Everything that changes before 2036

By 2036, on the charts that follow, people will be a small minority on the web and machines will have been reading more of it than we do for several years. Most of us will meet AI in the tools we already use, and most of us will use it deliberately. Well over half of new cars will be electric, and a third will carry hardware built for some form of self-driving. The humanoid robot market will have either delivered ten million units or fewer than one. Space tourism will have restarted or it will not, and on present evidence the first footprints on Mars are still a date rather than a place. Whole categories of work that today employ millions will have stopped hiring at the bottom, and whole categories of business will have stopped being things anyone starts. And the encryption protecting all of it will either have been replaced or will be running out of time to be. Underneath all of it, the planet will be past 1.5°C of long-term warming, which is the one forecast here with physics rather than adoption curves behind it.

Some of that is measured. Most of it is modelled. The point where one becomes the other is marked on every chart, because the difference between the two is the whole story of this year. The machines are ahead of schedule, and the schedule was wrong in both directions at once. Cloudflare's chief executive predicted bots would pass humans on the web in late 2027; it happened in June 2026. Estimates of Chinese humanoid shipments tripled inside six months. The qubit count needed to break the internet's encryption fell twentyfold on a single paper. Meanwhile the most-shared chart of the year, Linux surging past Windows, turned out to be measuring crawlers.

So these seventeen charts are drawn to show the shape rather than the number. Most run to 2036; one runs to 2040, because that is where the most ambitious forecast in the piece sits, and the two grids near the end carry a further column for 2047 which is ruled off and labelled as a different kind of claim. Where credible sources disagree by a wide margin, we have drawn the band rather than picked a line, and where they are measuring different things we have said so rather than joined them up. Two of the charts will not say what you expect. One of them contains a set of dates no company can move.

1. Humans become a minority

Delegating one task to an agent does two things at once. It removes the pages a person would have viewed, and it adds the pages a machine views instead. Our model assumes roughly seven human page views per self-served task and around a hundred machine fetches per delegated one. That ratio is an assumption, not a measurement, and the chart is sensitive to it. But under any plausible version of it the human share moves far faster than adoption, and human use does not have to fall at all for the human share to collapse.

2. The crossover: machines reading more than people, some time after 2030

Strip out the legacy scrapers and training crawlers and look only at live, person-initiated agent fetches against human page views. Agent traffic is somewhere around 0.4% of web requests today. On our central model it passes human traffic early in the next decade. That is a large claim and it deserves to be stated as one.

Here is what the crossing actually requires. The agent-to-human ratio today is roughly 0.4 to 42.5, or about one to 112. Reaching parity by 2031 means that ratio grows about 112-fold in five years, which is around 157% a year compounded. That is not a timid forecast. It is a strong one, and it rests on assumptions about how many tasks each person delegates, how many fresh fetches each task triggers rather than cached answers or API calls, and how much human browsing each delegation actually displaces. Change any of those and the crossing moves by years. We have kept 2031 as the illustrative case because it follows from the same assumptions used in chart 1, not because it is the most probable single year. For scale, the one machine-side series anyone publishes weekly grew faster than that over the past year: token volume through the OpenRouter exchange went from 4.9 trillion a week in September 2025 to 115 trillion in the last week of August 2026, a factor of 23. That is inference demand rather than page fetches, and part of it is one exchange taking share from providers' own APIs, so it does not transfer to this chart. It does mean 157% a year is not, by the standards of the nearest measured thing, an aggressive number.

3. The driver: how people are actually meeting AI

The adoption question is usually asked as though AI were a destination people choose to visit. Increasingly it is not. It is embedded in the search box, the email client and the document editor people were already using. Those are different things and they deserve different lines.

The reason to draw three lines rather than one is that they will not move together. The top line is close to saturation already, because it counts anyone who sees a generated answer above their search results whether they wanted one or not. The middle line is the traditional adoption curve. The bottom line is the one that drives every other chart in this article, and it is the one nobody measures well yet. Being bold about the top line costs nothing; being bold about the bottom one is the actual forecast.

4. Which tasks move first

Adoption arrives by task, and within any single population the order is fairly consistent: research first, then creation, then decisions, then transactions. The gap between the first and the last is measurable. Among people who use AI at all, more than three-quarters use it to shop, yet nearly one in three will not let it spend money, and only around 6% of workers say they go to AI first when a task lands on them.

