Humanoid Market Forecasts Keep Revising Upward While the Fielded Fleet Barely Moves

Goldman Sachs Multiplied Its Own 2035 Forecast by Six and Left the Year Alone

Sometime between an earlier, unpublicized estimate and its February 2024 report, Goldman Sachs Research decided the global humanoid robot market would be worth roughly six times what its own prior model had said — and kept the target year exactly where it was. The February 2024 figure is specific: a $38 billion total addressable market and 1.4 million annual humanoid shipments by 2035, a number a later industry compilation of the report describes plainly: Goldman Sachs “lifted its 2035 TAM sixfold in a single revision after AI progress and cost declines beat its model” ^[1]. Six times the market size, same decade-out deadline, and nothing about how many humanoid robots are actually walking around a factory floor changed on the day that revision was published, because a revision to a spreadsheet and a robot coming off an assembly line are not the same event. They are not even measured in the same units, which turns out to be most of this piece’s argument.

Goldman is not alone, and it is not even the most recent instance. Morgan Stanley’s 2025 report, “The Humanoid Economy,” puts the market at $5 trillion by 2050 ^[1]. Citi Institute, a year earlier, had already put a number on the same eventual outcome: a market growing “from zero today to $7 trillion by 2050” ^{[2] [1]}. And MarketsandMarkets — a market-research vendor whose entire business model is publishing and then updating these estimates — republished its own humanoid robot report in July 2026 with a bigger number attached: $5.41 billion in 2026, rising to $50.27 billion by 2035 at a 28.1 percent compound annual growth rate, both figures independently confirmed on the vendor’s own live report page and its own press release, dated July 6, 2026 ^{[3] [4]}. An earlier vintage of the same firm’s forecasting, preserved now only in a third party’s summary rather than on MarketsandMarkets’ own site, had put the 2025 figure at $2.92 billion rising to $15.26 billion by 2030 at a 39.2 percent CAGR ^[1] — a smaller total, a nearer horizon, and a page that no longer exists in that form, because the vendor overwrote it in place rather than publishing the new numbers alongside the old ones.

Four forecasting houses, four incompatible scopes, and one shared direction of travel. This piece is not about whether any one of those numbers is right — dollar totals for a market that does not fully exist yet are, definitionally, unfalsifiable at the time they are published. It is about a comparison none of the four reports make and no press coverage of them seems to ask for: how far ahead of the actually-disclosed, actually-verifiable 2026 humanoid fleet each of these forecasts currently sits, and whether that distance is closing or opening as newer vintages arrive. Set against the three most concretely documented 2026 humanoid programs — Figure’s production line at BMW, Agility Robotics’ warehouse fleet at GXO, and Tesla’s internal Optimus program — the distance is not closing. It may not even be measurable in a single number, because the fielded side of the comparison refuses to report itself in units a forecast can be divided against.

A Ratio Nobody Currently Publishes But Every Report Implies

Every one of the four forecasts above is, structurally, a claim about a future ratio: some dollar total or shipment count that presupposes a much larger fielded humanoid population than exists today, arrived at by some assumed growth path between now and the target year. None of the four reports states what that implied population is today, and none of them compares its own number against a verified current baseline. I want to make that comparison explicit rather than leave it embedded in an assumption.

Call it the forecast-to-fielded ratio: a cited long-horizon projection — a shipment count or a dollar total, whichever the report publishes — divided by the best available verified count of what is actually fielded, fully sourced from company disclosures rather than from another analyst’s estimate, at the time the forecast is read. Computed once, at a single point in time, the ratio is just a number. Computed across successive report vintages — as Goldman’s own sixfold revision and MarketsandMarkets’ back-to-back reports let it be, in a limited way, here — it becomes a trackable series: does the distance between the narrative and the verified deployment close as each new report arrives, or does it widen?

Two things this construct is explicitly not. It is not a claim that any of the four forecasts is wrong; an early-stage technology’s market can look wildly premature at every point along its own eventual, correct trajectory, a point this piece returns to directly below. And it is not a claim that a bigger denominator would make the ratio meaningless — a ratio of a million to one and a ratio of a thousand to one are different findings, and only computing the number, honestly, with its assumptions stated, can tell a reader which one currently holds. What follows is the attempt to compute it, and an equally honest account of where the computation runs into a wall the forecasts never had to face: a fielded side that will not report itself in a comparable unit at all.

