Tesla shuts down Dojo supercomputer project, will focus on new AI chip development

[DragonSlayer101]

What just happened? Tesla has scrapped its ambitious Dojo supercomputer project, which was designed to train the company's full self-driving neural networks. The decision marks a surprising change of course for Elon Musk, who had publicly touted Dojo as a key component in achieving Tesla's autonomous driving goals.

Dojo project head Peter Bannon is reportedly leaving Tesla, along with several key engineers and chip designers, to join a startup called DensityAI. The new company was founded last year by former Dojo lead Ganesh Venkataramanan and ex-Tesla employees Bill Chang and Ben Floering.

Roughly 20 members of the Dojo team – including some senior-level executives – have reportedly joined DensityAI, along with engineers from across the tech industry. The startup is developing both hardware and software to power AI data centers focused on robotics and automotive applications.

Musk later confirmed the report, stating that Tesla will shift its resources toward scaling its AI5 and AI6 chips rather than pursuing two separate AI chip designs. Musk said in a post on X that the current chips, along with their successors, will be "excellent for inference and at least pretty good for training."

It doesn't make sense for Tesla to divide its resources and scale two quite different AI chip designs.

The Tesla AI5, AI6 and subsequent chips will be excellent for inference and at least pretty good for training. All effort is focused on that.

– Elon Musk (@elonmusk) August 8, 2025

Musk previously stated that Tesla's next-generation AI5 chips will enter manufacturing at the end of 2026, with the AI6 expected to go into production sometime in 2027 or 2028. These chips are intended for use in Tesla's FSD-enabled vehicles and Optimus humanoid robots, though Musk has suggested they could also support broader AI applications.

Tesla's decision to shut down the Dojo project marks a significant strategic shift for the automaker. Musk had championed Dojo for years, describing it as central to Tesla's long-term goal of achieving full self-driving. However, despite his early enthusiasm, there had been growing signs that all was not well with the initiative.

In recent months, Musk had been noticeably quiet about Dojo, offering only a brief mention during Tesla's second-quarter earnings call earlier this year. Speaking to investors at the time, he said he was considering merging the Dojo 3 and AI6 inference chip designs to avoid spreading resources too thin.

With Dojo now shelved, Tesla's AI ambitions will be supported by another in-house supercomputer named Cortex. Currently under construction at the company's Austin data center, Cortex is expected to be powered by more than 100,000 Nvidia H100 and H200 accelerators once complete.

Permalink to story:

Tesla shuts down Dojo supercomputer project, will focus on new AI chip development

 

[seeprime]

I think the people that left knew that Musk's idea of camera only self driving was impossible without lidar and radar. There is no way cameras alone, no matter how many, or how good the software is, cannot have perfect views in all weather conditions, or all hours of the day. Unmarked gravel roads are a challenge also.

 

[Eldrach]

I think the people that left knew that Musk's idea of camera only self driving was impossible without lidar and radar. There is no way cameras alone, no matter how many, or how good the software is, cannot have perfect views in all weather conditions, or all hours of the day. Unmarked gravel roads are a challenge also.

Well, true - but the system can still be used in most weather conditions, it only needs to be good enough to tell the driver it cannot operate in those conditions. Lidar and radar works poorly on unmarked gravel roads and snowy roads aswell btw. No system can handle all conditions atm.

 

[godrilla]

Well, true - but the system can still be used in most weather conditions, it only needs to be good enough to tell the driver it cannot operate in those conditions. Lidar and radar works poorly on unmarked gravel roads and snowy roads aswell btw. No system can handle all conditions atm.

If they control the robotaxi or Tesla taxi they can have them operated in ideal conditions as well.
Imagine your ai taxi tells you to get out or wait for better weather conditions to complete the trip.

 

[Beerfloat]

If they control the robotaxi or Tesla taxi they can have them operated in ideal conditions as well.
Imagine your ai taxi tells you to get out or wait for better weather conditions to complete the trip.

This is a Musk AI. More worried about the taxi deciding to run over random kaffirs.

 

[Theinsanegamer]

I think the people that left knew that Musk's idea of camera only self driving was impossible without lidar and radar. There is no way cameras alone, no matter how many, or how good the software is, cannot have perfect views in all weather conditions, or all hours of the day. Unmarked gravel roads are a challenge also.

But if that were the case, why leave now? The demand for LIDAR less self driving came out what? A decade ago?

Well, true - but the system can still be used in most weather conditions, it only needs to be good enough to tell the driver it cannot operate in those conditions. Lidar and radar works poorly on unmarked gravel roads and snowy roads aswell btw. No system can handle all conditions atm.

Hate to tell you this, but a fully self driving car needs to be able to drive in all weather conditions, not just what they find acceptable in Arizona. Shutting down 5 months of the year anywhere north of the 45th parallel just isnt acceptable.

 

[seeprime]

But if that were the case, why leave now? The demand for LIDAR less self driving came out what? A decade ago?
Hate to tell you this, but a fully self driving car needs to be able to drive in all weather conditions, not just what they find acceptable in Arizona. Shutting down 5 months of the year anywhere north of the 45th parallel just isnt acceptable.

They left now because it was time. We don't know how long ago the new AI endeavor was planned. It certainly wasn't spontaneous.

 

[Eldrach]

But if that were the case, why leave now? The demand for LIDAR less self driving came out what? A decade ago?
Hate to tell you this, but a fully self driving car needs to be able to drive in all weather conditions, not just what they find acceptable in Arizona. Shutting down 5 months of the year anywhere north of the 45th parallel just isnt acceptable.

