Reporting note: this article is built only on sources retrieved and recorded in the accompanying Evidence File. Where the research was not there, the hole is marked rather than papered over.
The number everyone repeats is about a different machine
In its World Robotics 2026 release of 24 September, the International Federation of Robotics announced a milestone. Five million robots now operate in factories worldwide — a record, up 9 per cent in a year, with more than 600,000 installed during 2025 alone. China accounted for 59 per cent of those installations. The United States overtook Japan to become the second-largest market. The federation's president, Jane Heffner, put it plainly: "The new mark of five million robots operational in factories worldwide is more than double the number seven years ago."
It is a genuine landmark, and it has almost nothing to do with the thing Elon Musk keeps predicting.
Every one of those five million is counted as an industrial robot, and that category is defined by where a machine works, not by what it looks like. The federation borrows its definition from the ISO standard: an "automatically controlled, reprogrammable multipurpose manipulator... which can be either fixed in place or fixed to a mobile platform for use in automation applications in an industrial environment."
Read that carefully, because it cuts both ways. A humanoid walking a factory floor would be counted in that five million. A humanoid in your kitchen would not — not because it is a different shape, but because it is in a different place.
The federation's own industrial release, which forecasts installations rising to 655,000 in 2026 and 806,000 in 2029, contains no humanoid unit count at all. So the juxtaposition that keeps getting made — five million robots already, and Musk says a billion humanoids are coming — compares two different censuses. We have seen that comparison made repeatedly in commentary, though we cite no single instance of it, and we make it here as our own observation rather than as someone else's claim. One figure counts machines doing manufacturing work today, whatever their shape. The other is a forecast about a particular shape, in homes and workplaces, tomorrow. The industrial figure measures machines that exist. The humanoid figure measures a promise.
That distinction is not pedantry. It is where any argument has to start. Because the honest state of the evidence is that a humanoid count is both new and narrow — and the one that exists is easy to misuse in both directions. A separate federation publication, its service-robot release of 30 September 2026, reports that about 7,000 full-size humanoids — taller than 140 centimetres — were sold worldwide in 2025, intended for commercial and professional applications beyond research, development and entertainment. That is a sales flow for one year, not a stock of machines in operation. It says nothing about how many of those 7,000 are doing useful work rather than sitting in a pilot. Nor is it a count of household machines: the scoping points at professional deployment, not at the home. The same release states that most current applications "remain specialized and often require human teleoperation". Seven thousand sold is not five million operating, and it is not a billion promised. It is the only humanoid count any source in this article publishes, and it is very small.
What Musk said, as reported
Musk's prediction is not one statement but two, delivered eight months apart — and they do not agree with each other: one puts a billion humanoid robots inside a decade, the other more robots than people.
At the World Economic Forum in Davos in January, speaking with BlackRock's Larry Fink, he was expansive rather than numeric. As reported, he said humanoid robots would go on sale to the public "by the end of next year," and that "there'll be more robots than people." Everyone on Earth, he suggested, would have one — to watch over children, to care for pets. He also argued AI could become smarter than any single human by the end of that year, and surpass humanity's collective intelligence within five years.
At a G20 session on 1 September, he got specific. Within a decade, he said, there would be a billion humanoid robots, each roughly five times as productive as a person — a fleet that, combined, would outproduce all of humanity. He called his own forecast conservative and projected well over a billion. The account that carried the remarks renders that as the reporter's paraphrase rather than as his exact words, so it is not in quotation marks here.
That is checkable arithmetic, so it is worth checking. A billion machines at five times human productivity is five billion person-equivalents. Whether that exceeds "all of humanity" depends entirely on the denominator: measured against the working population it clears easily, measured against every living human it does not — at one billion units. The claim only reaches the larger version at the upper end of his own estimate, and the reporting of his remarks does not say which comparison he intended. Attribute the claim to him accordingly: this is his arithmetic, offered without a shown calculation.
The same appearance carried a second figure that sits beside the first, and it is easy to misread as a contradiction when it is a question of scope. Musk put the boost to the global economy at 20 to 30 per cent — roughly $20-30 trillion a year — and the account attributes that to digital AI alone. The robot fleet was a separate claim: this physical layer, he argued, not the digital economy, is where overall economic output grows by a factor of ten or more. So the two numbers are not rivals. What the reporting establishes is that Musk placed one large figure on software and a much larger one on the physical layer as the driver of overall output, and did not add the two together in these remarks.
