Data point

Robots installed or shipped per year, by category (units)

Comparison of actual global industrial robot installations in 2024 against Tesla Optimus's stated 2025 humanoid production target, its estimated 2025 actual output, and Goldman Sachs's 2030 base-case forecast for global humanoid shipments. Units are physical robots installed or shipped in the stated year.

Source: International Federation of Robotics, World Robotics 2025 - Industrial robot figure from IFR World Robotics 2025. Optimus target and shortfall reporting from Tesla's Q4 2024 earnings call as covered by Tech Times. Goldman Sachs 2030 base case from Goldman Sachs Research. The 2025 Optimus actual is an estimate since Tesla has not disclosed an exact unit count; it is shown to illustrate order of magnitude, not a precise figure.

Nvidia CEO Jensen Huang used a CES 2025 keynote to declare “physical AI” the next frontier: AI that can “perceive, reason, plan and act” in the physical world, rather than only producing text or images. Nvidia’s own glossary defines it the same way: AI combined with robotics, sensors and actuators so a system can sense an environment, decide, and take physical action. McKinsey uses a near-identical term, “embodied AI,” built around vision-language-action (VLA) foundation models. The label is new. The underlying idea, machines that sense and act in the world, is not; it is a rebranding of decades of robotics and embodied-cognition research for the current AI investment cycle.

What already exists, at scale. The International Federation of Robotics reports 542,000 industrial robots installed globally in 2024, the fourth straight year above 500,000, with China alone accounting for 54% of new deployments. This is the real, non-humanoid base case: fixed-purpose industrial automation has been compounding for years and does not depend on any “physical AI” breakthrough to keep growing.

What is still a forecast. Humanoid robots are a different, much earlier-stage category. Goldman Sachs Research now projects a $38 billion humanoid robot market by 2035 and more than 250,000 humanoid shipments by 2030 in its base case. That $38 billion figure is itself a sixfold revision of Goldman’s own prior $6 billion estimate, made inside about two years, which says as much about the volatility of these forecasts as it does about the market.

Where the gap shows up concretely. Tesla’s Optimus is the most visible humanoid program and also the clearest case study in target-versus-delivery. On Tesla’s Q4 2024 earnings call, Elon Musk targeted roughly 5,000 Optimus units for internal factory use in 2025. Reporting into mid-2026 indicates actual 2025 output reached only a few hundred units, under 10% of the stated target, and Musk has since declined to give a firm 2026 production number while still describing eventual production lines rated at millions of units per year. A target missed by more than 90% in one year, followed by a larger unquantified claim for the following years, is a pattern worth naming plainly rather than smoothing over.

For anyone thinking about exposure to this theme, the practical distinction is between two very different things: proven industrial robot deployment that is already large and growing on its own trend line, and humanoid-specific bets that are priced substantially on analyst forecasts and executive statements that have already moved sharply and missed badly within the same forecast horizon. This is a framework for thinking about the space, not a recommendation to buy or avoid any security.

Countercase

Several weaknesses limit how much confidence this note can support. First, selection bias: Optimus is used here as the headline example of humanoid delivery risk because it is the most publicly documented case, but that also means it is not necessarily representative. Other humanoid programs, including Chinese suppliers and firms like Agility Robotics and Figure AI, have less public reporting, so their own target-versus-actual gaps (or lack of gaps) cannot be verified from public sources in this note. Second, narrative risk: “physical AI” is a compelling story precisely because it promises AI leaving the chatbot and entering the physical world, which is a strong rhetorical hook independent of whether the underlying technical shift is as large as the label implies; framing it as a rebrand does not prove the technology is stalled, only that the label should not be mistaken for evidence of progress. Third, forecast reliability: every forward-looking number in this note, the $38 billion 2035 figure and the 250,000-unit 2030 base case, is an analyst projection, not a realized outcome, and Goldman’s own six-fold upward revision in two years shows these projections can move a long way before any of the underlying units actually ship. Readers should treat the industrial robot numbers as the only hard, already-realized data point here, and treat every humanoid number as provisional.

Sources