The Muskverse is a noisy place.Competing for our attention today are war information fogSome anti-semitic tropes“terminal monkey” and a seemingly ever-growing cast A child with a strange name.
It would take something big to cut through all this noise, or at least something over the top: like the 66-page research report written by Morgan Stanley analyst Adam Jonas and six of his colleagues.
Yes. . . Hahaha. . . fork!
Investors have long debated whether Tesla is a car company or a technology company. We think it’s both, but the biggest value driver is software and services revenue. We believe the same forces that drove AWS to 70% of AMZN’s total EBIT can also be at work at Tesla, opening up new addressable markets that extend well beyond selling vehicles at a fixed price. catalyst? Dojo, Tesla’s custom supercomputing effort over the past five years. The 12th version of Tesla’s fully autonomous driving system (OTA before the end of the year) and Tesla’s next AI Day (early 2024) are worthy of attention.
We believe Dojo can add up to $500 billion to Tesla’s enterprise value, reflected by faster adoption of mobility (robotaxis) and web services (SaaS). This change prompted our PT to rise to $400, compared to the previous $250. We upgrade to Overweight and rank Tesla as a top pick.
Maybe you haven’t heard of Tesla Dojo.Maybe you’ve been relying on something like this Tesla 2022 Annual Report and its 10-K Filingneither mentioned Dojo once.
Morgan Stanley applies a broader frame of reference:
This is the gist of the argument. Teslas “are sensor-equipped robots that make life and death decisions in highly unpredictable environments and situations.” Their next generation proprietary brain will be the Dojo chip, developed in-house by Tesla specifically to capture large amounts of data.
While regular chipmakers have to consider whether their new chips can still run Apache Spark and FIFA 23, Tesla’s GPU team, like Mr. Miyagi, is focused on advanced driver-assistance systems.
“One of the problems facing generative AI and ML (machine learning) training systems is that advances in hardware lag far behind software,” Morgan Stanley explained. “Tesla’s approach to solving this problem is to gradually develop its own innovative hardware architecture to provide the ideal architecture for running Tesla’s calculations. Dojo hardware is developed to accelerate itself without relying on external devices because each chip They are all their own supercomputers.”
Single-purpose specialization will make its mini-supercomputers even more superior, doing to artificial intelligence what mining ASICs did to cryptocurrencies:
Tesla’s experienced semiconductor team built a custom AI ASIC chip that can run more efficiently (energy consumption, latency) than leading cutting technologies because its core function is to process vision-based data for autonomous driving use cases. ). The edge general-purpose chips on the market (NVIDIA’s A100 and H100) may cost a fraction of the latter.
Tesla is not the first tech company to try to build a custom chip system in-house, but given the company’s deep understanding of ADAS (a pioneer in the electric vehicle market) and its vast and growing data network (400,000 FSDs are already on the road) ), a world-class design team, and extensive resources, in addition to the diverse potential needs of over-reliance on NVDA, we believe Dojo may prove to have Competitiveness.
Evidence of Dojo’s potential often comes from Tesla demos, such as the company’s 2021 and 2022 AI Days, where the architecture announced firstand what Musk said about the second-quarter conference call Project R&D spending exceeds US$1 billion over the next year.
Citing the updates, Morgan Stanley speculates that the Dojo chips will deliver six times the performance of the current generation-only Nvidia A100 GPU box, with a single Nvidia box currently costing $200,000 per unit. Fewer boxes may also reduce energy consumption because they require less cooling.
Even this analysis must add a caveat to Tesla’s record of achieving goals (or even just setting them). For example, the prediction given in Tesla’s Q2 update for Dojo to reach 100 exaflops (meaning 100 billion floating-point operations per second) would mean that the computing cost is almost double the computing cost proposed at AI Day 2022.
Morgan Stanley said that “considered as a whole, the falsification of numbers may reveal some definitional inconsistencies and broad interpretation by investors,” and some investors may feel that such courtesy is not deserved.