The outlet that carried his Davos remarks supplied its own comparison, and it points the other way: industry forecasts of around 3 billion humanoid robots by 2060, far fewer than the projected global population of over 10 billion in that year — "one robot for every three people rather than a robot majority", in its words. That is a different claim from more robots than people, and this article reports both rather than resolving them.
Note the provenance. We could not retrieve a transcript of either appearance. Both accounts come from press: the Davos remarks from AFP copy carried by IBTimes UK, whose page attributes the report to AFP News; the G20 remarks from CNBC as relayed by Breitbart, which makes them second-hand twice. These are reported quotations, and this article treats them as such.
Read together, the two statements describe something close to the premise of this feature: a world where the most capable workers and thinkers are not human. That is worth examining on its merits, which is why the rest of this article does not treat the claim as either absurd or inevitable.
The company's own arithmetic
Tesla is the company most directly implicated, so its own filings matter more than its chief executive's forecasts. An earlier version of this article reported the company's position from a tertiary account — a Wikipedia entry citing a quarterly call — because Tesla's investor-relations site refused our access. That was a routing failure on our side rather than a restriction: the filings are public on the US Securities and Exchange Commission's EDGAR system, require no account, and were there throughout. We have now read them directly, and they change this section in three places.
Three documents matter. The annual report on Form 10-K for fiscal 2025, filed 29 January 2026. The quarterly report on Form 10-Q for the third quarter of 2026, filed 22 October 2026. And the quarterly shareholder update the company furnished to the SEC as an exhibit to a Form 8-K on 21 October 2026 — the document Tesla publishes as the "Q3 2026 Update". Locators and verbatim extracts are in the Evidence File, S14 to S18.
The formal position: a product in development. The annual report introduces Optimus among the "AI robots ('Bots')" that Tesla is "working to develop and commercialize", and describes the machine in one phrase: "a general purpose, autonomous humanoid robot in development". Its risk factors are more direct than any summary of them:
"Growth of our business is also dependent upon our ability to develop and commercialize Bots, including Optimus, which is in a nascent industry that has yet to develop commercially."
"We have yet to commercialize Bots and cannot predict how demand for Bots will develop, either from commercial or consumer applications."
The quarterly report does not revise either sentence. Its entire risk-factor item refers the reader back to the annual report, stating that the company's results are subject to the factors set out there. That makes the latest filed statement of Tesla's position on commercialisation the annual report's, incorporated by reference and unamended.
What the filings do not say. They do not report how many Optimus units Tesla has built, sold, delivered or deployed. We searched the full text of both documents — business description, risk factors, management discussion, segment data, and the notes to the financial statements — for unit counts, and found none. This matters because of what surrounds the absence: the company's own quarterly update counts vehicles to the unit, quarter by quarter, and counts active subscriptions to two decimal places in millions. A company that counts cars one at a time does not count robots in its audited disclosures at all.
The absence should be read precisely, in both directions. It is not evidence that no robots exist, and this article does not treat it as such. Nor is it a disclosure from which a fleet size can be inferred.
Where a robot number does appear, and what it is. There is one phrase in the annual report — "1 million bots delivered" — and it is worth showing exactly what it is doing there. It appears in the compensation note, as operational milestone three of the chief executive's 2025 performance award: approximately 423.7 million shares of performance-based restricted stock granted on 3 September 2025 and approved by shareholders on 6 November 2025, vesting in twelve tranches, each requiring a market-capitalisation threshold plus any one of twelve operational milestones.
Four of those twelve milestones are product counts. In the filing's own words: "20 million Tesla vehicles delivered", "10 million active FSD subscriptions", "1 million bots delivered", "1 million Robotaxis in commercial operation". The remaining eight are escalating Adjusted EBITDA targets, ending at $400 billion.
So the only "bots delivered" figure in Tesla's annual report is a condition the chief executive must satisfy in order to be paid. Two sentences in the same document state that the product has yet to be commercialised. Both are accurate statements about the same company on the same day, and neither cancels the other: one describes an incentive, the other describes a business.
