Martin Viecha
The third question is, what is the current assessment of the pathway toward regulatory approval for unsupervised FSD in the U.S.? And how should we think about the appropriate safety threshold compared to human drivers?
Lars Moravy -- Vice President, Vehicle Engineering
I can start. There are a handful of states that already have adopted autonomous vehicle laws. These states are paving the way for operations, while the data for such operations guides a broader adoption of driverless vehicles. I think Ashok can talk a little bit about our safety methodology, but we expect that these states and the work ongoing as well as the data that we're providing will pave a way for a broad-based regulatory approval in the U.S. at least and then in other countries as well.
Elon Musk -- Chief Executive Officer and Product Architect
Yeah. It's actually been pretty helpful that the autonomous car companies have been cutting a path through the regulatory jungle. So, that's actually quite helpful. And they have obviously been operating in San Francisco for a while.
I think they got approval for City of L.A.. So, these approvals are happening rapidly. I think if you've got at scale, a statistically significant amount of data that shows conclusively that the autonomous car has, let's say, half the accident rate of a human-driven car, I think that's difficult to ignore because at that point, stopping autonomy means killing people. So, I actually do not think that there will be significant regulatory barriers, provided there is conclusive data that the autonomous car is safer than a human-driven car.
And in my view, this will be much like elevators. Elevators used to be operated by a guy with relay switch. But sometimes, the guy would get tired or drunk or just make a mistake and shear somebody in half between floors. So, we just get an elevator and press button, we don't think about it.
In fact, it's kind of weird if somebody is standing there with a relay switch. And that will be how cars work. You just summon the car using your phone. You get in, it takes you to a destination. You get out.
Vaibhav Taneja -- Chief Financial Officer
You don't even think about it.
Elon Musk -- Chief Executive Officer and Product Architect
You don't even think about it, just like an elevator. It takes you to your floor. That's it. Don't think about how the elevator is working or anything like that.
And something I should clarify is that Tesla will be operating the fleet. So, you can think of like how Tesla, think of it as combination of Airbnb and Uber meaning that there will be some number of cars that Tesla owns itself and operates in the fleet. There will be some number of cars and then there'll be a bunch of cars where they're owned by the end user. That end user can add or subtract their car to the fleet whenever they want, and they can decide if they want to only let the car be used by friends and family or only buy five-star users or by anyone at any time they could have the car come back to them and be exclusively theirs like an Airbnb.
You could rent out your guest room or not, any time you want. So, as our fleet grows, we have 7 million cars going -- 9 million cars going to eventually tens of millions of cars worldwide. With a constant feedback loop, every time something goes wrong, that gets added to the training data and you get this training flywheel happening in the same way that Google Search has the sort of flywheel, it's very difficult to compete with Google because people are constantly doing searches and clicking, and Google is getting that feedback loop. So, the same with Tesla.
But at the scale that is maybe difficult to comprehend, but ultimately, it will be tens of millions. I think there's also some potential here for an AWS element down the road where if we've got very powerful inference because we've got a Hardware 3 in the cars, but now all cars are being made with Hardware 4. Hardware 5 is pretty much designed and should be in cars, hopefully toward the end of next year. And there's a potential to run -- when the car is not moving to actually run distributed inference.
So, kind of like AWS, but distributed inference. Like it takes a lot of computers to train an AI model, but many orders of magnitude less compute to run it. So, if you can imagine a future, perhaps where there's a fleet of 100 million Teslas, and on average, they've got like maybe a kilowatt of inference compute. That's 100 gigawatts of inference compute distributed all around the world.
It's pretty hard to put together 100 gigawatts of AI compute. And even in an autonomous future where the car is, perhaps, used instead of being used 10 hours a week, it is used 50 hours a week. That still leaves over 100 hours a week where the car inference computer could be doing something else. And it seems like it will be a waste not to use it.
Martin Viecha
Ashok, do you want to chime in on the process and safety?
Ashok Elluswamy -- Director, Autopilot Software
Yeah. We have multiple tiers of validating the safety for in any given week, we train hundreds of neural networks that can produce different trajectories for how to drive the car, replay them through the millions of clips that we have already collected from our users and our own QA. Those are like critical events, like someone jumping out in front or like other critical events that we have gathered database over many, many years, and we replay through all of them to make sure that we are net improving safety. We have simulation systems that also try to create this and test this in close loop fashion. And some of this is validated, we give it to our own QA networks.
We have hundreds of them (QA networks) in different cities, in San Francisco, Los Angeles, Austin, New York, a lot of different locations. They are also driving this and collecting real-world miles, and we have an estimate of what are the critical events, are they net improvement compared to the previous-week builds. And once we have confidence that the build is a net improvement, then we start shipping to early users, like 2,000 employees initially that they would like it to build, they will give feedback on like if it's an improvement there or they're noting some new issues that we did not capture in our own QA process. And only after all of this is validated, then we go to external customers.
And even when we go external, we have like live dashboards of monitoring every critical event that's happening in the fleet sorted by the criticality of it. So, we are having a constant pulse on the build quality and the safety improvement along the way. And then any failures like Elon alluded to, we get the data back, add it to the training and that improves the model in the next cycle. So, we have this like constant feedback loop of issues, fixes, evaluations and then rinse and repeat.
And especially with the new V12 architecture, all of this is automatically improving without requiring much engineering interventions in the sense that engineers don't have to be creative and like how they code the algorithms. It's mostly learning on its own based on data. So, you see that, OK, every failure or like this is how a person chooses is how you drive the intersection or something like that, they get the data back. We add it to the neural network, and it learns from that trained data automatically instead of some engineers saying that, oh, here, you must rotate the steering wheel by this much or something like that.
There's no hard inference conditions. Everything is neural network. It's pretty soft. It's probabilistic, so it will adapt probabilistic distribution based on the new data that it's getting.
Elon Musk -- Chief Executive Officer and Product Architect
Yeah. And we do have some insight into how good the things will be in like, let's say, three or four months because we have advanced models that are far more capable than what is in the car, but have some issues with them that we need to fix. So, they are like, there'll be a step change improvement in the capabilities of the car, but it will have some quirks that are -- that need to be addressed in order to release it. As Ashok was saying, we have to be very careful in what we release to the fleet or to customers in general.
So, if we look at, say, 12.4 and 12.5, which are really -- could arguably even be Version 13, Version 14 because it's pretty close to a total retrain of the neural nets in each case, substantially different. So, we have good insight into where the model is, how well the car will perform in, say, three or four months.
Ashok Elluswamy -- Director, Autopilot Software
Yeah. In terms of scaling loss, people in the community generally talk about model scaling loss where they increase the model size a lot and then they have corresponding gains in performance, but we have also figured out scaling loss and other access in addition to the model side scaling, making also data scaling. You can increase the amount of data you use to train the neural network and that also gives similar gains and you can also scale up by training compute, you can train it for much longer and on more GPUs or more dojo nodes that also gives better performance, and you can also have architecture scaling where you count with better architectures for the same amount of compute produce better results. So, a combination of model size scaling, data scaling, training compute scaling and the architecture scaling, we can basically extrapolate, OK, with the continue scaling based at this ratio, we can predict future performance.
Obviously, it takes time to do the experiments because it takes a few weeks to train, it takes a few weeks to collect tens of millions of video clips and process all of them, but you can estimate what is going to be the future progress based on the trends that we have seen in the past, and they're generally held true based on past data.
