Why humanoid forecasts disagree: the 23x spread, computed and decomposed

Two serious research houses have published humanoid shipment forecasts for 2030. Their numbers differ by about 23 times. Omdia said 38,000 units, in July 2024. Goldman Sachs said 890,000 units, in September 2026. Neither number is a typo. This page computes the spread in the open, then explains it. The two houses count different things. They start from different 2025 actuals. They treat China differently. And they build on assumption sets that disagree more than their conclusions do. A fifth factor, forecast vintage, explains as much of the gap as any of the four.

The spread, computed in the open

Here is the division, symbol and number both shown, so nothing rides on trust:

spread = Goldman Sachs 2030 units / Omdia 2030 units
spread = 890,000 / 38,000
spread = 23.4

Both operands are real, dated sources. Omdia's press release, London, July 25, 2024, put global humanoid robot shipments above 10,000 units by 2027. It saw 38,000 units in 2030, an 83 percent compound growth rate from 2024. Goldman Sachs, in its September 2026 revision, put 890,000 units in 2030, up from 256,000. The same revision lifted 2035 to 6.48 million units shipped in 2035, worth 138 billion dollars.

For the year 2030, the widest named pair runs wider still. Bank of America said 1.2 million units for 2030 in March 2026. Against Omdia's 38,000, that is a spread of more than 30 times between named houses.

Dollars disagree too, for the same target year 2035:

Divide the ends and the 2035 spread in dollars is 300 / 15 = 20 times. Every row above is a serious publication. That is the puzzle this page takes apart.

Same vintage, same market, still 6 or 7 times apart

The 23x pair is 26 months apart in publication date. So part of the spread is just time passing. Hold the date still and the disagreement shrinks. It does not go away.

Goldman Sachs, in its February 27, 2024 research note, put the 2030 base case at more than 250,000 units. Omdia, five months later, said 38,000 for the same year. Two houses, the same summer, both on the record:

same-vintage spread = 250,000 / 38,000 = 6.6

Then one house moved. Goldman revised 2030 from 256,000 to 890,000, a 3.5 times lift, inside its September 2026 update:

full spread = 6.6 x 3.5 = 23

The 23x spread is therefore not one disagreement. It is a same-vintage disagreement of about 6.6 times, multiplied by one house revising its own number 3.5 times. Keep that decomposition in mind. It turns a shouting match into two separate questions. Why did serious houses disagree by almost 7 times in mid-2024? And why did one of them lift its number more than three times over after that?

Reason 1: they count different things

The biggest single error in reading these forecasts is stacking numbers that measure different quantities in one column. The table that follows separates them.

A shipment number counts machines sold in one year. An installed base number counts every machine still standing, across all years. Morgan Stanley's headline, about 1 billion humanoids in use by 2050 with about 90 percent doing industrial and commercial work, is a population count. Goldman's 6.48 million for 2035 is an annual sales count. Press coverage routinely places them side by side. They are not comparable rows.

The dollar figures split by scope, not just size. Morgan Stanley's 5 trillion dollar figure for 2050. In the bank's own words, it includes "sales from supply chains and networks for repair, maintenance and support". Interact Analysis counts commercial hardware shipments. The J.P. Morgan Private Bank number for 2035, roughly 300 billion dollars against about 2 billion in 2025, is a sales measure. Three different nets, three different catches.

Reason 2: they start from different actuals

Every forecast is a growth curve pinned to a base year. Change the base and you change the whole curve. The 2025 base itself is disputed.

All three rows describe the same year. The ratio between the top and bottom rows is 2.6 times before any forecasting happens. Pin an 890,000 target for 2030 to each base in turn.

implied growth = 890,000 / 13,000 = 68 times in five years
implied growth = 890,000 / 7,000 = 127 times in five years

A 68 times lift in five years means shipments must a little more than double every year. That is five years running. The IFR's September 2026 reading of the same market is the sober pole. Most current applications remain specialized and often need a human teleoperator. The lack of a strong business case, above all in industry, still limits adoption.

The near-term check is already on the calendar. Goldman's September 2026 revision put 2026 at 75,000 units, up from its own 51,000 earlier estimate. Against Omdia's counted 13,000 for 2025, that is roughly a 6 times year. The counted answer lands in early 2027. It will grade every curve in this page. For the counted numbers themselves, see the companion pages on the 2025 actuals and the 2026 test year. This page stays on the why.

Reason 3: China sits inside every number

There is no global humanoid market to forecast yet. There is a Chinese production surge with a much smaller rest-of-world market around it. Houses fold the two differently.

The J.P. Morgan Private Bank. On August 14, 2026, it put China at roughly 75 percent of both the humanoid and the autonomous vehicle markets. The same piece reports the three leading Chinese makers. They are on track to build more than 100,000 units in 2026. That is about five times their 2025 output. Interact Analysis projects China above 65 percent of real-world application shipments by 2035. China and the United States together sit above 85 percent of demand. Barclays, in its May 19, 2026 study, models up to 24 million humanoids in China's workforce by 2035. That is about 4 percent of the country's labor force.

Two consequences follow. First, a global units forecast is mostly a China production forecast wearing a global label. Second, the cost side is regional too. The same J.P. Morgan piece prices the 2025 parts cost of one humanoid in US dollars: 45,500 in China, and 131,800 outside it. At around a million units a year, the China figure falls to 16,300 by 2034. A house that assumes the Chinese cost curve goes global. It builds a very different 2030 than one that does not.

Reason 4: the assumption sets disagree more than the conclusions

Under every headline number sits a small model of the world. The models barely overlap.

