Tesla Killed Project Dojo — Here's the Real Reason Why

By Joe Provence

(Updated March 2026) 2:59 PM MST

In the summer of 2021, Elon Musk stood on a stage at Tesla's first AI Day and made a promise that sent analysts into a frenzy. Project Dojo, he declared, would make Tesla the most advanced AI company on the planet. The custom-built D1 chip at its core would free Tesla from dependence on Nvidia, slash training costs, and give the company an insurmountable edge in the race to full self-driving. Morgan Stanley ran the numbers and told investors that Dojo alone could add $500 billion to Tesla's enterprise value.

Four years later, Musk quietly killed it.

No press release. No AI Day eulogy. Just a post on X: "Once it became clear that all paths converged to AI6, I had to shut down Dojo and make some tough personnel choices, as Dojo 2 was now an evolutionary dead end."

That's the official story. The real story is more complicated — and more instructive for anyone who builds technology, bets on it, or simply follows how billion-dollar decisions get made inside the world's most watched companies.

What Dojo Was Actually Supposed to Do

To understand why Dojo's failure stings, you first need to understand what made it ambitious.

Most AI supercomputers are built like a sprawling office campus — processors, memory, and networking gear spread across different buildings, constantly shuttling data back and forth through congested corridors. It works, but the communication overhead creates bottlenecks that slow training times and drive up costs.

Dojo's architecture tried to solve that differently. Tesla designed a system-on-a-chip that compressed the entire operation — compute, memory, and interconnect — onto a single, tightly integrated die. The D1 chip was purpose-built for one job: processing video footage from Tesla's fleet at massive scale to train the neural networks behind Autopilot and Full Self-Driving.

The vision was elegant. Tesla collects more real world driving data than any other company on earth. Feed that data into a supercomputer purpose-built to process it, and you compound your advantage faster than any competitor relying on general-purpose hardware ever could. Dojo wasn't just an infrastructure project — it was supposed to be Tesla's AI moat.

Three Things That Went Wrong

1. Nvidia Refused to Stand Still

Dojo was conceived in an era when Nvidia's dominance in AI training felt like a vulnerability worth attacking. By 2021, Nvidia controlled roughly 95% of the AI training chip market, and every company in the world was competing for the same limited supply of H100s. The dependency felt dangerous.

But while Tesla spent years developing a custom chip to escape Nvidia's grip, Nvidia kept shipping. The H100 gave way to the H200. The H200 gave way to the Blackwell architecture. Each generation widened the performance gap that Dojo's D1 chip was supposed to close, and the Dojo team found itself in an arms race it hadn't budgeted for.

By the time Dojo 2 was approaching readiness, the calculus had shifted. Tesla would have been deploying a custom chip into a market where Nvidia had a multi-year head start, massive manufacturing scale, and an ecosystem of software tools that Tesla's team would have had to replicate from scratch.

2. Tesla's Own AI6 Chip Changed the Math

This is the part Musk's public explanation actually gets right, even if it's incomplete.

Tesla had been developing the AI6 chip in parallel — originally designed as an inference chip, meaning it was built to run trained models in real time inside vehicles and robots rather than train new models from scratch. Inference and training are traditionally different problems requiring different chip architectures. Training demands high numerical precision and sustained throughput over long runs. Inference demands low latency and energy efficiency at the edge.

But as AI6 matured through its development cycles, something unexpected emerged. The chip's architecture turned out to be surprisingly capable at training tasks — not world-class, but competitive. And in the economics of hardware, a chip that handles both jobs adequately and eliminates the need for a separate infrastructure can beat a chip that does one job brilliantly.

Tesla had invested in AI6 heavily enough that walking away from it wasn't an option. Scaling Dojo alongside it meant funding two separate chip programs, two separate manufacturing relationships, and two separate software stacks. The math stopped making sense.

3. The Team Walked Out

The version of events Musk prefers to tell positions the Dojo shutdown as a rational strategic pivot. The version Bloomberg reported tells a different story.

Twenty engineers from the Dojo team had already left before the official shutdown — not for other companies, but to found their own. DensityAI, led by Ganesh Venkataramanan (the original head of Dojo before his 2023 departure) and fellow ex-Tesla engineers Bill Chang and Ben Floering, launched in stealth to build data center chips for AI applications in robotics, autonomous driving, and agentic computing.

Peter Bannon, the chip architect who had overseen both Dojo and AI6, left Tesla around the same time the shutdown was announced. Whether he departed voluntarily or was part of Musk's "tough personnel choices" was never publicly clarified.

What's clear is that by the time the official shutdown order came down, the project had already been hollowed out by attrition. The decision to kill Dojo may have been as much a recognition of reality as a strategic choice.

What Tesla Is Doing Instead

The shutdown didn't mean Tesla is stepping back from AI infrastructure — it means the strategy shifted from building chips to buying them.

Tesla has signed a $16.5 billion deal with Samsung to manufacture its AI6 chips through 2033. TSMC is handling production of the AI5 chips. And in Austin, Texas, Tesla is building Cortex — a massive AI training supercluster powered by more than 100,000 Nvidia H100 and H200 GPUs. The company that spent four years trying to escape Nvidia's ecosystem is now one of its largest customers.

Musk has framed this as convergence rather than retreat: AI6 chips in a cluster configuration, handling both inference and training, potentially becoming what he called "Dojo 3 in spirit." Whether that framing holds up technically is debatable. What's not debatable is that the $500 million Tesla invested in a Dojo facility in Buffalo, New York now sits as a very expensive reminder of how quickly the AI hardware landscape can shift beneath a long term bet.

What This Means for Everyone Watching AI

The Dojo story isn't just a Tesla story. It's a case study in the compounding risk of vertical integration bets in a market moving faster than any single company can track.

When Tesla conceived Dojo, the logic was sound: own your critical infrastructure, reduce dependency on suppliers, build a moat. The same logic has driven Apple's chip program, Amazon's Graviton, and Google's TPUs. The difference is that Tesla was trying to out-engineer Nvidia in Nvidia's core market, with a smaller team, less manufacturing leverage, and a shorter runway than any of those comparisons.

The lesson isn't that vertical integration is wrong. It's that the bet has to be sized to the time horizon. Tesla's Dojo bet assumed Nvidia would remain a constrained, expensive commodity for long enough to make a custom alternative worthwhile. Nvidia didn't cooperate.

For small and mid-size businesses watching this story from the outside, the takeaway is quieter but just as relevant: the tools you build your operations around should be evaluated not just for what they do today, but for how they hold up when the market moves. The companies that navigate that question well — whether they're building supercomputers or choosing a website platform — are the ones that stay competitive without betting the farm on a single architectural decision.

A Note From Salt Creative

We write about technology and AI because our clients — small and mid-size business owners across Boise, Spokane, Portland, and Colorado Springs — are navigating these questions every day at a smaller scale. Which tools to adopt. Which platforms to build on. Which bets are worth making.

If you're thinking about your business's digital infrastructure and want a team that follows this space closely, we'd love to talk. Schedule a free strategy session with Salt Creative →