Tesla turns superchargers into AI data centers

**SPEAKER_1** (0:01)
When you need to build up your team to handle the growing chaos at work, use Indeed Sponsored Jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications and more. Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a $75 sponsored job credit at indeed.com/podcast. That's indeed.com/podcast. Terms and conditions apply. Need a hiring hero? This is a job for indeed sponsored jobs.

**SPEAKER_2** (0:30)
Whatever your thing, it could be anything. Canva helps you make that thing a thing.
Canva is a simple online tool thing. It's a way to design with our magic AI tool things. You can social media your thing, generate images or videos of your thing, make decks for presentations to show your thing. Whatever needs to be done for your thing, Canva can make it an even better and bigger thing. Canva, the thing that makes anything a thing.
Thanks for watching.

**Stage Zero** (1:30)
Tesla abruptly shut down its internal supercomputer project, DOJO, and immediately filed a trademark for something called Megapod. It's basically a strategy to turn their global charging network into a distributed artificial intelligence data center.

**SPEAKER_4** (1:46)
Right, which means they abandoned a custom-built supercomputer that analysts previously valued at half a trillion dollars. They fired the project lead, and now they're betting the company's computing future on hardware sitting in parking lots and charging stations.

**Stage Zero** (2:01)
So the real question you have to ask is how a company replaces a centralized supercomputer with hardware distributed across global parking lots and why electricity is suddenly the most valuable currency in computing.

**SPEAKER_4** (2:13)
You really have to look at exactly what they walked away from first. I mean, DOJO was not just some side project.

**Stage Zero** (2:19)
No, definitely not.

**SPEAKER_4** (2:20)
It was intended to be an exascale computer built entirely from scratch. They weren't buying off the shelf components for this. They designed their own silicon starting with that custom D1 chip.

**Stage Zero** (2:30)
And the architecture of it was incredibly rigid. It flowed upward in a very strict hierarchy. You started with CPU nodes which fed into the D1 die.
Every single D1 die had 354 computing cores.

**SPEAKER_4** (2:43)
Yeah.

**Stage Zero** (2:44)
Then they arranged 25 of those individual dies to create a training tile. Six of those tiles made up a system tray, two trays made a cabinet and then ten cabinets formed what they called an Exopod.

**SPEAKER_4** (2:54)
So by the time you reach that Exopod level, you have over one million individual computing cores producing roughly 1.1 exaflops of compute power.

**Stage Zero** (3:03)
Right.

**SPEAKER_4** (3:04)
And the reason they engineered it that way with that specific hierarchy was for one highly specific task. It was optimized purely for processing vast amounts of video data.

**Stage Zero** (3:14)
Because the cars on the road act as these sensor encrusted robots, they navigate highly unpredictable environments generating immense amounts of visual data while making split-second life and death decisions. Dojo was built to ingest all of that specific visual data and train the full self-driving neural networks. The theory was that a custom chip could process that exact type of data more efficiently, with lower energy consumption and latency than general purpose chips from external suppliers.

**SPEAKER_4** (3:43)
And the financial markets priced enormous expectations into that hardware. Morgan Stanley modeled Dojo as adding hundreds of billions of dollars to the enterprise valuation.

**Stage Zero** (3:53)
Which is just a massive number to attach to an internal server project.

**SPEAKER_4** (3:57)
It is, but the thinking was that if Dojo accelerated the autonomous driving software, it would eventually open up a software-as-a-service licensing business. You wouldn't just sell cars anymore.
You would license the vision processing software to any device with a camera.

**Stage Zero** (4:12)
Then Peter Bannon, the project lead, leaves. And around 20 engineering team members depart for an artificial intelligence startup. The company states publicly that dividing their resources to scale two very different chip designs is just inefficient. So they disbanded the Dojo team, reassigned the remaining staff, and completely shifted away from building custom chips for giant training supercomputers.
Now they're relying entirely on external partners, specifically Nvidia and AMD, for training models. Internally, their silicon engineering is focused strictly on the AI5 and AI6 chips, which are designed for edge inference inside the actual vehicles and robots.

**SPEAKER_4** (4:52)
Which makes sense when you consider that developing silicon is arguably the most complex engineering challenge on Earth. Maintaining two parallel development tracks requires splitting your absolute best engineering talent.

**Stage Zero** (5:04)