Goldman Sachs Deep Dive Report: Dojo Building Tesla's AI Empire with "Trillions of Dollars" in Returns Expected

Supercomputer Dojo, can it bring a qualitative change to Tesla?

On September 10th, in a report led by Adam Jonas, the chief analyst of Morgan Stanley Automotive and a "Tesla fan," the team stated that Tesla's supercomputer Dojo could increase Tesla's market value by up to 500 billion dollars, while Tesla's current market value is about 800 billion dollars.

Morgan Stanley believes that in the process of seeking solutions to autonomous driving, Tesla has developed an advanced Dojo supercomputer, which is referred to as the "mother of all AI projects" in the automotive industry. Dojo can open up "new potential markets" for Tesla:

Dojo is designed to handle large amounts of data in training autonomous driving systems, and it may bring "asymmetric advantages" to Tesla in this potential trillion-dollar autonomous driving field.

From now on, software and services become the biggest value drivers for Tesla.

The more we research Dojo, the more we realize that the value of Tesla's stock may be underestimated. We believe that Dojo can represent a significant change in the market's perception of Tesla.

Morgan Stanley predicts that by 2040, the revenue from Tesla's network services will reach 335 billion dollars, accounting for more than 60% of Tesla's overall EBITDA, replacing car manufacturing as Tesla's core business.

During the first-quarter conference call this year, Musk stated that Dojo is a "long-term bet" that could potentially bring "hundreds of billions of dollars" in returns if it develops as desired by the company.

Based on this, Morgan Stanley raised Tesla's target price from the previous 250 dollars to 400 dollars, upgraded its rating to "buy," and listed it as Morgan Stanley's "top pick" stock.

Stimulated by Morgan Stanley's report, Tesla's stock price rose by about 10% to 274 dollars on Monday, reaching its highest level since July 19th and the largest increase since January of this year.

Tesla has developed a custom-made AI chip for Dojo, which is said to cost only a fraction of Nvidia's chips but is significantly more efficient. Morgan Stanley estimates that Dojo can save Tesla 6.5 billion dollars in costs in the next few years.

Dojo can shorten the time required for training workloads from over a month to within a week.

Analysts believe that in the long run, Dojo can also help Tesla have cross-industry AI software and hardware capabilities, expanding its influence to more industries.

We believe that once Tesla makes progress in autonomous driving and software, third-party Dojo services can provide investors with the next stage of Tesla's growth story.

With the support of Dojo, Tesla could eventually become a cloud service provider like Amazon AWS.

Over time, these capabilities can be monetized through an "AWS-like" business model.

Why is Dojo so fast?

According to Tesla's claims, Dojo can shorten the workload of autonomous driving training from one month to one week.

Morgan Stanley analyst Joe Moore believes that vision-based training is not as complex as training large language models (LLMs), but it requires a larger amount of data (video input from 10 cameras continuously running on each vehicle) and higher computational power.

Information provided by Tesla earlier shows that Dojo integrates 120 training modules called ExaPODs, each containing 3000 D1 chips. It has over 1 million training nodes, and its custom hardware and software can further enhance the efficiency of D1 chips, enabling them to perform 1 EFLOP of computational power, which is equivalent to 1 quintillion floating-point operations per second.

Morgan Stanley states that D1 is Tesla's self-developed 7nm chip manufactured by TSMC. Their semiconductor team's investigation shows that Tesla has ordered approximately 40,000 to 50,000 D1 chips this year.

The Dojo D1 chip consists of small nodes, each containing a dedicated 64-bit CPU with over-spec cores. Based on TSMC's 7nm technology, each node is like a separate computer with a dedicated CPU, local memory, and communication I/O SerDes interface. This means that each core can run independently without relying on shared cache or register files.

These chips do not support virtual memory like other chips because management believes it takes up too much space, reduces bandwidth, and increases latency. Instead, each core has 1.25MB of SRAM as the chip's primary internal memory, with a loading speed of 400GB/s and a storage speed of 270GB/s.

According to Tesla's management, the speed of one Dojo Tile (integrating 25 D1 chips) is 30 times that of 24 GPUs. Performing specific operations on 25 D1 chips takes 5 microseconds, while it takes 150 microseconds on 24 GPUs.

Therefore, analysts believe that compared to current GPUs, Dojo's potential efficiency, speed, and cost advantages can significantly shorten Tesla's autonomous driving development time. According to Tesla's plan, its Palo Alto data center will build 7 Dojo ExaPODs. By February 2024, Tesla's computing power will rank among the top five globally, and the total scale in October will reach 100 Exa-Flops, equivalent to the combined computing power of 300,000 NVIDIA A100 graphics cards.

