Goldman Sachs — a forecast four times higher

Goldman Sachs has dramatically revised its estimates for the humanoid robot market, raising its 2035 forecast from 1.38 million to 6.5 million units delivered annually, corresponding to an estimated $138 billion market, nearly four times larger than the previous $38 billion valuation, according to a research report titled “Physical AI,” published by analyst Eric Sheridan. The bank also raised its interim milestones: the 2030 forecast increased from 256,000 to 890,000 units, while the forecast for the current year rose from 51,000 to 75,000.

Humanoid robots have started looking for jobs. Sheridan identified warehousing and logistics as the main early deployment scenario. Amazon has already installed more than one million robots (these are not humanoid robots) across more than 300 locations, and Goldman Sachs estimates that automation could save the company approximately $72 billion in service costs by 2030. The automotive industry is seen as the next major wave, with Tesla, Hyundai and Toyota, the latter estimated to produce between 190,000 and 540,000 humanoid units by 2035. Each humanoid robot is expected to generate approximately $3,000–$6,000 in semiconductor demand, a market that could reshape the balance of power between traditional industrial automation suppliers such as Siemens, Rockwell Automation and Schneider Electric, and new entrants specializing in robot software.

Musk: one billion robots in ten years

If the Goldman Sachs forecast already seemed bold, it was overshadowed at the beginning of the month by a statement from Elon Musk. Speaking remotely at the G20 Ministerial Meeting on Innovation in Chapel Hill on September 1, Musk said that the world would have “at least one billion” humanoid robots within ten years, each producing five times as much as a human employee, a figure he described as “a conservative bet that I would bet serious money on,” according to CNBC, which reported from the event. He also predicted that artificial intelligence could increase global economic output by 20 to 30 percent annually (equivalent to $20–30 trillion) and warned of an energy shortfall of around 15 gigawatts that could affect data centers starting in 2027.

Musk’s statements came, not coincidentally, amid protests taking place near the summit venue, where demonstrators expressed concerns about the impact of artificial intelligence on jobs.

The contrast is relevant: while the Tesla leader talks about robots that will produce more than all of humanity combined, on the ground the technology remains far from being able to perform complex tasks without constant human supervision. Humanoid robots can walk, run or imitate human movements. But what matters is whether they can turn these demonstrations into autonomous, repeatable work that is inexpensive enough to make economic sense.

The reality inside Chinese factories: “dance disguised as work”

At a training center in Liuzhou, in southern China, more than one hundred humanoid robots stand in orderly rows, guided by human instructors through elementary tasks such as sorting crates or making coffee. According to an extensive Reuters investigation, the machines remain slow and clumsy: an inexperienced instructor might obtain just one usable movement out of 300 attempts, while even an experienced one needs around 50 attempts for the same result. The robots perform simple tasks at only about 20% of the speed of a human worker.

China nevertheless dominates global production, accounting for 95% of the approximately 20,000 units delivered last year, according to BofA Global Research, while the Ministry of Industry and Information Technology estimates that production will exceed 100,000 units this year. The problem, executives and researchers in the industry say, is not hardware but intelligence: robots struggle with any situation that requires intuition or deviation from a programmed routine. “The robots’ IQ is too low. Much of what we see is dance disguised as work,” summarized Tang Wenbin, CEO of startup Yuanli Lingji. Counterpoint Research data confirm this gap: entertainment and performance accounted for 33.6% of humanoid robot deliveries in the first half of 2026, while smart manufacturing barely reached 12.8%, and warehousing and logistics just 4.9%.

