Humanoid robots: bet on content per robot, not unit forecasts

Goldman says 250,000 humanoids ship in 2030. Bank of America says 1.2 million. That 5x gap between serious desks is the whole argument for owning what goes inside the robot instead.

Goldman Sachs expects more than 250,000 humanoid robots to ship in 2030. Bank of America expects 1.2 million in the same year. That is a 5x disagreement between two serious research desks on a single number, and it tells you the unit forecasts are not something to build a position around. The figure that behaves better is how much content goes into each robot, whatever the count turns out to be.

The banks are not close to agreeing

Humanoid robots have almost no shipment history to extrapolate from. Every forecast is therefore a judgement call about when a technology starts working, not a projection of an existing trend. That shows up as enormous dispersion.

Two caveats on that table. The Deutsche Bank 2030 figure comes from trade press coverage of a June 16, 2026 note by Iris Zheng rather than a Deutsche Bank document we could open directly, so treat it as secondary sourced. And a figure of 3,030 humanoid units in 2024, widely shared on social media, does not survive checking, so it is not here.

Even the near term numbers move violently. Deutsche Bank raised its own 2026 estimate from 17,500 units to roughly 50,000 in a single revision. Goldman raised its 2035 market size 6x, from $6 billion to $38 billion, citing a roughly 40 percent drop in build costs. A forecast that moves 6x is a forecast telling you it does not know.

What a chip company actually sells into a robot

Here is the number that behaves better, because it comes from a company quoting its own product rather than modelling someone else's demand.

On the Analog Devices Q3 FY2025 earnings call on August 20, 2025, chair and chief executive Vincent Roche said:

"Our content in a humanoid robot is likely to be several thousands of dollars, that's basically a 10x increase over the content in today's cutting-edge AMRs."

The reason he gave is the useful part. Roche pointed to the explosion in sensor and actuator counts. A humanoid has 30 or more joints, and every one needs position sensing, precision motor control and power management. An autonomous mobile robot, the wheeled kind already working in warehouses, has a handful of motors and drives around a flat floor. The humanoid has to hold a wrist steady while carrying weight.

More joints means more sensing. More sensing means more chips. That relationship holds whether the world builds 250,000 robots or 1.2 million.

Carry one caveat into any model built on it. "Several thousands of dollars" is Analog Devices' own content, not the total semiconductor content of a robot, and it is a forward looking company estimate rather than a measurement of shipped product. Roche is talking his own book.

Build cost is the best evidenced part of the story

Bank of America puts a China built humanoid bill of materials at $35,000 in 2025, falling below $17,000 by 2030. Morgan Stanley's price path runs from roughly $200,000 per unit in 2024 to about $150,000 by 2028 and roughly $50,000 by 2050. Goldman's forecast raise was driven by a 40 percent cut in build costs.

Different firms, same direction. This is the one place the evidence is consistent.

Cost decline matters more than it sounds. Robots do not get adopted when they become capable. They get adopted when they become cheaper than the labour they replace. Every $1,000 off the bill of materials moves another tier of work into range.

There is a catch. Falling build cost is good for robot buyers and awkward for robot sellers' margins. Component suppliers with rising content per unit sit on a different side of that trade than assemblers competing on price.

Physical AI is an inference demand story

Robots do not think once. They have to run a model of the world continuously to move through it, which turns each deployed robot into a recurring compute customer rather than a one time hardware sale. Nathan Labenz put the mechanics plainly on The Cognitive Revolution on July 9, 2026, describing a robot that "will need to create like this simulation 30 times a second," and the volume of tokens that burns.

The demand side is already climbing for other reasons. Bargo's Token Demand Index sat at 248.9 on July 19, 2026, up 32 percent in 30 days and up 149 percent since the series began on May 6, 2026. Weekly token volume runs at 58.5 trillion. The effective blended price has fallen to $1.97 per million tokens as usage shifts toward open source models, now 53.6 percent of volume.

That combination is the important part: demand rising while the per token price falls means the market is expanding, not that margins are collapsing.

On the supply side, the Compute Tightness Index read 47.5 on July 20, 2026, a balanced regime, up 2.1 points over 30 days and tightening. H200 capacity is the tight spot at 66.6 while H100 has loosened to 38.7.

None of that is robot demand yet. It is the pipe robot demand would flow through, and it is already filling.

The strongest argument against the whole thesis

Jürgen Schmidhuber, one of the founding figures of modern deep learning, argued on the Unsupervised Learning podcast on July 9, 2026 that robot hardware "is really inferior compared to human bodies" and that closing the gap "will take another few decades maybe."

He extended it to the money, arguing that investors putting "a thousand billion dollars into GPUs for data centers today within the next five years they are going to lose 900 billion dollars."

Take him seriously for one reason above all: he has no robot company to promote. Roche sells chips into robots. Tesla sells the robot. The warehouse automation incumbents have their own reasons to say humanoids will not work. Schmidhuber has none of those incentives, and he is the bear.

There is a quieter version of the same warning in the forecast table. When credible desks differ by 5x on the same year, the honest reading is that the error bars are wider than most positions assume.

Worth holding alongside it: the robotics businesses making real money today are not humanoids. Intuitive Surgical runs a profitable surgical robot business at scale, and it still fell 14.1 percent to $345.58 on July 17, 2026 after its Q2 report, despite US procedure growth the company described as healthy. Even good robotics numbers get sold in a risk off tape.

What to watch

• Content per robot, not unit counts. It rises with joint count regardless of how many robots ship.
• Bill of materials disclosures. Falling build cost is the most consistently evidenced part of the story and the clearest adoption trigger.
• Token demand and compute tightness as leading indicators of whether physical AI is actually consuming inference.
• Who is talking their own book. The chip executive quoted here is; the sharpest sceptic is not.
• The gap between humanoid optionality and robotics that already earns money. Surgical and logistics robotics trade on entirely different drivers.

This article is informational and is not investment advice.

Sources

• Goldman Sachs, The global market for humanoid robots could reach $38 billion by 2035 — goldmansachs.com
• Morgan Stanley, Humanoid robot market expected to reach $5 trillion by 2050, April 2025 — morganstanley.com
• Bank of America Institute, Humanoid Robots, April 29, 2025 — bankofamerica.com
• Bank of America Institute, Physical AI Part 2, March 12, 2026 — bankofamerica.com
• Analog Devices Q3 FY2025 earnings call transcript, August 20, 2025 — fool.com
• Analog Devices Q3 FY2025 results release — investor.analog.com
• Morgan Stanley, The Humanoid 100: mapping the humanoid robot value chain — morganstanley.com
• Bargo Token Demand Index and Compute Tightness Index, July 2026 — Bargo Token Demand Index

More research at bargo.ai/research.