Four Forecasting Houses, Four Different Scopes, One Direction

Line the four estimates up and the first thing that stands out is not their size but their incompatibility. Goldman’s $38 billion is a 2035 total addressable market figure paired with an explicit unit count — 1.4 million shipments a year by that date ^[1]. Morgan Stanley’s $5 trillion and Citi’s $7 trillion are both 2050 figures, fifteen years further out than Goldman’s, describing what the compilation summarizing them calls a “humanoid economy” rather than a hardware TAM — a category that, on its face, could include the value of labor humanoids perform rather than only the value of the robots sold, a scope difference the underlying reports would need to state explicitly and this piece cannot resolve without them ^[1]. MarketsandMarkets’ figures are narrower still: a hardware market-size estimate with no shipment count attached at all, currently $5.41 billion for 2026 growing to $50.27 billion by 2035 ^{[3] [4]}.

None of these four numbers can be legitimately averaged, chained, or cross-validated against one another — they are not measuring the same thing, over the same period, using the same method. That incompatibility is itself worth stating plainly rather than glossing over, because most coverage of “the humanoid market” cites whichever of these figures fits the point being made in that article, without noting that the number two paragraphs later in a competing article is describing something else entirely. A $38 billion hardware TAM and a $7 trillion economic-value estimate are not in tension with each other; they are answers to different questions wearing the same units.

What they do share, reliably, is direction. The clearest single illustration is not a comparison across firms but a comparison of MarketsandMarkets against itself. The vendor’s own live press release states its current headline as “the humanoid robot market size is estimated to be USD 5.41 billion in 2026 and is projected to reach USD 50.27 billion by 2035, expanding at a CAGR of 28.1% from 2026 to 2035,” published July 6, 2026 ^[4]. A version of the same underlying report, cited via a secondary industry compilation and no longer independently confirmable at its original address because the vendor updates the same report page in place rather than versioning it, had earlier put the number at $2.92 billion in 2025 rising to $15.26 billion by 2030 ^[1]. Two things happened in that update, not one: the near-term figure moved up (from a 2025 base of $2.92 billion to a 2026 base of $5.41 billion, an increase larger than one year of even a 39-percent compound growth path would explain on its own), and the horizon moved out (from 2030 to 2035, buying five additional years for the bigger number to be reached). A revision that simultaneously raises the number and extends the deadline is a specific kind of upward revision — one that makes the eventual figure larger while making the near-term claim less falsifiable, not more.

Reading that table by row misses the point of building it. Read down the MarketsandMarkets pair specifically: the same company, describing the same market, on the same live URL, moved its own long-run figure up by more than 3x and pushed the horizon out five years, inside roughly a year and a half — and did so without publishing anything that functions as a reconciliation, a “here is what changed since our last estimate” note of the kind Goldman’s own compilation at least gestures toward with its “lifted its 2035 TAM sixfold in a single revision after AI progress and cost declines beat its model” framing ^[1]. A forecast that revises upward because a real, verifiable event occurred — a documented cost decline, a documented capability jump — is doing something different from a forecast that revises upward with no stated cause visible to anyone reading the report after the fact. This piece cannot tell, from what MarketsandMarkets has published where it can currently be checked, which of those two things happened to its own number. That absence of a stated reason is itself the finding, not a gap in this piece’s research.

What’s Actually Running in 2026 Doesn’t Even Share a Unit With the Forecast

Set those four dollar-and-shipment projections against the three 2026 humanoid programs with the most public operational detail, and the first problem is not that the fielded numbers are small. It is that none of the three report their own scale in a form that divides cleanly against a forecast’s shipment count or dollar total.

Figure is the most transparent of the three, and its own April 29, 2026 production update is worth reading for exactly how much detail a company can disclose about a ramp without ever stating a total customer-facing fleet size. The company reports having built “over 350 of our third generation humanoid robots” as of that update, alongside a production-rate increase “from 1 Figure 03 per day to 1 per hour” — a change the company itself frames as “a 24x throughput improvement in under 120 days” ^[5]. The same update discloses “over 9,000 actuators across more than 10 distinct SKUs,” a battery line that has “shipped over 500 battery packs,” an end-of-line first-pass yield “now over 80%,” and a battery-line first-pass yield of 99.3 percent ^[5]. {{figure:cadence-dial-hand-crossing-from-day-to-hour}}

That is a genuinely detailed manufacturing disclosure — more granular, unit for unit, than anything Goldman, Morgan Stanley, Citi, or MarketsandMarkets publish about their own inputs. It is also, by its own wording, a count of robots built, not a count of robots deployed to a paying customer’s floor and running unsupervised; Figure’s own June 30, 2026 update states that “Figure 02 contributed to the assembly of 30,000 cars at BMW last year” and that “Figure 03… arrived in Hall 52, one of the assembly and logistics halls at BMW Group Plant Spartanburg,” language that confirms a live, ongoing deployment without stating how many of the 350-plus units built are the ones actually working that line ^[6].