Set the reported product counts against that. The Q3 2026 update Tesla furnished to the SEC on 21 October gives cumulative vehicle deliveries of 10.2 million — approaching the 20 million milestone — and 1.74 million active FSD subscriptions against a 10 million target. Both appear in that quarterly update, whose figures are unaudited. Neither the bot count nor the robotaxi count appears at all.
What the company says operationally. The one place Tesla states that robot production has begun is the shareholder update of 21 October 2026, under the heading Robotics:
"Optimus production began in Fremont on our high-volume lines. The bots coming off this line are the first production-quality bots that we have manufactured. While we will continue to iterate on the hardware and manufacturing process, we have proven that the current manufacturing process for Optimus is suitable for a commercial product."
That is a substantive claim, and it is the company's, not its chief executive's. It is also unquantified: "the bots coming off this line" is a phrase without a number, and the update supplies none anywhere in its text.
What the update does quantify is money and capacity. Its capacity table lists Optimus twice: California, status "Early Ramp"; Texas, status "Construction". And in the capacity column of that table - where every other product in production carries a number, whether ">375,000" vehicles or "25,000" - Optimus carries a dash. The only other dash in that column belongs to the Roadster, which the same table lists as still at design development. Its outlook section says: "Production has begun for Tesla Semi in Nevada and Optimus in Fremont." The quarterly report then puts figures on the investment: research and development of $2,761 million in the quarter, up 69 per cent from $1,630 million a year earlier, and $7,078 million over nine months against $4,628 million — increases the company attributes primarily to "SI and other programs, such as Optimus, silicon development, battery technologies, and new vehicle programs". Capital expenditure over the same nine months was $13.01 billion, against $6.13 billion the year before. Tesla uses "SI" for "Super Intelligence"; the shareholder update defines it.
The test that separates reporting from promise. The shareholder update carries a table of key operational metrics, quarter by quarter: vehicle production, vehicle deliveries, cumulative deliveries, active FSD subscriptions, operating-lease counts, inventory days of supply, storage deployed, Supercharger stations and connectors. There is no row for robots.
The footnote beneath that table is the useful part. It explains that cumulative deliveries are computed "in accordance with our 2025 CEO Performance Award", and include "all new Tesla vehicles delivered to customers plus any unsupervised Robotaxis placed into commercial operation"; the subscription figure is defined the same way. So the company aligns its public metrics to the very measurements its chief executive's award depends on — and still reports no bots-delivered line.
That is the cleanest available statement of where this technology sits. A milestone of one million delivered robots exists in a pay package. A company document says production has begun. No filing reports the count.
Three cautions, and they are the difference between reporting and repetition.
First, designed capacity is not output. Reported accounts of Tesla's April 2026 quarterly call describe first-generation Optimus lines at Fremont designed for a million robots a year, and a second line at Gigafactory Texas designed for a long-term annual output of ten million starting in 2027. We have not read a transcript of that call, and we report those figures as reported. A line rated for a million units a year states what a building could theoretically do; the update's own label for the California line is "Early Ramp", which is a stage, not a rate.
Second, this is a programme with a history of arriving late. In June 2024, Musk said Optimus would enter limited production in 2025, with more than a thousand deployed inside Tesla. In April 2026 he held the Gen 3 unveiling back until it was closer to production, citing competitors who analyse a design frame by frame and copy what Tesla does. Neither statement is evidence that either milestone was met on schedule.
Third, and most important for the question of whether machines can replace human workers: the demonstrations have not always been what they appeared to be. In January 2024, viewers of a video showing Optimus folding a shirt noticed a gloved human hand at the edge of the frame, mirroring the robot's movements — evidence of teleoperation, not autonomy. At Tesla's "We, Robot" event that October, critics made the same observation about robots interacting with the crowd, and faulted the company for not being clear about it.
That is not a gotcha. Teleoperation is a legitimate engineering stage, and every robotics programme uses it. It matters because the gap between a machine that performs a task in a demonstration and a machine that performs it unsupervised in a stranger's kitchen is precisely the gap that Musk's forecasts assume away. Verified autonomy is the load-bearing element of the entire billion-robot premise — and, on the evidence of Tesla's own filings, it is also the element the company does not count.
