There's been a lot of talk about physical AI, much of it very optimistic. Physical AI is happening, but progress is slower than many expected. It won't rival general artificial intelligence (GenAI) anytime soon.

We will analyze humanoid robots, fully autonomous vehicles, and then all specialized robots (ASRs), concluding with a rough estimate of the semiconductor TAM for robotics from 2025 to 2035.

Morgan Stanley (May 14, 2025) predicts that by 2050, the humanoid robot market (which we call GPRs: general-purpose robots) will reach $5 trillion, with 1 billion units deployed. This means that one in every 10 people will be a humanoid robot. They indicate that robot adoption will be relatively slow before 2035, but will accelerate in the late 2030s.

At a TSMC Technology Symposium in Taiwan this past May, TSMC's Asia Pacific Business Development Director, Ray Wan, stated, "Following generative AI and agent-based AI, physical AI will be the next key trend..." He predicted that by 2030, the global AI robotics market will exceed $35 billion, with 10% of cars expected to be autonomous. By 2035, he predicts 1.3 billion AI-powered robots will be deployed worldwide, rising to 4 billion by 2050, including 650 million humanoid robots.

Elon Musk said earlier this year that Tesla plans to produce 10,000 Optimus Prime robots this year, with thousands performing useful tasks by the end of the year. The project's leader recently left Tesla. Will Lockett is extremely skeptical, as evidenced by his recent article, "Tesla's Robots Are Pathetic." Tesla recently opened a drive-thru restaurant in Santa Monica, where an Optimus Prime robot delivers popcorn. Internet reports indicate that it is remotely operated by a human and freezes when the connection is lost. At the drive-thru, food orders are delivered by waiters on roller skates—a feat that would be particularly appealing if multiple Optimus Prime robots could deliver food on roller skates.

But as economist Rudiger Dornbusch observed, "…things take longer to happen than you think, but they happen faster than you think."

Humanoid robots are difficult

Humanoid robots will not be a significant market driver over the next 10 years. Humanoid robots are general-purpose robots (GPRs). They are highly agile, but this also makes them more difficult to develop.

The Bismarck Brief estimates that there are currently 1,000 humanoid robots in the world, most of which are prototypes.

Brad Porter is the CEO of Collaborative Robotics, a company backed by Sequoia Capital, Khosla, and Lux. Before joining Cobot, he was Vice President of Amazon Robotics, leading a team of 10,000 people. His article is worth reading, but the key points are as follows:

• Artificial intelligence cannot yet handle a robust balance system. (It takes humans years to learn balance.)

• For most production tasks, humanoid robots are not the right design solution.

• Having wheels with three to four points of contact with the ground and a payload within a cone of stability is the correct design solution.

• Safety is equally important. GenAI can make mistakes, but that's just talk. A humanoid robot capable of lifting a person could also cause death if operated improperly. The use of robots in the automotive and industrial sectors requires rigorous adherence to very detailed safety regulations. The required level of safety testing and design is based on worst-case scenarios. An entertainment system malfunction wouldn't kill someone, but a brake failure or a flailing robotic arm might.

Take self-driving cars, for example. Google (now Waymo) began testing on the streets of Mountain View in 2010. It took more than a decade for the first fully driverless vehicles on city streets. The reason lies in safety. Waymo can reliably handle six-sigma anomalies and, based on millions of miles of driving data, has proven itself to be significantly safer than human drivers.

Humanoid robots will eventually emerge, but it may be a decade or more away. While technical challenges need to be overcome and safety rigorously verified through extensive trials, ASR will remain a more economical alternative.

Self-driving cars will be the first robots most of us encounter

If you go to San Francisco (or Los Angeles, Phoenix, Austin, and several other US cities coming soon), you'll find Waymo cars everywhere. Waymo, owned by Google, has been operating for a decade, but only recently began offering paid rides to passengers.

In San Francisco, Waymo captured 26% of the ride-sharing market share in April 2025, surpassing Lyft for the first time (source: YipitData). Uber's market share exceeded 50%. Waymo initially launched in San Francisco, then expanded to Los Angeles, Phoenix, and now Austin. They've already logged approximately 70 million miles! Over the next 6 to 12 months, they'll move south from San Francisco to San Jose.

The current Waymo vehicle, pictured above, has an estimated total cost of $140,000 and uses $10,000 NVIDIA datacenter-grade GPUs along with more expensive sensors (cameras, radar, and lidar). The upcoming next-generation Waymo vehicle is expected to cost $85,000, without sacrificing safety or features. Morgan Stanley predicts that Waymo's revenue will reach approximately $2.5 billion by 2030, based on 1 billion miles driven by its approximately 17,000 Waymo vehicles. Even with Tesla included, however, it accounts for less than 0.1% of all autonomous vehicle miles driven in the United States. However, Goldman Sachs notes that this represents 7% of all shared ride-hailing miles driven in the United States.

The global auto market, with annual sales of 90 million vehicles (compared to 100 million for trucks), is worth $2.1 trillion. Precedence Research and Coherent Market Insights estimate that the global automotive semiconductor market will be worth $51 billion to $77 billion by 2025. The average semiconductor cost per vehicle is approximately $600. High-end vehicles with more cameras and features cost around $2,000.