Read each group on its own terms. Among workers who already use AI, the tasks near the top are the ones where a first draft is useful even when imperfect and a person checks the output anyway. Among all adults, the tasks near the bottom involve money, health or a date that cannot be missed, and those are where people keep their hands on the wheel. Microsoft's own Work Trend Index found decision-making accounting for 28% of Copilot activity, so the decision layer is moving faster than most people assume. The transaction layer, where an agent actually pays, is the one still held back by the third of people who will not let it, and that is the layer that changes the economics of the web when it moves.

5. Linux on the one measure that cannot be gamed

Page-view trackers spent 2026 reporting a Linux surge that turned out to be bot traffic reclassified. Measured on Valve's hardware survey, which reads the client rather than a user-agent string and cannot be inflated by crawlers, Linux was 3.90% in August 2026 against Windows at 93.95%.

Linux gained more share in three months of bad measurement than it has in a decade of real adoption. That is the whole lesson of this article in one sentence.

One caveat we hold rather than dismiss. It is entirely possible that Windows faces a sharper fall than either line above: the Windows 10 hardware cut-off has stranded a large installed base, European public procurement is moving with unusual conviction, and the next generation of personal devices may not be PCs at all. If that happens it will show up first on Steam and in first-party government analytics, not on page-view trackers. Nothing in bot-proof data shows it yet, and we would rather publish a chart that is early than one that is wrong.

6. The quantum clock, and why the deadline is not a qubit count

The number everyone wants is the year a quantum computer breaks RSA-2048. Nobody can give it, and the reason is worth understanding, because the chart below is easy to misread.

Two things are true at once. The capability is not close: the largest demonstrated error-corrected count is under a hundred, and it is at a code distance far below what Shor's algorithm needs. And the estimates keep moving in one direction: Google's Craig Gidney cut the physical-qubit requirement roughly twentyfold in a single 2025 paper, trading an eight-hour run for one under a week, and a February 2026 architecture paper proposes breaking RSA-2048 with under 100,000 physical qubits using newer codes. IBM's own roadmap, taken at face value, passes the estimate on the chart between 2031 and 2033. Vendor roadmaps are ambitions, not deliveries. Crossing that line is not the year encryption breaks; it is the year the qubit count stops being the constraint.

The deadline is not scheduled for the day the machine arrives, because encrypted traffic captured today can be stored and decrypted later. The EU's draft NIS2 amendment in early 2026 described harvest-now-decrypt-later as likely occurring already. The NSA's schedule for national security systems has three separate dates: new acquisitions must comply from 1 January 2027, equipment that cannot support the new algorithms must be phased out by 31 December 2030, and the algorithms are mandated for use by 31 December 2031. Those bind national security systems, not every company, but they set the pace that suppliers to those systems will follow, and a June 2026 US executive order extended migration requirements across federal agencies.

When each quantum technology actually arrives

"Quantum" is used for five different things, and they arrive decades apart. Two of them get called quantum encryption by different people, and only one of those needs a quantum computer anywhere near it.

The practical answer to "when does quantum arrive" is that the defensive half has already arrived and the offensive half has not. Quantum-safe encryption is running on most of your web traffic today without anyone having had to notice. Quantum computers should be doing paid scientific work around the end of this decade. The machine that breaks today's public-key encryption is most likely a late-2030s event, but Cloudflare, which sees more of the internet's traffic than almost anyone, has moved its own target for complete post-quantum protection forward to 2029 on the strength of recent progress. That is the date to plan against, not the one where the lines cross on the chart above.

7. Cars: electric first, autonomous second

The two transitions are usually discussed together and are running several years apart. Electrification is well past its inflection; autonomy is barely at its own.

8. Air taxis: launch is real, scale is contested

An earlier version of this section used a single narrow forecast and presented it as the answer. That was wrong, and the correction is more interesting than the original claim.