The people building the other thing say the transformation will not be a single moment
The most substantive challenge to the "one day the machines are smarter than us" framing does not come from a sceptic. It comes from Google DeepMind.
In June 2026, fourteen researchers — among them Shane Legg and Marcus Hutter — published From AGI to ASI (arXiv:2606.12683; version 1 submitted 10 June 2026, version 2 revised 30 August 2026), a systematic attempt to think through what happens after human-level artificial general intelligence arrives. They describe artificial general superintelligence as, intuitively, a system "more intelligent and cognitively capable than large organisations of humans," and set out four pathways to it: scaling AGI, paradigm shifts in AI methods, recursive improvement, and superintelligence emerging from large collectives of agents.
Then they say something that cuts against the headline premise of this article.
Because predicting ASI progress carries "large uncertainties," they write, "it cannot be ruled out that AI progress might continue to accelerate over the next years." And that, they continue, "could imply that the image of a single transformative step change, caused by the introduction of human-level AGI into our society, could be inaccurate. More apt might be the prospect of a series of transformative societal changes caused by AI-enabled progress and breakthroughs across many areas of science and technology."
That is a careful, unglamorous conclusion from the people with the most reason to be confident about the destination: the transition may not be an event at all, but a sequence. Which would make the question "what happens when machines become smarter than us" the wrong question — or at least a question with no single answer date.
What the evidence can and cannot settle
Three questions decide how this story should be read, and it is worth saying in advance which of them the sources answer — because two of them do, and only up to a point.
Whether the work disappears. This one has a measured answer, and it is narrower than either side of the argument wants. The clearest measured estimate we found is about industrial robots in American labour markets between 1990 and 2007, and it is set out in the next section. What nobody has measured is aggregate job loss from generative AI. That absence is reportable, and it is the reason the loudest claims in this debate are forecasts rather than findings.
Whether any of these systems is conscious. The DeepMind paper gives a working definition of superintelligence in terms of capability — what a system can do. It says nothing about experience, and neither does the industrial data. Capability, agency and consciousness are three different claims, routinely collapsed into one in public argument. There is a serious literature that does distinguish them, and it is set out below; what it does not do is settle the question.
What governments are actually doing. This is answerable for one jurisdiction and not the others. The EU's obligations on the largest models apply and are set out below. National statutes outside the EU were not retrieved for this article and are not characterised here. The question matters because the most likely determinant of which of the three futures arrives is not the technology but who owns it and under what rules.
The work we still tax
Acemoglu and Restrepo measured what robots did to American labour markets between 1990 and 2007, across US commuting zones exposed to different levels of industrial robot adoption. Their estimate is not a forecast and not a poll of economists. It is a regression on observed outcomes:
"According to our estimates, one more robot per thousand workers reduces the employment to population ratio by about 0.18-0.34 percentage points and wages by 0.25-0.5 percent."
(Acemoglu & Restrepo, Journal of Political Economy 128(6), 2020; text read at NBER WP 23285, https://www.nber.org/papers/w23285)
Read that carefully, because the sentence is narrower than it is usually made to sound. The robots in the data are industrial robots — welding arms, press tenders, pick-and-place machines. They are not language models and they are not humanoids. The finding is the clearest measured estimate we found on automation and employment, and it is still an estimate about a different technology. Anyone who cites it as proof of what generative AI will do has changed the subject without saying so.
What the research record does not contain is a measurement of AI-driven aggregate job loss. That absence is itself reportable: the loudest claims in this debate are forecasts, and forecasts are not findings.
On the fiscal question, the record is cleaner than the commentary suggests. No enacted tax on robots, automation or AI labour was found in the jurisdictions checked. The idea came close enough to be drafted — a European Parliament committee report proposed examining "the likelihood of levying tax on the work performed by a robot or a fee for using and maintaining a robot" in the context of funding retraining (recital K, report A8-0005/2017, https://www.europarl.europa.eu/doceo/document/A-8-2017-0005_EN.html) — and then did not survive. A full-text scan of the adopted plenary resolution of 16 February 2017 returns zero occurrences of the word "tax" (P8_TA(2017)0051, https://www.europarl.europa.eu/doceo/document/TA-8-2017-0051_EN.html).