Reality sometimes seeps through the cracks, as when Morgan Stanley semiconductor analyst Joe Moore makes a cameo in the film to explain what generative AI chips actually are. He began by describing the market in two distinct areas—training, the time-consuming and power-consuming compilation of data models; and inference, the process of calling those models to do things.
ASICs for artificial intelligence are coming soon, with Google, Amazon, Microsoft and Meta all announcing projects of their own. But Nvidia has proven difficult to pivot. Making a lot of money from the high-volume business of gaming cards means Nvidia can regularly iterate on its existing designs and reap immediate returns on its $8 billion in annual R&D spending.
Competitors start everything from scratch. Any technological lead they find is often temporary, as Nvidia never lags behind. Inference chips are the boring workhorses of the AI space, and performance mostly means efficiency, so assuming customers can adapt to software compatibility, there may be some competition. But training, Dojo’s domain, is a different story.
Building machine learning models requires huge upfront costs, so task optimization and expertise are important. Among Nvidia’s challengers, only Google, the AI pioneer that invented the transformer less than a decade ago, understands its customers’ needs well enough to occupy a place in the construction of the AI world.
Could Tesla do the same with its object labeling space? Moore spoke highly of the company’s internal chip team, which was built by industry superstar Jim Keller between January 2016 and April 2018 when he left Intel. Moore also noted that Tesla has an advantage over many underfunded startups because it understands both hardware and software. The final product “may prove competitive in its customized use cases,” he said:
Tesla isn’t competing to make better chips. Tesla is optimizing for a single purpose, increasing overall production with greater efficiency and lower cost. NVIDIA develops the world’s most powerful GPU chips to harness the demanding performance of gaming. Can Tesla capitalize on autonomous vehicle/FSD demand and become the global leader in custom AI chips?
Yes!Jonas says et al.. Yes, it can! look!
look!
look!
According to Morgan Stanley’s forecast, Tesla’s electric vehicles are worth 28.3 times 2025 earnings before interest, taxes, depreciation, and amortization (EBITDA). It’s more expensive than Nvidia’s 25.3 times Ebitda. “WWe believe the growth in Dojo synergies in the second half of the decade justifies the valuation,” the team said:
As Tesla begins to unlock Dojo synergies in the second half of the decade and beyond 2030, we expect meaningful EBITDA margin expansion. We forecast that network services will achieve 65% EBITDA margins and account for 62% of Tesla’s total EBITDA by 2040. Therefore, we can imply that the company’s EBITDA margin will be 35% in fiscal 2040, which is higher than 15% in fiscal 2023 and 24% in fiscal 2030.
Edward Stanley, head of equity strategy at Morgan Stanley, was also invited to provide a separate note on the Tesla Dojo’s possible contribution to Bessebinderism:
As we’ve written before, just 2.3% of stocks generated $73 trillion in net shareholder returns over 30 years. All investors should be interested in finding moonshots that are plausible, scalable, and not yet completely undervalued by the market. We believe Tesla’s Dojo project ticks all the boxes. (…)
Although Dojo is still in its early stages of development, we believe that if Moonshot is successful, its applications could expand beyond the automotive industry in the long term. Dojo is designed to process visual data, laying the foundation for vision-based artificial intelligence models such as robotics, healthcare, and security. We believe that third-party Dojo services can provide investors with the next phase of Tesla’s growth story once Tesla makes progress on autonomous driving and software. Humanoid robot technology and its potential to replace workers and social workers is one of our initial and adjacent moonshots, but Dojo’s success will significantly accelerate that goal.
Jonas has previously made various suggestions to Tesla, including battery production, charging station and insurance, where AI has not previously had the same traction, he accompanied the downgrade to “equal weight” on June 22, 2023, explaining:
While we understand why Tesla receives serious mention in the AI conversation, we believe a re-rating on this topic falls into the realm of irrefutable bull cases. We believe that autonomous driving and generative artificial intelligence remain two distinct technical disciplines. While the market may want to dream about artificial intelligence themes, we’re prepared to wake up to the sound of blaring car horns.
Beep.
Further reading:
Tesla’s Springfield Canyon Trail (FTAV)
Morgan Stanley Adam Jonas’s (FTAV) Best Works
Svlook