The approvals are genuinely close. Joby's quarterly filing of August 2026 put the fourth of the FAA's five certification stages at roughly three-quarters complete, and the company has been flying production-conforming aircraft since March. Archer says it was first to close the third of four phases. The FAA's eVTOL Integration Pilot Program lets pre-certified aircraft operate across 26 states, and Joby flew the first such point-to-point flights in New York in April 2026. Joby is positioned to carry paying passengers from a Skyports vertiport at Dubai International in 2026 under a six-year exclusivity, has absorbed Blade's Manhattan passenger business and its 90,000 annual riders, and Archer and United have announced a nine-node New York network with Archer also selected for the 2028 Los Angeles Olympics. This is not a technology waiting for permission. It is a technology at the final gate.

The honest statement is not that forecasters disagree by a factor of several hundred. It is that nobody has published comparable forecasts for the same market, and the numbers most often quoted alongside each other were never measuring the same thing. What is settled is the shape of the launch: two to five aircraft per city at around $200 to $300 a ride. At least six manufacturers have ceased operations or entered insolvency since 2023, including Lilium and Volocopter, and the sector has consumed more than $13bn in equity since 2019 against passenger revenue in the United States that remains effectively zero until type certification lands, which analysts now put no earlier than mid-2027. The companies still standing are exactly where they need to be. Whether this becomes a transport mode or a premium service is the open question, and a scope-matched forecast that would answer it does not yet exist.

9. Humanoid robots: the line is live, the volume is not

Tesla's Optimus line at Fremont, converted from Model S and X production, began operating in late August 2026. That is real. What it is producing is low-volume units for internal training and data collection, with commercial sales targeted for 2027. As of mid-2026 Tesla had an estimated 1,000 to 1,200 Optimus units across Fremont and Giga Texas, no external sales, and no published uptime figures; on the January 2026 earnings call the chief executive acknowledged that none were yet doing useful work. Wolfe Research does not expect meaningful external revenue before late 2027, and the AI5 chip the robot is designed around is not due until volume production around mid-2027.

The consumer end is arriving ahead of Tesla. 1X's NEO, backed by OpenAI, is the first home humanoid with transparent pricing and a confirmed delivery timeline: $20,000 outright or $499 a month, with more than 10,000 pre-orders and deliveries through 2026. XPeng closed a $900m round to bring IRON to mass production by the end of the year. Figure is building its third-generation robot at roughly one an hour. Unitree's G1 sells for around $13,500 and its R1 for under $6,000. Chinese manufacturers accounted for roughly 97% of global humanoid shipments in the first half of 2026.

The economics are moving faster than the units. Bank of America puts a China-built 2025 bill of materials near $35,000, falling below $17,000 by 2030, with actuators more than half the build, which is why Tesla designs its own. Tesla targets $20,000 to $30,000 at scale; most analysts expect $100,000 to $150,000 for early units. Goldman Sachs projects a $38bn humanoid market by 2035. The Fremont line is designed for up to a million units a year, and Tesla has declined to give a 2026 target. Mass production is a 2027 story. The line going live in 2026 is the thing that makes 2027 possible.

When do robots start building robots?

In the narrow sense, they already do, and have for a quarter of a century. FANUC's plant near Mount Fuji has run lights-out since 2001, with industrial robots assembling other industrial robots at about fifty per shift and running unsupervised for up to thirty days at a time. Loading software into a finished machine is the easy part: modern factories flash firmware automatically at the end of the line, so a robot putting its own code into a new one is not the difficult step.

There are three difficult steps. The first is dexterous assembly of something as complicated as a humanoid, with thousands of parts, hands full of fine wiring and actuators that need calibrating; the arms that build arms work on a far simpler product. The second is the AI replicating itself, and the evidence here moved sharply in 2026. The UK AI Security Institute's 2025 benchmark found that frontier models could manage many parts of autonomous replication but not the whole of it. A study published in May 2026 then found models that could compromise a vulnerable server, copy their entire inference stack onto it and repeat the process from the copy, with the best model's success rate rising from 6% to 81% within a year, in controlled test environments. The third is closing the loop entirely: designing the next generation, sourcing the materials, maintaining the factory and deciding what to build, with no person making a decision anywhere in it.