Agility Robotics discloses a different axis entirely. Its own November 20, 2025 announcement states that “Digit… has moved over 100,000 totes at GXO’s Flowery Branch facility” ^[7] — a cumulative task count, not a robot count, and the page itself does not name how many individual Digit units produced that total. A hundred thousand totes could be the output of a handful of robots working continuously for months or a larger fleet working shorter shifts; both are consistent with the one number Agility has actually published, and nothing in that disclosure lets a reader distinguish between them. A separately reported multi-year commercial agreement between GXO and Agility could not be independently loaded during this drafting pass on repeated attempts, so this piece treats its specific terms as unconfirmed and does not draw on them.

Tesla discloses the least of the three, and its own words are the most direct evidence for why. On the company’s Q4 2025 earnings call, reported January 28, 2026, Elon Musk described Optimus as “still very much at the early stages” and “still in the R&D phase,” adding that while “we have had Optimus do some basic tasks in the factory,” the robots are “not in usage in our factories in a material way” and that “it’s more so that the robot can learn” ^[8]. Musk went on to say Tesla would not “expect to have any kind of significant Optimus production volume until probably the end of this year” — meaning, as of that January call, sometime later in 2026 ^[8]. No fleet size was given in that transcript. Widely circulated estimates elsewhere in the press put Tesla’s internal Optimus population somewhere in the low thousands or fewer, but this piece could not independently confirm a specific figure from a Tesla-sourced disclosure during this drafting pass, and rather than repeat an unconfirmed number, treats Tesla’s own fleet size as undisclosed.

Three programs, three units of account: a manufactured-unit count that does not confirm deployment, a cumulative task count that does not confirm fleet size, and an explicit refusal to give either number at all, attached to a plain statement that the robots in question are not yet doing material work. None of the four forecasts above had to clear that bar. A dollar total for 2035 or 2050 does not need to specify how it was measured in the interim; a company disclosing what it actually built this quarter does, and the three companies making that disclosure have each chosen a different axis to be specific about.

The Ratio, Computed as Honestly as the Data Allows

Compute it anyway, because refusing to compute a number on the grounds that the inputs are imperfect is its own kind of dishonesty when the alternative is letting the forecast stand unchallenged by any comparison at all. The largest single verified manufactured-unit count among the three named 2026 programs is Figure’s “over 350” ^[5] — larger than any number Agility or Tesla has disclosed in comparable terms, though almost certainly not the largest true figure across the industry once platforms like Unitree’s commercially sold G1, which this piece has not independently re-verified a current shipment total for in this pass, are included. Divide Goldman’s 1.4-million-a-year 2035 shipment target by that 350-unit figure and the forecast-to-fielded ratio comes out to exactly 4,000: Goldman’s model implies an annual production rate, nine years from now, four thousand times larger than the single largest cumulative unit count any of 2026’s most closely covered humanoid programs has itself disclosed.

That number needs three caveats stated in the same breath it is offered, because each one cuts a different direction. First, it compares an annual run-rate nine years in the future against a cumulative count-to-date today — not a like-for-like comparison, and if Figure’s production line sustains anything close to its newly disclosed one-robot-per-hour cadence, its own annual output within a year or two would already be a large multiple of 350, shrinking the ratio fast on the fielded side alone, independent of anything Goldman does. Second, 350 almost certainly understates the true 2026 global humanoid population, since it excludes Tesla’s undisclosed internal fleet, Agility’s undisclosed Digit count, and any commercially shipped units from Unitree or other manufacturers this piece did not re-verify current figures for — meaning 4,000-to-1 is very likely an overstatement of the true current gap, not an understatement. Third, and cutting the other way, Figure’s 350 is a manufactured count, and the plain language of the company’s own disclosures does not confirm that all 350 units are doing productive work rather than sitting in testing, R&D, or qualification — exactly the distinction Tesla’s own executives draw explicitly for Optimus and that Figure’s own materials never quite resolve either way.