Bank of America's March 2026 model assumes humanoids replace 20 percent of the industrial workforce and 50 percent of the service workforce. One robot does the work of 2.5 industrial workers or 1.5 service workers. At the far end sit about 0.7 units per household. Its 2060 ceiling, 3 billion humanoids in operation by 2060, is roughly 80 percent of everyone on Earth who works, against a world labor force of about 3.74 billion people, the World Bank figure for 2025. Citi's December 2024 path to 648 million humanoids by 2050 runs on regional rates. It assumes 30, 20, and 10 percent of industrial jobs by development tier. For households it assumes 22, 11, and 5 percent. Morgan Stanley prices its curve in US dollars per unit: about 200,000 in 2024, about 150,000 by 2028, and 50,000 by 2050. United States household penetration lands between 3 and 33 percent, depending on income.

The price assumptions alone span a multiple of ten. Goldman's February 2024 note recorded a 40 percent cost drop inside a year. It went from a range of 50,000 to 250,000 US dollars per unit in 2023 to 30,000 to 150,000 in early 2024. Goldman's September 2026 numbers imply a blended 138 billion dollars over 6.48 million units, about 21,000 per machine. Real deals sat far higher. UBTECH booked about 113 million dollars of humanoid revenue on 1,079 units delivered in 2025. That is roughly 105,000 per robot. Same industry, same years, a five times spread on price per unit. Every adoption curve is a payback sum over a price. The unit forecasts inherit the disagreement.

Forecasts move: the vintage problem

A forecast's publication date belongs in every sentence that quotes it. These numbers are not standing still.

Goldman Sachs, between February 2024 and September 2026, moved 2030 from more than 250,000 to 890,000. It moved 2035 from 1.4 million to 6.48 million. It moved 2026 from 51,000 to 75,000. Its own explanation for the first big lift was blunt: "AI progress surprised us the most." Bank of America raised its 2030 to 2035 run rate from 400,000 to 1 million units between its March 2025 and March 2026 vintages. MarketsandMarkets replaced its April 2025 edition, 15.26 billion dollars by 2030, with a July 2026 edition, 50.27 billion by 2035, at the same URL.

Revision is not always upward. Interact Analysis cut its mobile robot shipment growth forecast from 6 percent to 4 percent a year for 2025 to 2030. The cut followed the warehouse automation downturn. McKinsey pushed its expectation for urban autonomous driving pilots from 2030 to 2032 in January 2026. Forecast error in robotics runs in both directions. That is why a spread between a July 2024 number and a September 2026 number is not proof that one house is careless.

The honest summary of the September 2026 Goldman revision came from trade coverage, not from a rival bank. It called the number a forecast, not an order book. Neither of the reports behind it identifies confirmed customer orders near the projected volumes.

How to read any humanoid forecast

Run every headline number through six questions before you quote it.

A number that survives all six questions can be quoted with its date. Most headlines fail question one.

What the disagreement means if you build robots

The unit forecasts disagree wildly. Almost every house agrees on one structural thing: a software layer sits inside the machine market. It scales with units. Goldman's own robotaxi work, April 2026, splits a roughly 2 trillion dollar 2035 autonomous vehicle market into about 300 billion dollars of AI software and subscriptions, a 15 percent share. Run the same share on Goldman's own 138 billion dollar humanoid market for 2035 and the software layer comes out near 20 billion dollars a year. Our bottom-up cost runs land in the same place by a different road. The robot AI software layer in 2035 is a 10 to 30 billion dollar per year market. That holds whether 2030 lands near Omdia's 38,000 or Goldman's 890,000.

That layer is where this site works. Every unit in every table above needs a policy, a brain that turns cameras into motor commands. It also needs the data to train it. A pi 0.5 class policy runs a full manipulation loop. In our own October 2026 cost runs, serving a pi 0.5 class policy for one robot working one hour cost about 6.48. The cost per robot hour page shows that math in full. The per-robot-hour glossary defines the unit. The entry point costs less still. An ACT policy trains on a used desktop card in about an hour. The tier 1 builder budget guide prices that path from zero dollars up.

So use the spread, do not fear it. The low pole tells you what a hardware tracker will back today. The high pole tells you what a bank's growth model prices in if cost curves cooperate. Your build plan lives between them. Your serving costs are knowable to the cent today. Live rates sit on the pricing page. You can open an account at signup.

Frequently asked questions

Which 2030 humanoid forecast is right? Nobody can know yet. Omdia's 38,000 is a hardware tracker's view from July 2024. Goldman's 890,000 is a September 2026 model that its own authors frame as a high growth scenario. The counted shipments of 2027 through 2030 will grade both. The nearest checkpoint is Goldman's own 75,000 for 2026, testable against the full-year counts in early 2027.

Is Goldman Sachs or Omdia more credible here? They do different jobs. Omdia tracks hardware shipments bottom up and published its 38,000 with an explicit warning. The technology remains in its infancy, it wrote. Large-scale implementation is a few years away. Goldman builds scenario models around cost curves and AI progress. It has already lifted its 2030 number 3.5 times in 30 months. Credibility is the wrong axis. Read each with its method and its date.

Why do trillion dollar numbers appear next to 15 billion dollar ones? Scope and horizon. Morgan Stanley's 5 trillion is a 2050 ecosystem figure. It includes supply chains, repair, and support. Interact Analysis's 15 billion is 2035 commercial hardware revenue. Neither is wrong. They measure different nets over different years. Neither is a unit count.

Do any robot forecasts go down? Yes, and that matters for calibration. Interact Analysis cut mobile robot shipment growth from 6 percent to 4 percent a year for 2025 to 2030. McKinsey pushed urban autonomous driving pilots out two years in January 2026. Revision direction in robotics runs both ways.

Last verified: 2026-10-26