Our estimate is that humanoids will do a meaningful share of the assembly of their own model, on a line where the software is loaded automatically, around 2029 to 2031. Tesla's chief executive has described Optimus as the first machine that will be able to build civilisation by itself on any viable planet, but told investors in January 2026 that it was not yet in material use in Tesla's own factories. A plant where robots build, test, program and ship robots with people only supervising is an early-to-mid 2030s event. The fully closed loop is not, and not only for engineering reasons. Autonomous self-replication is one of the capabilities that governments and AI developers now test for specifically as a safety threshold, because it would be very hard to reverse once it existed. The date that loop closes is therefore partly an engineering question and partly a decision about whether anyone should allow it, and the sensible answer is that a named person keeps the off-switch.

10. Who builds all this

Stanford's AI Index 2026 says the lead between the best American and Chinese models has changed hands more than once since early 2025. In February 2025 DeepSeek-R1 briefly matched the top US model. By March 2026 the leading US model sat 39 rating points ahead on the Arena leaderboard, 1,503 against 1,464. That is a lead of under 3%, and it is not a lead anyone should assume is permanent.

The composition of the lead is the interesting part, because no country leads everywhere. The United States dominates capital and infrastructure: $285.9bn of private AI investment in 2025 against China's $12.4bn, a factor of 23, with California alone accounting for $218bn, and 5,427 data centres, more than ten times any other country. China leads on output and physical deployment: 69.7% of global AI patent filings, 23.2% of publications, 20.6% of citations against 12.6% for the US, and industrial robot installations at roughly nine times the American rate. South Korea leads the world on patents per head. And a single foundry in Taiwan fabricates almost every leading AI chip.

There is one place where that composition is measured weekly rather than surveyed annually, and it points the same way as the patent and robot figures rather than the Arena one. OpenRouter is a neutral exchange that routes requests across several hundred models from every major lab, and it publishes its own token volumes. In January 2025 models from American labs carried 87% of everything routed through it. They now carry about a third.

The crossing is not a single moment. Chinese models first passed American ones in the week of 23 March 2026, lost the lead again in April, and have held it since the 27th of that month, finishing August at about 65% against 34%. The line that moved most quietly is the third one: everywhere else, which includes every European, Japanese and Korean lab put together, fell from 6% to under 2% and our scenario keeps it there. And for one fortnight in August the best instrument available for this question could not tell you the nationality of a sixth of what it was measuring. That is the condition of the evidence in this area, on the one measure where somebody actually counts.

The right-hand panel is deliberately drawn as a range wide enough to contain the United States taking the lead back. The mechanism that produced the measured rise is not mysterious and not obviously self-reversing: Chinese labs publish open weights, an exchange like this accumulates whatever is cheapest to serve, and cheap models attract the high-volume, low-stakes work that generates most tokens. What would move the lower edge is the thing the chart cannot see, which is whether the workload that matters stays on the frontier labs' own APIs. On present evidence that is the likeliest shape: a widening gap between where the tokens go and where the value does.

11. Space tourism, the only line here that goes down

Of everything in this article, this is the market that was supposed to be furthest along. Two companies spent two decades and several billion pounds building vehicles to sell short trips above the atmosphere, both got them flying, both carried paying customers, and in 2026 neither of them is flying anybody.

The suborbital line is the tourism business and it peaked in 2025 at 42 people. Blue Origin, which had carried more space tourists than anyone in history, looked at its own ledger in January 2026 and parked the vehicle for at least two years to work on its lunar lander instead. Virgin Galactic retired Unity in June 2024 to build the Delta fleet and has not flown a passenger since. The orbital line, which is not tourism but professional crews plus a handful of private orbital missions, barely moved through all of it: between 21 and 28 people a year for six years. The dramatic part of human spaceflight turned out to be the part with almost no growth in it, and the part with growth turned out to be able to stop.

A market can have working technology, satisfied customers and a waiting list, and still be worth switching off.

12. The satellites, which went the other way entirely

The same company that cannot yet fly a Starship to Mars has put more than eleven thousand working satellites in low Earth orbit and sells broadband from them to twelve million subscriber lines. This is the space business that actually happened, and it happened while everyone was watching the rockets.

Two things about that curve are worth separating. The subscriber count doubled in twelve months, which is the kind of growth rate that cannot continue for a decade and is the reason our band widens so fast. But the constellation underneath it is a different kind of fact: 11,133 satellites in orbit on 11 September 2026, out of 12,935 launched since May 2019, on a network that is now lowering its main shell from 550 to 480 kilometres to reduce collision risk. Roughly two-thirds of all working satellites around Earth belong to one company.