I don’t think we’ll reach that particular point the next 20 years. Full self driving will be limited to regular roads and will also most likely have a «approved grid» wherever it’s allowed

 

[cachy]

I think the people that left knew that Musk's idea of camera only self driving was impossible without lidar and radar. There is no way cameras alone, no matter how many, or how good the software is, cannot have perfect views in all weather conditions, or all hours of the day. Unmarked gravel roads are a challenge also.

Your comment has nothing to do with the article.

In regards to lidar, do you have lidar in your head or do you drive on vision alone?

 

[cachy]

They left now because it was time. We don't know how long ago the new AI endeavor was planned. It certainly wasn't spontaneous.

They formed a new startup.

 

[cachy]

I don't think most folks understand the goals of AI in general. It does not need to be good at it. It just needs to be as good as a human.

 

[cachy]

I don’t think we’ll reach that particular point the next 20 years. Full self driving will be limited to regular roads and will also most likely have a «approved grid» wherever it’s allowed

You must mean driver-less taxis not full self driving that's already in Teslas because that worked great when I had my free trials. Even the regular cruise works flawlessly on highways. Stays in the middle of the lane and adjusts speed given your preset distance to the car in front. It won't change lanes but on highways I usually just stay in the same lane.

 

[integr8d]

But if that were the case, why leave now? The demand for LIDAR less self driving came out what? A decade ago?
Hate to tell you this, but a fully self driving car needs to be able to drive in all weather conditions, not just what they find acceptable in Arizona. Shutting down 5 months of the year anywhere north of the 45th parallel just isnt acceptable.

That's like saying all planes need to be able to fly in all conditions. And that's insane... There are plenty of times (Intense rain, snow, fog, etc) that human drivers need to pull over and wait for better conditions. And anything short of SWIR cameras will not change that for human or software.

 

[Eldrach]

You must mean driver-less taxis not full self driving that's already in Teslas because that worked great when I had my free trials. Even the regular cruise works flawlessly on highways. Stays in the middle of the lane and adjusts speed given your preset distance to the car in front. It won't change lanes but on highways I usually just stay in the same lane.

You’re using the service in the tiny area of the world where it’s been specifically trained - talking about global coverage. Cruise control works fine

 

[Plutoisaplanet]

You’re using the service in the tiny area of the world where it’s been specifically trained - talking about global coverage. Cruise control works fine

That’s not how Tesla’s FSD works. Tesla owners use it today anywhere in North America. Here’s an example of it working somewhere in YouTube 5 months ago:

Supervision is still required because they’ve been developing it and are now validating its completion one area at a time (while I’m sure making fixes). It operates in Austin, TX and that service area has been expanded twice in the last month alone with a region the size of Waymo’s. It also now operates in a large portion of the SF bay area (far larger than Waymo’s), but California is asking for a trial with someone sitting behind the driver’s seat for regulatory approval.

It’s expanding way faster than specifically trained regions. Otherwise you’d see Waymo operate in way mo’ places.

 

[seeprime]

They formed a new startup.

That was in the article.

 

[seeprime]

Your comment has nothing to do with the article.

In regards to lidar, do you have lidar in your head or do you drive on vision alone?

Lidar and cameras are safer than cameras alone due to lidar penetrating poor visibility conditions. Your comparison of s human is not valid.

 

[Eldrach]

That’s not how Tesla’s FSD works. Tesla owners use it today anywhere in North America. Here’s an example of it working somewhere in YouTube 5 months ago:

Supervision is still required because they’ve been developing it and are now validating its completion one area at a time (while I’m sure making fixes). It operates in Austin, TX and that service area has been expanded twice in the last month alone with a region the size of Waymo’s. It also now operates in a large portion of the SF bay area (far larger than Waymo’s), but California is asking for a trial with someone sitting behind the driver’s seat for regulatory approval.

It’s expanding way faster than specifically trained regions. Otherwise you’d see Waymo operate in way mo’ places.

Well - We are a few years away from global then as those locations you mentioned generally all have sunny bright weather optimal for cameras .
Don’t get me wrong - when it becomes available in my country, I’ll upgrade my Juniper..but..we’ll be seeing 2030 before that happens

 

[brucek]

While "all road, all condition*" full self driving is obviously even more appealing, I think there is still plenty of market demand for reliable full autonomous driving in major market areas under normal vs extreme conditions. Personally I'd rather have more safety under more tightly defined situations, than more willingness to try anything and see what happens. (*Obviously there are some conditions under which no one should be driving, and those apply to AI too.)

Same for the current self-driving taxi rollouts, which come with some pretty tight limits for where they'll go and I think some or all still refuse to even go on the highway. Again even with those limitations that still covers plenty of the trips needed so better to have that than nothing.

 

[Plutoisaplanet]

Well - We are a few years away from global then as those locations you mentioned generally all have sunny bright weather optimal for cameras .
Don’t get me wrong - when it becomes available in my country, I’ll upgrade my Juniper..but..we’ll be seeing 2030 before that happens

It was snowing in Idaho in the video I linked to lmao. You clearly didn’t watch the video. And no, Austin TX gets rain throughout the year, erratic weather, and has a humid subtropical climate. Meanwhile San Francisco is coastal and very unlike what you describe. It gets a lot of fog, is usually overcast, and definitely gets more rain than most of California.

Anyways, let’s clear up this myth about cameras being incapable of allowing self driving. What’s the difference between good cameras (what they have on Teslas) and your own eyes? Would you be incapable of driving if watching video footage? Because it looks easy for me to do. The only reason it wasn’t possible before is image processing and AI capabilities weren’t up to par for the task.

And if you think that there’s not enough dynamic range to cameras compared to our eyes, well I’ve seen Autopilot drive straight into low sunlight with zero issues. It recognizes oncoming cars with the sun right behind them with no problem.

 

[bviktor]

This is the same person that NV CEO Jensen called "superhuman" for the way he builds stuff.