The largest suppliers are Infineon, NXP, STMicroelectronics, and Texas Instruments. Nvidia accounts for only 3% to 4%. Computing power per unit cost will continue to improve, so over time, more and more cars will have the computing power required for fully autonomous driving—although the additional cost of lidar and radar will slow this progress. Features currently available on high-end cars—partial automation/driver assistance—will become standard on every car by 2035. By 2035, high-end cars will have reached Level 3 autonomy, capable of operating autonomously in specific situations, such as controlled-traffic highways and clear weather. Waymo-level autonomous driving will remain a niche market in 2035, primarily used for robotaxis.

But children born today may not need to take a driving test by 2040.

Application-specific robots are the future of the next decade

Industrial robots are nothing new. The image below shows a robotic arm from Kuka Robotics, used to assemble solar panels at Nanosolar. I briefly ran the company (to help my venture capital firm, which invested in Rambus). Note that these robotic arms are housed in cages, as striking a person with the rotating arm could be fatal.

The International Federation of Robotics (IFR) states that there are already 4.2 million industrial robots worldwide. By 2023, nearly 400,000 new robots will be installed worldwide, with Asia accounting for 70%, Europe for 17%, and the Americas for only 10%.

There are numerous online predictions for the industrial robot market by 2025, ranging from $38 billion to $55 billion. This implies a cost of approximately $100,000 per robot. The largest industrial robot manufacturers include Mitsubishi Electric, ABB, FANUC, Kuka, Yaskawa, and Kawasaki. These companies are all based in Japan or Europe. Industrial robots/factory automation are used to reduce costs, remove humans from hazardous environments, and improve process controllability (output/specifications).

The International Federation of Robotics (IFR) announced that the global automotive industry will deploy one million robots by 2023. The five countries with the largest industrial robot deployments are China (by far the largest), Japan, the United States, South Korea, and Germany. China, with its largest domestic market, is expected to become the largest manufacturer of industrial robots.

Future Market Insights predicts that the industrial robot market will grow to $291 billion by 2035. Furthermore, there is a market for non-industrial robots. The surgical robot market is worth approximately $10 billion annually, but it's very expensive, costing between $500,000 and $2 million per system. Self-driving cars are another example. By 2035, the robotics market is projected to reach around $350 billion.

What impact might this have on semiconductor sales? We'll estimate annual production and the semiconductor content per device.

Perhaps the largest single user of robotics is Amazon, which recently announced it has deployed 1 million robots in its logistics operations. (Its total workforce across all business units exceeds 1.5 million.) Approximately 75% of Amazon's deliveries are at least partially assisted by robots. Amazon has been involved in this area for years, acquiring Kiva Systems (founded in 2002) for $775 million in 2012. You can see some of the robots in action here.

Amazon's capital expenditures on robotics are estimated to be $7 billion to $8 billion in 2024 (Seeking Alpha), out of a total capital expenditure budget of $77 billion (primarily for data center AI).

Assuming they deploy 100,000 new robots in 2024, each robot costs $70,000. They have nine different models, but most are large, with numerous motors and mechanisms, making $70,000 per robot a reasonable price.

I know from experience that for safety and price reasons, robotics companies tend to use processors and GPUs developed for automotive applications. These processors have the safety features and certifications that end customers demand, and large companies can produce them in high volumes at competitive prices. High-end cars are equipped with seven cameras, high-resolution vision AI, numerous motor controllers, and more. This is similar to the equipment used in typical robots, which cost around $2,000 each.

Therefore, if the total robotics market revenue in 2035 is $376 billion and the robot price is approximately $100,000, then the number of robots sold will reach 3.5 million (a tenfold increase from 2025), with semiconductor costs per robot at $2,000, and the total robotics semiconductor market size reaching $7 billion. Even assuming that the price of robots falls to around $50,000, the average price of a car, semiconductor consumption will still reach $14 billion. (A $350 billion robotics market would be significantly smaller than the $2 trillion to $3 trillion automotive market, with similar semiconductor content per robot.) By 2035, the total semiconductor market size will exceed $1.5 trillion. Therefore, the robotics market will remain small for the next decade.

NVIDIA's sales by segment for fiscal year 2025 (ending January 2025) are as follows:

• Data Center 88%

• Gaming 9%

• Professional Visualization >1%

• Automotive (and Robotics) >1% (approximately $2 billion/year)

• Other <1%

NVIDIA's automotive sales account for a relatively small share of the automotive semiconductor market ($2 billion/year, 4% of the $50 billion/year total), but they possess the most powerful technology. Front-facing cameras typically have the highest pixels and frame rates, processed by NVIDIA's automatically optimized GPUs; security cameras for front-facing inspections are next, while all other cameras and sensors use processors from Infineon, TI, Renesas, or NXP. NVIDIA's automotive sales also encompass robotics. Robotics applications often use automotive chips due to their safety features, high volume pricing, and availability of large suppliers.

In summary, robotics is a growing market, but it's still a long way from consuming the same amount of semiconductors used in data centers for AI, at least for the next decade. If humanoid robots reach one billion by 2050 (as predicted), and each uses semiconductors costing $1,000, this would represent a $1 trillion semiconductor market. But that's still 25 years away.

Source: Content compiled from semiengineering

Reference link: https://semiengineering.com/physical-ai-chip-sales-wont-rival-genai-anytime-soon/

View more at EASELINK