However, Morgan Stanley pointed out that the current data about Dojo is all from Tesla's own promotional materials, so the actual situation still needs to be verified:

Although it is difficult to verify Tesla's claims about the cost and performance of Dojo, we believe that with Tesla's track record of innovation and capability, the company has the opportunity to launch competitive custom solutions.

Dojo has been operational since July this year, and we believe that continued promotion and subsequent company announcements will serve as catalysts for investors to understand the potential of Dojo.

Cars + Software: Dojo Enables Tesla's "Dual Power" Mode

Morgan Stanley's report states that Dojo boosts Tesla's valuation mainly in two aspects: cars and autonomous driving software, that is, Dojo is the "intersection between hardware and software" for Tesla.

For years, we have been trying to get investors to focus on Tesla's transformation of car owners into "users," thereby generating high recurring (high-profit) revenue potential.

So far, Tesla has achieved some success in this regard, but we believe that Tesla's internal computing power could significantly accelerate the speed of network effects and data collection/analysis/learning. By 2027, the total mileage driven by vehicles using Tesla FSD (Tesla + third-party) globally will reach 1 billion miles per day.

Analysts believe that Dojo will help accelerate the adoption of Tesla's autonomous taxi plans, and Tesla can also sell software and hardware to third parties, achieving the "dual flywheel" development of cars + software:

We will focus on the potential of Dojo, which is to provide autonomous driving and network service revenue with faster adoption rates and higher average revenue per user (ARPU), thereby significantly increasing our valuation.

With the computing power enhancement brought by Dojo, Tesla's monetization of car software can be realized faster, and higher recurring revenue can be obtained.

Morgan Stanley predicts that by 2040, Tesla's revenue from network services will reach $335 billion, compared to the previous forecast of $158 billion, and it is expected that this segment will account for more than one-third of Tesla's total EBITDA by 2030, and more than 60% of the overall EBITDA by 2040 (previously expected to be 38%). .**

We will leverage Dojo's additional value in software services (by increasing ARPU and adding third-party licenses), Tesla's automotive business (by increasing the size of the long-term fleet and profit margin), and third-party battery business, as we believe that charging and FSD transactions will also bring higher hardware value-add.

We have conducted a comprehensive evaluation of Tesla's valuation based on new profit forecasts. By the 2025/26 fiscal year, Tesla's profit forecast has increased by about 20%. We believe that this will provide a solid foundation for our target stock price of $400, compared to the growth and valuation multiples of related technology peers.

Is Tesla Becoming an AI Cloud Service Provider?

Dojo's impact on Tesla and the entire industry may go beyond just autonomous driving.

Morgan Stanley believes that Tesla is likely to offer Dojo-based visual machine learning systems in the future, extending its influence beyond the automotive industry to any industry that requires a visual focus.

If another original equipment manufacturer incorporates Tesla's FSD system into their vehicles, Dojo's scalability will greatly increase. Tesla could become one of the best visual and video processing machine learning training systems in the market.

Compared to the currently preferred general-purpose GPUs, Dojo's design use cases are more favorable for transformation-based neural network training. The chip is designed to advance and accelerate autonomous driving and can be utilized in any scenario involving a large amount of complex video recognition data.

Similar to Nvidia dominating the GPU market and Qualcomm dominating the mobile semiconductor market, Tesla may have the ability to become a leader in the market for machine learning training using visual data, with potential profits exceeding those from producing autonomous vehicles.

While many industries can benefit from accelerated visual artificial intelligence, we believe that Dojo could be applied in certain industries to accelerate safety and efficiency through visual data AI applications, such as robotics, aviation, and security.

However, Tesla seems more inclined to offer cloud services in the future rather than selling chips and systems.

Previously, Tesla's Senior Director of Hardware, Ganesh Venkataramanan, stated that Tesla's immediate priority is to run Dojo internally but eventually it could be opened to non-Tesla users as well. During the first-quarter conference, Musk also mentioned that Dojo's future could be like Amazon AWS, a cloud service provider.

Therefore, Morgan Stanley has referenced the development path of Amazon AWS and suggests that Tesla's network services could also become a revenue pillar in the future.

As an internal cloud service, AWS provides developers with a shared software infrastructure that helps improve their efficiency. With the growth of bookstore and online marketplace businesses, the company realized that the technology used to support its core business could be monetized by renting it out to third parties.

Today, AWS accounts for 70% of Amazon's pre-tax profits (according to our Internet team's 2023e data), while the retail business contributes only 20% of Amazon's revenue.

We believe that the potential of Tesla's "cloud" is a long-term "moonshot" that ranks low on the company's priority list. However, at the same time, we recognize that Tesla is building important internal capabilities that can be monetized over time through an "AWS-like" business model.

We do not believe that Dojo will replace AWS. However, we do believe that AWS provides a strong example of how leveraging core capabilities (specialized data processing and software) can create tremendous value drivers.