Beijing’s approach follows the same industrial-policy pattern previously used for electric vehicles and solar panels: massive subsidies, rapid capacity building and cost reduction through competition. The essential difference is that, unlike those already mature industries, humanoid robots remain an early-stage technology, with unresolved problems in dexterity and reliability. Chinese government entities have already spent at least $230 million on humanoid purchases in the first half of 2026, compared with just $62 million a year earlier, while some industry founders already describe the sector as a potential “speculative bubble” that will, sooner or later, go through a painful consolidation. Unitree founder Wang Xingxing was similarly sober at the World Robot Conference in Beijing, estimating that the “ChatGPT moment” for embodied intelligence in robots is at least two or three years away, and perhaps even a decade.

How far have we actually come with humanoid robots?

The comparison with the classic industrial robot, a fixed robotic arm programmed rigorously for a single repetitive task, puts things into perspective. A traditional industrial robot is still considerably faster, more precise and more reliable than any humanoid for well-defined tasks, which is why it remains the standard on automotive assembly lines or in body welding. The advantage of humanoid robots lies not in raw speed or precision, but in versatility: theoretically, they can move through spaces designed for humans, handle a variety of objects and be reassigned between different tasks without the costly reconfiguration of an entire production line. This is why Amazon and Walmart have placed major bets on them in warehouses, where the variety of tasks exceeds what traditional fixed automation can efficiently handle.

Progress compared with the previous generation is nevertheless real and measurable, even if it remains insufficient for general-purpose work. A concrete example comes from China itself: the TianGong Ultra humanoid robot ran 100 meters in 8.64 seconds at the World Humanoid Robot Games in Beijing in August (faster than Usain Bolt’s human record) and is now being tested, in a new version, at an engine factory in Beijing, where it moves 8–12 kilogram boxes in controlled repetitive tasks. As I wrote extensively in the Aidapted article “The Great Champion and World Record Holder Now Works in a Factory in China,” the team behind the robot emphasized that the real challenge was not sprinting speed, but the transition “from high performance to high reliability and usability.” It is the same gap described by the Reuters report from Liuzhou, between an impressive spectacle and economic utility. The difference from a few years ago is that humanoid robots can now maintain dynamic balance at high speeds, recover from minor disturbances and execute complex motor sequences, capabilities that were almost nonexistent in generations before 2023. Nevertheless, they remain far behind a simple robotic arm when it comes to perfect repeatability and consistent speed on strictly defined tasks.

A piece of the larger puzzle of AI fears

The contrast between financial enthusiasm for robots and technical skepticism on factory floors is not isolated; it fits into a broader pattern that I have documented in other recent analyses. During the same period when Goldman Sachs and Musk were talking about billions of robots and trillions of dollars in additional productivity, leaders in the artificial intelligence industry were debating the opposite scenario: the need for a deliberate pause in frontier AI development, out of concern that capability progress is already outpacing the pace of safety measures (“Frontier AI Is Being Asked to Slow Down, But No One Wants to Be the First to Hit the Brakes” — Aidapted). Almost simultaneously, China is building its own regulatory barriers against a scenario involving loss of control over artificial intelligence (“China Accuses the US of AI Hypocrisy While It Builds Its Own Barriers” — Aidapted).

The two stories — physical robots that can barely keep up with elementary warehouse tasks, and language models that some warn could surpass the combined intelligence of humanity within a decade — may seem to come from different worlds, but they are actually two sides of the same coin. The “physical intelligence” on which humanoid robots rely depends directly on the progress of the AI models that control their movements and decisions; if warnings about excessively rapid AI development are well founded, they could accelerate precisely the component that humanoid robots currently lack: intelligence. And if the AI industry actually slows down, as some of its leaders are calling for, Musk’s projection of one billion robots in ten years could remain a distant horizon.

Between a financial bet and operational reality

Two stories therefore remain true at the same time: capital markets are already treating humanoid robots as the next major investment theme, and the sums being discussed, from Goldman Sachs’ $138 billion to the trillions envisioned by Musk, are large enough to change the strategies of entire industrial giants. At the same time, reality remains, for now, much more modest: machines moving at one-quarter of a human’s speed and requiring hundreds of attempts just to learn a single useful movement.

Sources: