Autonomous Driving: The Anatomy of Autonomy (Part II)

Autonomous driving (AD) companies are racing to shape commercialization opportunities, which rest on developing cost-effective and scalable AD technology stacks. Those stacks can be divided into three key segments: sensors, on-board processors, and data centers.

In each segment, companies will be in hypercompetition to win their share. The result will be a booming new ecosystem of auto industry suppliers, some of which will be legacy suppliers that have adjusted. Others will be new entrants.

So far, the world’s most mature AD supplier ecosystems are in the United States and China. When it comes to commercial viability, AD is not just a two-way competition between Waymo and Tesla—China’s Baidu and Huawei should be thrown in the mix as well. For example, Baidu recently debuted its $30,000 RT6 robotaxi, which is possible largely due to China’s cost-effective supply chain. Plus, China, like the United States, has a legion of auto companies looking to deploy AD.

Indeed, robotaxi fleets are the most obvious AD business model and are already regionally deployed in cities like Austin and Wuhan. That means the race is on to secure the optimal AD technology stack—a prerequisite for successful mass commercialization. In part two of our autonomous driving series, we look closely at competition in the supplier ecosystems that have sprouted up in China and the United States.

No Lidar or Some Lidar?

Autonomous vehicles use three main types of sensor hardware to perceive the external environment: cameras, lidar, and millimeter wave radar. Waymo’s sixth generation robotaxi, for example, is equipped with 13 cameras, 4 lidars, 6 radars, and an array of external audio receivers. This sensor-heavy stack largely aligns with other AD companies, especially competitors in the Chinese market.

Figure 1a. Waymo’s Sensor Hardware SuiteSource: Waymo.

However, Tesla stands out because it insists on a camera-centric approach, which gained buy-in in 2023 when Chinese researchers published a paper demonstrating that a self-driving model called the BEVFormer (Bird’s Eye View + Transformer) can generate a bird’s eye view of route planning using only cameras (see Figures 1a & 1b).

Figure 1b. Tesla’s Camera-Centric Approach Stands Out Note: Baidu’s lidar supplier, Zvision, priced its EZ6 lidar at just ¥2,000 yuan (~$276) per unit, reducing lidar’s cost contribution to the RT6 to just over 7%.
Source: Company announcements.

Yet even as the per unit cost of lidar has dropped precipitously since 2019, Tesla is holding fast to its camera-centric end-to-end AD system: No lidar is present in the company’s Full Self Driving (FSD) hardware. Elon Musk seems to to believe lidar is more of a liability than an asset to AD because it adds unnecessary complexity to the FSD data engine without providing sufficient performance gains to justify the downside.

If Tesla has its way—and drags a begrudging industry along with it—then the already struggling lidar industry will continue to falter. Global market leader Hesai Technology, as well as US firms Ouster and Luminar, have all seen their share prices decrease by more than 50% compared to early 2023. Only Hesai, the market leader, boasts positive operating cash flow (see Figure 2a & 2b). It doesn’t help that domestic Chinese players like Huawei are increasingly copying Tesla’s approach. For instance, Huawei’s ADS 3.0 software, launched this year, completely eliminated lidar and removed six millimeter wave radars.

Figure 2a. Non-Chinese Lidar Suppliers Are Uncompetitive…
Source: Yole Intelligence.

Figure 2b. …And Also Face Financial Woes
Note: Data in $, mn based on FY 2023 results.
Source: Company reports, Yole Intelligence.

In contrast, South Korean giants Samsung and LG, which both supply Tesla with high-definition cameras, are seeing success—thanks to mature camera technology previously honed for the smartphone supply chain (LG supplies Apple’s camera modules for the iPhone). While Chinese companies have made progress in indigenizing image sensor components for cameras, they have yet to capture significant camera market share from Korean suppliers. They also face stiff competition from Tier 1 US and European suppliers like ZF, Bosch, and Continental.

In radar, the largest market of the three AD sensor types, millimeter wave (mmWave) is beating ultrasonic due to its more durable performance and longer range that can extend hundreds of meters. Both Waymo and Tesla use mmWave radar exclusively in their latest AD stacks, an area where Chinese suppliers see an opportunity to close the gap.

As AD technology evolves, the optimal sensor suite will shift to continue balancing performance and cost. These sensors generate huge amounts of data, spurring competition in another vital component to the AD stack: on-board semiconductors.

Self-Driving Silicon: Buy Off the Shelf or Design Your Own?

While sensors take in the visual information of the external environment, on-board chips are needed to process terabytes of data per hour internally. These multifunctional systems-on-a-chip (SoCs) integrate core GPUs, AI accelerators, and CPU cores—all in real time.

Competing chip suppliers are optimizing for two standard criteria: efficiency and latency. For autonomous vehicles, power efficiency is even more important than it is for power-hungry data center GPUs. Since the future of autonomous vehicles is likely all electric, AD systems rely solely on battery power. If the AD system depletes the battery by, say, 15%-20% every hour, the entire car would need to be redesigned to accommodate much larger battery packs. Latency is also crucial. To create a safe AD system, the time it takes for a visual input to be processed into a response needs to be as fast as, if not faster than, human reaction time.