Put those three caveats together and the honest statement is not “the gap is four thousand to one.” It is that no comprehensive, cross-company, verified 2026 humanoid unit count exists anywhere this piece could locate, that the largest single verified figure available produces a computed ratio in the low thousands, and that the true ratio — if every company disclosed its fleet size in the same unit, at the same moment, with the same definition of “fielded” — could plausibly sit anywhere from several hundred to several thousand, a range this piece states rather than collapses into false precision. This is the epistemic layer worth naming directly: the 4,000-to-1 figure is derived arithmetic, computed transparently from two dated, cited numbers, but the number it derives from on the fielded side is a lower bound standing in for a true total nobody — not this piece, not Goldman, not any of the three companies — currently has in hand.

The Adoption-Curve Defense — and Where It Runs Out

There is a real objection to everything above, and it deserves to be stated in its strongest form rather than waved off. Technology-adoption curves are canonically exponential-looking only in hindsight. A forecast that looks absurdly premature relative to today’s fielded reality is fully consistent with that same forecast eventually proving accurate at its stated horizon — semiconductor unit forecasts from the 1970s, mobile-phone forecasts from the 1980s, and electric-vehicle forecasts from the 2000s all looked wildly ahead of their contemporaneous shipment data at some point along a path that, in each case, eventually got there or came close. A wide or widening forecast-to-fielded ratio in 2026 is not, on its own, evidence that Goldman’s 1.4-million figure or MarketsandMarkets’ $50.27-billion figure is wrong. It is evidence that the two numbers are not yet convergent, and early in a genuine adoption curve, that is exactly what the data should look like.

What would actually falsify this piece’s narrower claim — not that the forecasts are wrong, but that the gap between forecast and verified deployment is currently widening rather than narrowing — is a specific, checkable pattern in the next two or three report vintages: verified fielded counts growing faster, in percentage terms, than the forecasts themselves are being revised upward. Figure’s own disclosed 24x production-throughput improvement in under 120 days is, on its face, exactly the kind of fielded-side acceleration that could close a ratio like this quickly if it continues and if a meaningful share of the resulting units go into productive, disclosed deployment rather than testing and internal use ^[5]. If Figure, Agility, or Tesla publish a comparably large jump in verified, productive-deployment units over the next one or two quarters — while Goldman’s, Morgan Stanley’s, Citi’s, or MarketsandMarkets’ next report vintages hold roughly steady rather than revising upward again — the “keeps revising upward while the fielded fleet barely moves” verdict in this piece’s own title would need to be retired in favor of a convergence story. Nothing reviewed here shows that pattern yet. Every forecast vintage checked in this piece moved up; the most concretely disclosed fielded figures moved up too, but from a base small enough, and reported in units inconsistent enough, that no reader can yet tell whether the fielded side is closing the gap or merely producing bigger numbers on its own incompatible axis.

What the Next Forecast Should Have to Show Its Work On

None of this requires believing any of the four forecasting houses got their arithmetic wrong, and this piece does not make that claim. Discounted cash flows, adoption-curve models, and total-addressable-market estimates are built to describe a world that does not exist yet, and being far ahead of today’s verified deployment is not a flaw in that kind of model — it is close to the model’s entire job description. What none of the four reports currently do, and what this piece proposes as a standing check on the next report vintage from any of them, is publish their own forecast-to-fielded ratio alongside the headline number: state, explicitly, what current verified deployment their model’s implied trajectory is being measured against, in a unit that a reader could independently check against a company’s own disclosure.

That check would have cost Goldman nothing to include next to its 1.4-million-shipments figure, and it would have cost MarketsandMarkets nothing to include next to either its old $15.26-billion or its new $50.27-billion figure — a single sentence naming the verified 2026 or 2030 baseline the growth path is measured from, sourced to a company disclosure rather than another analyst’s estimate. Its absence is not a rounding error in an otherwise rigorous report. It is the one number that would let a reader tell, on the report’s own terms, whether “the market is about to arrive” and “the market has not yet shown up” are describing the same evidence or two different ones — and, on the evidence collected here, from Figure’s own production floor to Tesla’s own earnings call to a MarketsandMarkets report page that no longer shows what it said eighteen months ago, that check would currently come back showing a market whose forecasts move a great deal faster than its factories do.