Tesla’s (Driverless) Route to Dominance?

There’s an old saying in Tennessee – I know it’s in Texas, probably in Tennessee – fool me once, shame on… shame on you. Fool me – you can’t get fooled again.

-George W. Bush

In 2016, Elon Musk promised every Tesla (TSLA) shipped would “have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver.”

It sounded grand.

But who’s fooling who?

Fast forward to June 2023, and National Highway Traffic Safety Administration data revealed 736 car crashes involving the company’s autopilot software.

Most troubling, 11 of the 18 autopilot fatalities have occurred since May 2022.

And this is likely not a coincidence.

In May 2022, Tesla’s “Full Self-Driving” (FSD) mode was part of a beta-offering expansion, from 12,000 to over 400,000 cars.

This “rollout” was a long time coming.

However, it produced more problems than triumphs… But Tesla recently made a move that could revolutionize self-driving cars and the company forever…

Fool Me Once

The core issues behind this technology are hardware and software limitations to effectively analyze environments.

While FSD stays within lines on the highway and can even change lanes, those are two grains of sand on an enormous beach of driving variables.

Like any use case, the artificial intelligence (AI) that drives this process – no pun intended – requires training.

Enter another tech darling…

Nvidia’s (NVDA) booming success stems from the relative power of its chips. The tech world uses these to train AI models. And due to demand far outpacing supply, companies in need are being left in the lurch.

But rather than wait around for the next GPU shipment, Tesla took matters into its own hands. And some feel this step could alter the trajectory of the company forever.

It took steps to design – and finally roll out – its Dojo supercomputer. Most important of all, this supercomputer is powered by in-house microchips

Morgan Stanley (MS) recently made headlines when it raised its price target on Tesla from $250 to $400.

And the reason is Dojo.

The Wall Street giant feels this collection of cabinets filled with 3,000 microchips is the catalyst for Tesla’s future dominance.

And not just in the automotive market…

But in a world barreling headfirst into AI.

Yet there are others who don’t want to be made a fool once again. And they see this as the latest in a series of overpromises from a CEO who routinely puts his carriage in front of the horse…

DIY or Die

To be reliant on other companies presents a fundamental issue…

If something goes wrong with a supplier, you’re out of luck.

If other companies rely on it as well, supply becomes an issue.

The latter is a struggle Tesla grew weary of with Nvidia. While many companies are content to simply wait, Tesla – understanding the scope of its need – decided to act.

The first Dojo supercomputer officially went online in July 2023. Tesla claims these supercomputers (and accompanying D1 chips) represent a six-fold performance improvement compared to Nvidia’s A100 GPUs.

But the key difference between these chips and Nvidia’s is they’re specifically designed to process visual data, helping to train autonomous driving.

FSD mode is currently hamstrung by the ability of the Tesla fleet to collect information.

Well, by designing a chip and supercomputer specifically for visual analysis, Tesla’s supposedly created far more efficient technology than Nvidia… but at a fraction of the cost.

Here’s where it really gets interesting… a large part of Morgan Stanley’s bullishness stems from the applications of this tech beyond self-driving cars.

It feels the Dojo supercomputer could forever change “everything with a camera that can process data to make decisions.”

One-Way Ticket to New Heights

The biggest pain point in AI development and training is “labeling.” It involves adding context to raw data analyzed by AI.

To oversimplify, by labeling a hundred different images of stop signs as “stop signs,” the AI will be able to recognize stop signs in the real world.

This is a painstaking process.

But it’s essential.

Consider the following…

As of September 2022, Tesla had three supercomputers powered by a combined 14,000 Nvidia A100 GPU units. Almost a third of those GPUs are dedicated to auto-labeling.

According to Tesla, its Dojo supercomputer no longer requires labeling at all.

That’s groundbreaking.

But let’s look at it from another angle…

Those 14,000 GPUs are nothing to scoff at. In fact, one of Tesla’s supercomputers is ranked as the seventh largest in the world.

Now that we have some perspective…

Tesla projects supercomputer performances on par with one powered by 300,000 Nvidia GPUs by the end of 2024.

That implies the current capabilities of its Dojo supercomputers will increase more than 20-fold in the next 15 months.

But not only that, it would alter the AI landscape that Nvidia dominates… and could be the long-awaited thing that transforms Tesla from a car company to a true tech company. Both hardware and software.

Unfortunately, we must take phrases like “Tesla says” or “according to Tesla” with a grain of salt… We’ve been fooled more than once before…

Fool Me Twice?

If there’s one thing we’ve learned about Elon Musk, he isn’t afraid to make promises.

Of course, following through on those promises has been a separate story.

This makes jumping headfirst into those lofty numbers and projections above tricky.

Clearly, the capricious billionaire was a bit off on his projected timeline of FSD.

On top of that, the Morgan Stanley analyst who led the bank’s report, Adam Jonas, is notoriously one of the largest Tesla bulls on Wall Street.

So, it shouldn’t come as a surprise that he jumped headfirst. But when you see disclaimers like the ones below from Morgan Stanley’s research, it’s fair to wonder…

“It is difficult to explicitly validate many claims Tesla has made about Dojo’s cost and performance.”

Or, “We’d like to heavily caveat these cost-saving estimates by noting that this is what the company expects to achieve.”

And, “We note here that Tesla does not provide a performance comparison to Nvidia's next-gen solution, H100.“

For what it’s worth, Tesla is already scaling back its purchases of Nvidia GPUs. It’s moving forward with plans to install six more Dojo “ExaPODs” at its Palo Alto data center.

So, the wheels are turning in earnest.

And if we’re to believe Tesla, those wheels will be on a driverless car sooner than later.

But only time will tell who’s the fool.

Until then,

Kyle Ottenheimer