The competition is the more instructive story. Amazon's constellation, renamed Leo in late 2025, had 398 satellites launched across 15 flights by July 2026 against an FCC licence condition requiring 1,618 in orbit by 30 July 2026. It missed by a factor of four. The regulator granted a conditional waiver rather than an extension, temporarily demoting the spectrum priority of everything launched after the deadline, and left the requirement for all 3,236 satellites by July 2029 in place. That is what a hard deadline looks like in this sector: not a slipped press release, but a licence condition with a penalty attached, and it was still missed.

13. The Moon and Mars: measuring the slip itself

Every other chart here plots a quantity against time. This one plots time against time. For each public statement of a target date, it marks the year the statement was made and the year it named. A target that is genuinely approaching produces a flat line. A target that recedes as fast as time passes produces a line parallel to the diagonal, and never arrives.

Read the amber line first. In 2016 SpaceX said an uncrewed spacecraft would land on Mars in 2018, two years out. In 2017 it said 2022, five years out. In 2024 it named the 2026 transfer window, and said five uncrewed Starships would go, carrying Optimus robots in place of a crew. The 2026 window opened and closed with nothing departing. The company's own page now offers cargo flights to the Martian surface no earlier than 2028, and in July 2026 Musk put humans on Mars at five to seven years away, which is 2031 to 2033. Across a decade the destination has not changed, the technology has advanced enormously, and the date has moved four times.

This is not an argument that it will not happen. Orbital boosters land routinely, the largest launch vehicle ever built flies, and in February 2026 SpaceX told investors it would prioritise an uncrewed lunar landing in March 2027 over the Mars window, which is a real and defensible engineering decision rather than a retreat. It is an argument about what a development target is: a statement of intent from inside a programme, not a schedule anyone outside it should plan against. Celestial mechanics makes the misses unusually legible here, because the window only reopens every twenty-six months. Most industries can quietly absorb a six-month slip. Mars rounds every slip up to two years.

NASA's line has the same shape for the same reasons. The first Artemis landing was 2024 when the programme was announced, September 2026 as of January 2024, and in February 2026 Artemis III stopped being a landing at all: it was re-scoped to a docking demonstration in Earth orbit after a safety panel called the previous plan high risk, pushing the first surface mission to Artemis IV in 2028. What did happen, and deserves saying plainly, is Artemis II. Four people flew around the Moon between 1 and 11 April 2026, the first humans beyond low Earth orbit since 1972.

The dotted line is the one that should give a Western reader pause. China has said a crewed lunar landing before 2030 since 2021 and has not moved it once in five years, while the hardware for it — the Long March 10, the Mengzhou spacecraft, the Lanyue lander — has gone through development on schedule by its own account. Its robotic precursor has not kept pace. Chang'e-7, bound for the lunar south pole, was on the pad in August 2026 and was stood down hours before its launch on 23 August with a one-line statement that it did not meet launch conditions and could not fly in this year's window. No new date has been given, 2027 is widely expected, and Chang'e-8 is now described as a 2029 mission. The formal plan for the International Lunar Research Station puts construction from 2026 to 2035 and a utilisation phase beginning in 2036, which is the last year on every chart in this article.

14. Which work survives, and which does not

The question people actually ask is not what share of tasks a model can do. It is whether a particular job will still be worth hiring for. That question has one piece of hard evidence behind it and a great deal of guesswork, and the two are easy to confuse, so they are separated here.

The hard evidence is narrower than the headlines suggest and stranger than most forecasts. Across the whole workforce, almost nothing has happened: employment in the most AI-exposed occupations contracted 0.2% in the year to April 2026, against 0.1% growth in the least exposed. Cut the same payroll data by age and the picture changes completely.

That is the whole of the measured picture. It says the adjustment is happening at the point of entry rather than the point of exit, which matters more than it sounds: a profession does not disappear when people are sacked, it disappears when nobody new is hired into it and the existing cohort ages out. It also says the effect follows automation rather than exposure in general. Anthropic's own usage data, which is the closest thing to a task-level census, puts augmentation ahead of automation at roughly 55% to 42% and found that about half of all jobs have seen at least a quarter of their tasks touched by Claude. Where the interaction is collaborative, employment holds. Where it is directive, it does not.