Consequently, creating a SoC for self-driving that achieves performance and energy efficiency is at the heart of AD’s hardware-software integration challenge. AD chips are typically bespoke products designed to function within a specific software architecture. Therefore, companies that own their chip design and talent in-house have a leg up because they exert more end-to-end control from the hardware to the driving experience.

This is precisely the route Waymo and Tesla are adopting, because it allows them to better control innovation at the system level, optimizing interactions between software and silicon. They are also able to better align their chip design timelines with the rest of their AD product cycles—representing one of the key advantages they have over off-the-shelf chip suppliers.

But in-house SoC development is costly and technically out of reach for most AD companies, including automakers and autonomous truck startups, making them reliant on third-party SoC suppliers. Since customers’ AD software architectures are constantly evolving, third-party suppliers have to respond quickly to shifting market demand rather than count on one-size-fits-all solutions.

For these suppliers, agility is central. Their ability to iterate rapidly and keep up with the fast-changing and specific needs of customers is more telling than how capable their chips are today.

By that metric, China’s Horizon Robotics and US’ Qualcomm seem to be leading the pack (see Figure 3a). Both are late entrants into the AD SoC space, but Qualcomm in particular has notched major design wins since 2022 to supply the likes of BMW and Mercedes Benz, and it has touted a robust automotive chip pipeline worth over $30 billion (see Figure 3b).

Figure 3a. Product Iteration Cycle Is a Good Indicator of Competitiveness in AD Processors
Note: Data show average months between AD SoC products reaching mass production over last four product generations (five in Mobileye’s case).
Source: Company announcements.

Figure 3b. Intense Competition Results in Overlapping Customer-Supplier Relationships
Note: Customers are not exhaustive.
Source: Company announcements and presentations.

Nvidia’s strong SoC market position is largely due to a first-mover advantage (its first AD solutions were released at CES 2015). But it falls into the off-the-shelf supplier category for AD chips, as its GPUs are not naturally optimized for on-vehicle processing. In an indication that Nvidia might be struggling in the AD segment, the company cancelled its Atlan AD chip in 2022 and has delayed its next-gen Thor chip twice so far. Moreover, Nvidia began to lose market share in China to Horizon Robotics during the second half of 2023 and responded by launching a hiring spree in China focused on recruiting AD engineers (see Figure 4).

Figure 4. Horizon Robotics Threatens Nvidia in China (2023)
Note: Market share for chips used in AD systems.
Source: GGII (高工产业研究院).

While successful AD chip suppliers are emerging in both the United States and China, competition in the next layer of the AD supply chain—supercomputing—might be stacked against one side.

Scale and Supercomputing

Whether they’re driving in San Francisco, Austin, Wuhan, or Beijing, robotaxis are still in their “pilot” phase. AD firms’ expansion plans will place much more demand on the last crucial piece of the AD technology stack: data centers. These supercomputing clusters, which are needed to train vision language models (VLMs) far away from where the rubber meets the road, have to process increasing volumes of data as more autonomous vehicles hit the streets.

Powerful VLMs are essential to scaling because they enable robotaxis to operate on any roads in any conditions, instead of remaining limited to pre-mapped areas (which is the case for Waymo’s current model). That’s why massive amounts of compute are especially vital to Tesla, which constantly processes driving data collected by the millions of vehicles already running FSD software to improve AD performance.

According to Musk, Tesla’s Dojo project, an AI supercomputing cluster in Austin, Texas, will reach 100 EFLOPS this year. That would be attributable to tens of billions in investment, including Tesla’s purchase of the equivalent of 90,000 Nvidia H100 GPUs. If it lives up to those projections, Dojo will dwarf current compute capabilities available to Chinese leaders in the AD space like Huawei, which boasts 3.5 EFLOPS via its cloud, while Xpeng’s 8 EFLOPS and NIO’s 1.4 EFLOPS, respectively, are accessed via Alibaba Cloud services (see Figure 5).  

Figure 5. Tesla’s Dojo Would Far Exceed Chinese AD Players’ Compute Capacity
Source: Company announcements.

If Chinese firms continue to follow Tesla’s end-to-end approach, limited access to advanced compute will become a more obvious constraint. The US restrictions on access to semiconductors will likely hinder China’s progress in VLMs more than LLMs because the multimodal data (camera images) for VLM training is much more compute intensive. Though supercomputing remains a strategic priority for Beijing, China’s three leading clusters (run by state telecom firms China Mobile, China Telecom, and China Unicom) only sum to 53 EFLOPS, just over half of Dojo’s projected capacity.

Restricted access to compute is just one geopolitical thorn in the side of the AD industry. China and the United States are the world’s two largest auto markets, but AD companies on both sides of the Pacific face significant regulatory hurdles to developing and operating AD businesses in their markets. In part three, we’ll explore how these geopolitical constraints will likely end up bifurcating the AD market.

AJ Cortese is a senior research associate at MacroPolo. You can find his work on industrial technology, semiconductors, the digital economy, and other topics here. 

The author would like to thank Cedar Liu and Guanzheng Sun for excellent research assistance.