Physical AI Chip Sales Won’t Rival GenAI Anytime Soon

Fig. 1: When can we send our humanoid to pick up some milk at the store? Source: ChatGPT image

There has been a lot of talk about physical AI, much of it very optimistic. Physical AI is happening, but more slowly than many expect. It will not rival GenAI soon.

We’ll analyze humanoids, fully autonomous vehicles, then all application-specific robots (ASRs), ending with a rough estimate of the semiconductor TAM for robots from 2025 to 2035.

Morgan Stanley (May 14, 2025) predicted a humanoid robot market (let’s call this GPR: general-purpose robots) of $5 trillion by 2050, with 1 billion deployed. That would be ~1 humanoid per 10 people. They say adoption will be relatively slow until 2035, accelerating in the late 2030s.

At the May TSMC Technology Symposium in Taiwan, Wan Jui-yang, TSMC’s head of Asia/Pacific business development, said, “Following generative AI and agent AI, physical AI will become the next crucial trend…” He forecasts the global AI robot market to surpass $35 billion by 2030, with 10% of cars being autonomous by then. By 2035 he anticipates 1.3 billion AI robots deployed globally, rising to 4 billion by 2050, including 650 million humanoid robots.

Elon Musk, at the start of the year, said Tesla plans to build 10,000 Optimus robots this year, with several thousand doing useful things by year-end. The leader of the project recently left Tesla. Will Lockett is extremely skeptical, as evidenced by his recent article, titled “Tesla’s Robot is Utterly Pathetic.” Recently, Tesla opened a Diner & Drive-In in Santa Monica that, among other things, had an Optimus Robot serving popcorn. Internet reports are that it was tele-operated by a human and frozen when the link dropped. At the Drive-In, food orders are delivered by waiters on roller skates – when multiple Optimus Humanoids can deliver food on roller skates, that will be compelling.

But as economist Rudiger Dornbusch observed, “…things take longer to happen than you think they will, and then they happen faster than you thought they could.”

Humanoids are hard

Humanoids are not going to be a significant market driver in the next 10 years. Humanoids are general-purpose robots (GPRs). They are very flexible, but that makes them much harder to develop.

Bismarck Brief estimates there are 1,000 humanoid robots in the world today, mostly prototypes.

Brad Porter is the CEO of Collaborative Robotics, backed by Sequoia, Khosla & Lux. Prior to Cobot he was Amazon’s vice president for robotics, leading a team of 10,000. His article is worth reading, but the key points are:

1. AI can’t yet handle robust balancing systems. (It takes humans years to learn balance).
2. Humanoids are not the right design solution for most production tasks.
3. Wheels are 3 to 4 points of contact with the ground, and the payload inside the cone of stability is the right design solution.

Safety is just as important. GenAI makes mistakes but it’s just words. A humanoid robot capable of lifting a person is capable of killing a person if it mis-operates. Automotive and industrial applications of robotics require strict adherence to very detailed safety protocols. The level of safety testing and design required is based on the worst-case scenario. An entertainment system failure won’t kill anyone, but brake failures or flailing robot arms can.

Consider autonomous vehicles. Google (now Waymo) started testing on Mountain View streets in 2010. It took over 10 years for the first fully driverless ride on city streets. The reason is safety. Waymo can handle the 6-sigma unusual events reliably, and it has demonstrated it is much safer than a human driver based on data gathered across millions of miles and rides.

Humanoids will happen, but likely not for a decade or more. The technical challenges need to be surmounted, safety needs to be rigorously proven in extensive trials, and ASRs will remain a cheaper alternative in the meantime.

Autonomous vehicles will be the first robot most of us encounter

If you visit San Francisco (or Los Angeles, Phoenix, Austin, and soon several other US cities) you can find many Waymo vehicles on the streets. Waymo is owned by Google and has been operating for a decade, but only for a couple years with paying passengers.

Fig. 2: A Waymo I hailed in San Francisco – my initials flash on top. Very cool and safe ride.

In San Francisco, Waymo had 26% share of the ride-sharing market in April 2025, eclipsing Lyft’s market share for the first time (Source: YipitData). Uber has >50%. Waymo first launched in San Francisco, then Los Angeles, Phoenix, and now Austin. They have reached ~70 million miles driven! They are moving south from San Francisco to San Jose over the next 6 to 12 months.

The current model Waymo, shown above, is estimated to have a total cost of $140,000 using $10K of NVIDIA data-center class GPUs and even more $ in sensors (cameras, radar, lidar). The next generation Waymo, due soon, is estimated to cut costs to $85K with no loss of safety/features.

Morgan Stanley predicts Waymo will reach ~$2.5 billion of revenue in 2030 based on 1 billion miles driven across a fleet of ~17,000 Waymo vehicles. But even with Tesla included, it will account for less than 0.1% of total US miles driven. Nevertheless, Goldman Sachs points out that amounts to 7% of total US rideshare miles.

The world car market is 90 million units/year (100 million with trucks), equal to $2.1 trillion in revenue. The global automotive semiconductor market size is estimated to be $51 billion to $77 billion in 2025 (Precedence Research, Coherent Market Insights). The average semiconductor/car is ~$600. High-end cars with more cameras and features are more like $2,000.