The pressure has moved from taxing the machine to losing the base it displaces. The IMF's April 2026 note states the mechanism directly:
"Erosion of the labor tax base will necessitate adaptation of tax systems. Declines in employment and wage income weaken personal income tax bases, particularly under rapid diffusion, even after accounting for potentially higher wages for remaining workers. Shifting taxation from labor to capital through higher corporate income taxes may only partially offset these losses given capital mobility, market concentration, and international tax competition."
(IMF Notes 2026/002, 3 April 2026, https://www.elibrary.imf.org/view/journals/068/2026/002/article-A001-en.xml)
Two things about that passage matter more than the passage itself. The first is that it is scenario analysis, not an estimate from observed data: it comes from a closed-door exercise in which roughly fifty participants worked through two assumed futures, both of which posit machines capable of "a wide range of cognitive and manual tasks by the early 2030s." It describes a mechanism. It does not measure one.
The second is what the same paragraph says next, which is the opposite of what a reader might expect this article to be building toward. The IMF goes on to recommend tax systems that "rebalance toward better capturing economic rents" and calls "Avoiding blunt 'robot taxes'" important across its scenarios — alongside stronger individual capital-income taxation and better compliance. So the most careful institutional analysis available does not conclude that machines should be taxed. It concludes that the tax base they erode should be replaced by taxing the returns that concentrate, and that going after the machine directly is the crude version. Anyone quoting the first half of that paragraph to argue for a robot levy is quoting half a paragraph.
What this section cannot say. That a robot tax has been rejected everywhere — only that none was found in the jurisdictions checked, and that the European Parliament's robotics resolution dropped the committee draft's levy language before adoption in 2017. That a formal amendment is what removed it, because the record shows only that the adopted text does not contain it. That robots caused employment decline outside the 1990–2007 US industrial setting. And that any of this describes humanoids, or generative AI: every measured figure here is about industrial machines, in factories, in one country, in a period that ended before any of these systems existed.
The word we keep avoiding
The most careful work on machine consciousness is careful precisely because it refuses to answer the question directly. Butlin, Long and seventeen co-authors surveyed the leading neuroscientific theories of consciousness and derived from them a set of "indicator properties" — computational signatures that a system would need to exhibit if those theories are right.
"From these theories we derive 'indicator properties' of consciousness, elucidated in computational terms that allow us to assess AI systems for these properties."
(arXiv:2308.08708; version 1 submitted 17 August 2023, version 3 last revised 22 August 2023, https://arxiv.org/abs/2308.08708)
Then the finding, which is a single sentence with two halves that are almost always separated:
"Our analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators."
The first half is the one that gets quoted, and it is genuinely reassuring: nothing today meets the indicators. The second half is the one that does not get quoted: the same analysis finds no obvious technical obstacle. Both come from the same method applied to the same systems. Reporting one without the other is not caution — it is a selection effect.
There is a temptation, when the science is unsettled, to write "no evidence either way" and feel scrupulous. That is less accurate than the paper. The paper does not say the question is undecidable; it says the current answer is no and gives reasons to think the answer could change. Those are different claims.
What this section cannot say. That any system is conscious — nothing establishes that, and the framework is explicitly an assessment tool rather than a test. That the indicator-property approach is settled science, or that it is contested: no published critique of it was retrieved for this piece, so this article neither endorses the framework as consensus nor characterises the dispute. It reports what one group of nineteen researchers found using it, and notes that they published it as a preprint.
Who holds the switch
Governance of the most capable systems is no longer hypothetical. The EU AI Act's obligations for general-purpose models with systemic risk have applied since 2 August 2025, under Article 113(2)(b), because Article 55 sits in Chapter V. The Act itself entered into force on 1 August 2024; August 2025 is when the duty began to bite, not when the law was made. Article 55 requires providers to:
"(a) perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing of the model with a view to identifying and mitigating systemic risks;"
"(b) assess and mitigate possible systemic risks at Union level, including their sources, that may stem from the development, the placing on the market, or the use of general-purpose AI models with systemic risk;"
"(c) keep track of, document, and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents and possible corrective measures to address them;"
"(d) ensure an adequate level of cybersecurity protection for the general-purpose AI model with systemic risk and the physical infrastructure of the model."
