Tesla's $16.5 Billion AI Chip Order Is Not What You Think It Is.

The company is playing catch-up and failing.

Photo by Bram Van Oost on Unsplash

A few days ago, Musk proudly announced that Samsung’s mysterious $16.5 billion AI chip order was commissioned for Tesla and would be used to train their Full Self-Driving (FSD) AI. This self-important bragging might have impressed those deeply invested in the Tesla bubble, but for anyone actually paying attention, this was an admission of just how deranged Tesla’s self-driving push has become. It also proves that Musk isn’t doing anything to address the actual problems at Tesla. Let me explain.

This $16.5 billion deal is to supply AI-optimised chips to Tesla from 2026 to 2033. It doesn’t include construction costs, installation costs, energy costs, AI training costs, or maintenance costs — all of which are several times the cost of the chips themselves.What’s more, AI chips only have a lifespan of one to three years, given that these chips are pushed hard and often break under load.

So this means Tesla is actually planning on spending around $50 billion on AI over the next decade, or roughly $5 billion a year.

That is a gargantuan expenditure, but Tesla desperately needs it.

They have already spent well over $10 billion, possibly close to $20 billion, developing FSD. Yet, public data of FSD users (using the same hardware and software version as the Robotaxis) shows it travels an average of 444 miles between critical disengagements, when the driver needs to step in and stop illegal driving or an accident. However that data is a little skewed, as FSD customers are aware of its limitations and don’t use it in more challenging settings. Third-party tests, which don’t have this bias, show that the actual distance between critical engagements when using FSD as a full self-driving system is a mere 31 miles.

That is nowhere near safe enough for a driver assist system, let alone a system design, to be used fully autonomously.

For comparison, back in 2022, Waymo had spent around $10 billion developing its self-driving cars, yet its average distance between disengagements was 7,800 miles — which is 251 times better!

Since then, Waymo has spent around $1.5 billion per year on developing its AI systems, updating its fleet, and expanding its operational area. And this has massively paid off. The cars are now much safer, with an average critical disengagement rate of over 17,000 miles, the number of rides they offer skyrocketing to over 250,000 per week, and their operations finally breaking even.

They are the safest, largest, and most financially healthy robotaxi company out there. And they achieved this while spending a third less than Tesla is planning to.

So, does this mean that Tesla will catch up to Waymo?

Hell no!

I have addressed this topic before, but Waymo has redundancy, and Tesla doesn’t. tThat’s the problem.

Waymo uses lidar, radar, ultrasonic sensors, and an array of cameras to understand the world around it. It also has detailed 3D maps of its operational area to check this data against and uses several AIs running on different data, which it cross-references to remove anomalous results. This gives it huge redundancy. If the cameras are obscured by harsh light, or the computer vision AI fails to interpret the world correctly, if the lidar fails, or if the self-driving AI gets something wrong, there is something in place to mitigate that error and allow the car to drive safely on. This enables it to operate comparatively safely, without its AIs needing to be fully accurate.

Tesla is the complete opposite. It only uses a handful of cameras to understand the world around it and doesn’t have a 3D map of its operational area to reference that data against. As such, its computer vision AI needs to be almost 100% accurate, and its self-driving AI must also be nearly completely accurate for it to drive even remotely safely. These systems aren’t accurate at all, and that is why Tesla is so unsafe compared to Waymo.

But, there is this concept I’ve written about before called the efficient compute frontier. Basically, AI faces a significant issue with diminishing returns. We are already observing this with OpenAI, whose models are orders of magnitude larger than before but only provide a tiny increase in performance. Naturally, getting an AI to near 100% accuracy is not possible, as it is a Sisyphean task.

This means that it costs Tesla significantly more to increase its self-driving safety than it does Waymo for the same improvement — which explains why Tesla is having to dump $5 billion a year into AI alone, while Waymo is building a gargantuan near-profitable robotaxi company, vehicle infrastructure and all, for a third of that. Musk shouldn’t be proud of this $16.5 billion deal. It highlights just how far behind they are and just how hard it will be for Tesla to catch up with Waymo.

Meanwhile, Tesla’s sales continue to fall, and the company is set to slip into unprofitability soon. This isn’t just a backlash to Musk’s political ambitions, but because Tesla’s ageing lineup is too restricted, expensive, and not technologically advanced anymore. Tesla competitors are launching cars that outclass them in every way, in far more sectors than Tesla can dream of. So, why spend $50 billion on an AI project that we know will go nowhere when that money could be spent getting Tesla’s core business, its EV, up to par with the rest of the market? It’s almost as idiotic as Tesla’s approach to self-driving.

---

Thanks for reading! Don’t forget to check out my YouTube channel for more from me, or Subscribe. Oh, and don’t forget to hit the share button below to get the word out!

---

Sources: CNBC, Will Lockett, Inside EVs, RTA, Electrek, Teslaratti, The Information, Will Lockett, ARXIV, Will Lockett, TechCrunch, Will Lockett, CNBC