The biggest vendors are Infineon, NXP, STMicroelectronics, and TI. NVIDIA is only a 3% to 4%. The computing power delivered per dollar will continue to improve so that over time more and more vehicles will have the compute required for fully autonomous driving – although the extra cost of lidar and radar will slow things down. What is available now on a high-end car will become standard on every car by 2035 — partial automation/driver assist. By 2035, high-end cars will be Level 3 with autonomous driving in certain circumstances, such as controlled-access highways, sunny weather, etc. Waymo-class autonomous driving will still be a niche in 2035, primarily used for robo-taxis.

But a child born today likely will not need to take a driver’s test in 2040.

Application-specific robots are the next decade

Industrial Robots are not new. Below are robotic arms from Kuka Robotics used for solar panel assembly at Nanosolar, which I briefly ran (helping out VCs that had backed me at Rambus). Note the robot arms were in cages, because if an arm was rotating and hit someone, they would be killed.

Fig. 3: Robotic arms for solar panel assembly. Source: Nanosolar, 2010

The IFR (International Federation of Robotics) said there were 4.2 million industrial robots worldwide. Almost 400,000 new robots were installed in 2023 — 70% in Asia, 17% in Europe, and just 10% in the Americas.

Fig 4: Installed robots. Source: International Federation of Robotics, Sept. 24, 2024

There are multiple forecasts on the internet for the industrial robots market for 2025, ranging from $38 billion to $55 billion. This implies each robot costs ~$100,000. The largest industrial robot manufacturers are Mitsubishi Electric, ABB, FANUC, Kuka, Yaskawa and Kawasaki. All of them are either in Japan or Europe. Industrial robots/factory automation are used to reduce cost, take humans out of dangerous situations, and increase process controllability (yield/specifications).

The International Federation of Robotics (IFR) announced in 2023 that 1 million robots were deployed in the car industry worldwide. The five largest countries deploying industrial robots are China (largest by far), Japan, the US, South Korea, and Germany. China is expected to become the largest industrial robot maker because it has the largest home market.

Future Market Insights predicts industrial robotics will grow to $291 billion in 2035. There are non-industrial robotic markets, too. Surgical robotics are ~$10 billion/year, but these are very expensive at $500,000 to $2 million per system. Autonomous vehicles are another. Perhaps the total robotics market in 2035 will be in the ~$350 billion range.

What is the potential impact for semiconductor sales? We’ll estimate units/year and semiconductor content per unit.

The largest single user of robotics is probably Amazon, which recently announced they have 1 million robots deployed in their logistics operations. (They have over 1.5 million employees in all of their businesses.) About 75% of Amazon’s deliveries are assisted at least partially by robotics. Amazon has been doing this a long time. They acquired Kiva Systems in 2012 for $775 million (Kiva started in 2002). You can see some of their robots in operation here.

It’s estimated that Amazon’s capital expenditures on robotics in 2024 were $7 billion to $8 billion (Seeking Alpha) out of a total CapEx budget of $77 billion (dominated by data center AI).

Assuming they deployed 100,000 new robots in 2024, that is $70,000/robot. They have 9 different models, but most of them are physically large with a lot of motors and mechanics, so $70K per robot is in the ballpark.

I know from experience that for reasons of safety and price, robotics companies prefer to use processors and GPUs developed for automotive applications because they have safety features and safety certification required by the end customers, and they are made in high volume at good prices by big companies. A high-end car has seven cameras, high-resolution vision AI, numerous motor controllers, etc. This is similar to what a typical robot will use, or ~$2,000/robot.

So if the total robotics market in 2035 is $376 billion in revenue, and robots sell for ~$100K, then there will be 3.5 million units sold (10X unit growth from 2025) with $2,000/robot of semiconductors, for a total robotics semiconductor market of $7 billion. Even if you assume a robot’s price declines to the average car price of ~$50K, the semi consumption would be $14 billion. (Robots at $350 billion will be a much smaller market than cars at $2 trillion to $3 trillion, with similar semiconductor content per unit.) By 2035, the total semiconductor market will be >$1.5 trillion. So robotics will be a small market segment for the coming decade.

NVIDIA’s sales by segment for FY2025 (ended January 2025) were:

• Data Center 88%
• Gaming 9%
• Professional Visualization >1%
• Automotive (& Robotics) >1% (~$2B/year)
• Other <1%

NVIDIA’s automotive sales are a relatively small share of the automotive semi market ($2 billion/year of a $50 billion/year market = 4%), but they have the most powerful technology. The front-facing camera, which often has the most pixels and the highest frame rate, is processed by an NVIDIA auto-optimized GPU, then the safety camera that checks it, and all of the other cameras and sensors are processed using Infineon, TI, Renesas, or NXP processors. NVIDIA’s automotive sales include robotics. The robotics applications generally use automotive chips for their safety features, high-volume prices, and large suppliers.

In conclusion, robotics is a growing market is a long way from matching data center AI semiconductor consumption, at least in the coming decade. If there are a billion humanoid robots by 2050, as predicted, and each uses $1,000 of semiconductors, that is a $1 trillion semiconductor market. But that’s still 25 years away.