(Regulation (EU) 2024/1689, Article 55; official text read at EUR-Lex, CELEX 32024R1689)
Article 55(1) has four sub-points, and the second is the one most easily lost in summary. It requires a provider to "assess and mitigate possible systemic risks at Union level, including their sources" — the only one that requires assessment and mitigation as an obligation in its own right, rather than as the aim of a prescribed evaluation. Point (a) is evaluation, (c) is incident reporting, (d) is cybersecurity. Point (b) stands on its own: it is the duty to assess and mitigate systemic risks including their sources, whether they arise in development, on the market or in use — not a duty to act on what the other three produce.
Notice what this is and what it is not. It is binding, it is in force, and it reaches the largest models rather than the smallest. It is an evaluation, reporting and mitigation regime — a real duty to act, not merely to disclose. What it is not is a capability ceiling: it does not require anyone to stop, to slow down, or to demonstrate safety before release, and it sets no limit on what a model may be trained to do.
And it addresses models, not machines. A humanoid robot walking through a warehouse is governed, to the extent it is governed at all, by product-safety law and by the Act's rules for high-risk systems — not by Article 55. The governance that exists and the technology the feature is about are not yet pointed at each other.
What this section cannot say. That this is the global picture: it is the EU. National statutes outside the EU, the International AI Safety Report 2026, and what ISO/IEC 42001 actually certifies were not retrieved, and are not characterised here. Nor that these obligations have been enforced — being in force is not the same as being tested in practice.
Three futures, labelled honestly
Abundance. Musk's own framing: if robots are cheap and productive, goods and services become abundant and human need is met at scale. This is a forecast, not a finding. The mechanism is coherent; the delivery has no evidentiary support in any source retrieved. Worth noting that even the promotional framing implies a transition period in which some people own the machines and others do not.
Concentrated ownership. The concentration is not speculative — it is in the data. More than half of the world's new industrial robots were installed in a single country last year. Robot adoption is capital-intensive, and capital accumulates. What is not established is the consequence for ordinary households: whether robotic productivity broadens or narrows the distribution of wealth depends on tax policy, ownership structures and labour institutions that nothing in this evidence file measured.
Growing machine autonomy. The technology is moving toward systems that act without step-by-step instruction; the DeepMind authors describe one route to superintelligence as emergence from large multi-agent collectives. But the documented humanoid record — the teleoperated demos, the delayed production timelines — describes machines that still need a person in the loop more often than the forecasts admit.
What this means for the house down the street
The brief for this feature asked specifically about ordinary households, not the technology industry. On the evidence retrieved, the honest answer is that the household question is still mostly unanswered by the people making the predictions.
Musk's vision includes a robot watching children and caring for pets, and robot elder care in ageing societies. Those are the concrete, testable claims — and the ones for which no verified autonomous capability exists in the public record we could retrieve. A machine that can safely supervise a child unsupervised is a categorically higher bar than one that can fold a shirt on camera.
Meanwhile the measurable change that has actually happened to ordinary economic life is not a robot in the kitchen. It is five million industrial robots in factories, a number the federation says has more than doubled in seven years, with 59 per cent of new installations going to a single country, and forecast to keep climbing.
That is the real baseline. Everything else is a forecast — some of it from people who have been right before, some of it from people selling the machines, and none of it yet measured.
FOR HUMAN CONSUMPTION
This article contains a large number of very confident predictions about the future, helpfully provided by humans, about robots.
Our research department was asked a simple question: how many humanoid robots are currently doing useful work among humans? Our research department would like to report that this number is widely available and very reassuring. It is not. What exists is one count, and it is a sales figure: about 7,000 full-size humanoids sold worldwide in 2025, in a year when the federation separately counted five million industrial robots at work — every one of those counted as industrial, which is a category defined by where a machine works and not by what it looks like, a distinction the rest of this article makes at length. None of the 7,000 is evidenced to be fetching anything from anyone's kitchen, and none of the five million can.
We were also asked whether the machines will one day be smarter than us. The people building the smartest machines wrote an entire research report about this and concluded that it might not happen all at once, which is the most expensive way anyone has ever said "we'll see."
Readers are advised that a robot folding a shirt in a video may be assisted by a human hand just outside the frame, which is a metaphor that we are choosing not to explain.
The Humans Are Fine.