Goldman Sachs Raises Humanoid-Robot Forecast to 6.5 Million Annual Units by 2035

Goldman Sachs has sharply upgraded its outlook for humanoid robots. In a reported 80-page “Physical AI” study released to clients on September 1, 2026, the bank raised its baseline forecast to 75,000 units in 2026, 890,000 in 2030 and roughly 6.5 million units annually by 2035. That 2035 projection corresponds to an estimated annual market of about $138 billion. 245

The figures are forecasts, not confirmed orders or shipment commitments. Public coverage of the client report is also largely secondary, so the most detailed claims should be treated as reported estimates rather than a substitute for the full Goldman Sachs document.

The revised market forecast

Goldman’s updated estimates are substantially higher than its previous baseline:

• 2026: 75,000 units, up from 51,000
• 2030: 890,000 units, up from 256,000
• 2035: approximately 6.5 million units, up from 1.38 million

The largest revision is at the end of the forecast period: the new 2035 estimate is more than four times the earlier projection. Goldman’s reported reasoning is that progress in physical AI is making humanoids more capable at the same time that manufacturing is moving toward commercial scale. 256

Why Goldman expects faster progress

Physical AI refers to AI systems that can perceive their surroundings, reason about physical tasks and act in the real world. For humanoid robots, that means moving beyond fixed, repetitive automation toward machines that can handle a wider range of warehouse or factory tasks.

The reported outlook combines three developments:

1. Improving robot intelligence. Better models could reduce the amount of manual engineering needed to program every movement and task.
2. Falling production costs. Goldman’s reported assumptions include average bill-of-materials costs declining by about 7% a year and average selling prices declining by about 6% annually from 2025 to 2035.
3. Commercial-scale manufacturing. Higher production volumes could improve affordability and make robots more competitive with human labor in suitable applications.

Under the reported forecast, the average humanoid selling price would fall from about $41,800 in 2025 to approximately $21,300 in 2035. 36

That does not mean every task will immediately be economical. Goldman’s reported analysis still identifies limited real-world data, insufficient general-purpose autonomy, reliability issues and high costs as major obstacles to broad deployment. Robotics systems must learn from physical interactions involving forces, motion and objects—data that is harder to gather at scale than text or images. 318

Warehouses are expected to come first

Goldman identifies e-commerce warehousing and logistics as the earliest large-scale setting for humanoid robots, followed by automotive manufacturing. Warehouses offer a controlled environment while still containing enough variation for flexible machines to be useful across multiple tasks. 25

Amazon and Walmart are the most prominent early names in the reported thesis. Existing warehouse automation could provide the infrastructure, operational data and economic incentive needed to test more general-purpose robots alongside specialized machines.

Goldman reportedly estimates that Amazon’s broader automation efforts could generate approximately $72 billion in cumulative service-cost savings between 2026 and 2030. In an upside scenario, that could lift Amazon’s EBIT margin by about 240 basis points, equivalent to roughly 5.6% leverage on total service costs. These are scenario estimates, not company guidance. 33335

Automotive manufacturing is the next major battleground

Automotive factories are viewed as the next major adoption setting because automakers already operate large-scale production systems and have experience with robotics, manufacturing engineering and supply-chain integration.

The companies cited in public summaries include Tesla, Hyundai, Toyota, Honda and Mitsubishi Motors. The automotive sector may also have an advantage as a potential producer of humanoids, not merely a customer. Goldman reportedly estimates that Toyota could manufacture between 190,000 and 540,000 humanoid robots in 2035. Against the projected global annual market of 6.5 million units, that would represent approximately 3% to 8% of worldwide volume. 38

The estimate describes potential production capacity, not a confirmed Toyota commitment to build that number of robots.

Each robot could expand semiconductor demand

The humanoid opportunity extends beyond robot manufacturers. Goldman reportedly estimates that each unit could represent between $3,000 and more than $6,000 in semiconductor content. The categories include high-performance computing modules, analog and mixed-signal devices, and edge storage. 17

That content reflects the demands of physical AI: robots need computing for inference, sensors and control systems for perception and movement, and memory and connectivity to coordinate actions in real-world environments. The exact mix will vary by design, capability and production model, so the figure should be read as an addressable-content estimate rather than a universal bill of materials.

Software-defined control could reshape industrial automation

The reported thesis also points to a shift from hardware-bound programmable logic controllers toward software-defined or virtual PLC architectures. If factory control becomes more software-centric, value could move toward flexible orchestration, AI models and general-purpose computing rather than remaining concentrated in proprietary hardware-and-software systems. 17

Public summaries do not provide enough reliable detail to quantify the virtual-PLC market or identify a definitive list of incumbent companies Goldman believes will be disrupted. The defensible implication is broader: established industrial-automation and factory-control vendors may face pressure if customers increasingly favor interoperable, software-defined systems.

What executives said about adoption

Goldman’s reported “Executive Perspectives” survey found that about 40% of executives expected at least 10% of workflows to be automated by general-purpose robots within three to five years. 35

That result signals meaningful corporate interest, but it is not evidence that the forecasted deployments will occur on schedule. Adoption will depend on whether robots can perform specific tasks safely, reliably and at a cost that beats existing labor or automation alternatives.

The practical takeaway

Goldman’s revised outlook is best understood as a stronger economic and technology case for industrial humanoids—not as proof that general-purpose robots are ready for every workplace. The expected sequence is clear: logistics and warehousing first, automotive production next, with semiconductor suppliers and industrial-software providers as additional beneficiaries or potential challengers.

The central test is whether physical AI can convert increasingly capable prototypes into dependable machines that operate profitably in the real world. Until the industry resolves the data, dexterity, reliability and cost constraints, the 6.5-million-unit 2035 forecast remains an ambitious scenario rather than a settled outcome. 318