/laptops around the globe? My guess is that there are impossible latency issues to overcome, not to mention occasional destruction of someones real valuable data from a variety of accidents. A Cybertruck burns down and destroys several families distributed digital photo albums with it on a good day and a bunch of hospital medical records on a bad day? Do you want to store your business records on Kid Rock’s Model S plaid?