The next ten years, which is judgement rather than evidence

Everything in the grid below is our forecast. It is informed by the measured data above; by Anthropic's March 2026 observed-exposure measure, which weighs what models could theoretically do against what they are actually being used to automate at work, and which puts computer programmers at 75% of their tasks covered and data entry keyers at 67%; by the declining and growing roles in the World Economic Forum's survey of more than 1,000 employers; and by sector evidence on adoption, regulation and hiring, some of which is set out in the five points below the table. The cells themselves are a view, not a finding.

Each cell carries two things. The number is our estimate of how much less human working time the category needs to produce the same output as it did in 2022, counting software, AI and robots together. It measures replacement, not employment: a category at 40% can still employ more people than it did in 2022 if demand for its output has grown faster than the machines have, which is roughly what is happening in radiology and in care. The state is read straight off the number, so every row can be checked against the rule. Human, under 15%, means demand for people holds or grows and AI is a tool in the hands of the person doing the job. Shared, 15% to 39%, means the machine does a real share of the work while a person directs, checks and carries the accountability, and rising demand usually absorbs the saving. Thinning, 40% to 69%, means the saving outruns demand, fewer people are needed per unit of output, and hiring at the bottom stops first. Machine, 70% and over, means the default path no longer has a person on it, and human involvement is the exception rather than the rule. Every figure is rounded to the nearest five, because anything finer would claim a precision nobody has.

• Human-led, under 15%
• Shared with the machine, 15–39%
• Thinning, 40–69%
• Mostly machine, 70% and over

Five things are worth drawing out, because they are the parts we would defend if the grid turns out to be wrong everywhere else.

The first is that seniority, not sector, is the dividing line. Almost every row that reaches the machine state does so through its junior end. Entry-level software development and first-line helpdesk go before senior engineering and security architecture, paralegal review goes before contract negotiation, audit sampling goes before audit judgement, CAD detailing goes before the engineer who signs the drawing. That is exactly what the payroll data already shows, projected forward. It creates a problem nobody has solved: the senior roles that survive are currently staffed by people who learned the job by doing the junior one.

The second is that accountability is a moat and capability is not. Contract negotiation, financial modelling, security architecture and structural engineering stay out of the machine column for at least a decade not because a model cannot do the work, but because somebody has to be answerable for it, carry the insurance and sign the thing. A chartered engineer's signature on a structure and a named incident commander during a breach are the same kind of object: a person who can be held responsible. That rule is already written into the places that regulate this work. When the Solicitors Regulation Authority authorised the first AI-driven law firm in England and Wales in 2025, it did so on the basis that named regulated solicitors remain accountable for what the system does. Conveyancing is a reserved legal activity that only an authorised person may carry out. And from December 2027 the EU AI Act requires human oversight of AI used in recruitment, creditworthiness, life and health insurance pricing and access to public benefits. Every one of those rows sits a column later than capability alone would put it. Where the signature has legal or financial weight, the human stays in the loop long after the machine is better at the task. Where it does not, the human goes quickly.

The third is that being the field where AI is strongest does not mean being the field where AI replaces you. Protein and peptide design is the clearest case in the table. Machine learning has been the state of the art in structure prediction and de novo design for several years, further ahead of unaided human capability than in almost any other discipline here, and the row still sits in shared until the 2030s. The reason is that the output has to be made, expressed and tested in a wet lab before anyone believes it, and the bottleneck moved to the bench rather than disappearing. The bench is the slower of the two to automate. Self-driving laboratories that design and run their own experiments exist, but they remain a small market with a young evidence base, and regulated manufacturing expects fixed, validated procedures that a continuously learning system does not naturally provide. So the laboratory row, the routine end of the same building, thins no sooner than the design row does. Electronic design shows the opposite shape because its test bench is a simulator: verification, which is software all the way down, is being handed to AI agents well ahead of the architectural decisions, and that row thins from 2029.

The fourth is that physical work is slower but not safe, and that mechanising a job is not the same as replacing the people in it. Care is the row that has been argued over most, and it is two jobs rather than one. Clinical judgement, escalation and the legal responsibility for a patient behave like the professional rows above and stay human into the 2030s. The physical work of care, the lifting, transfers, night monitoring and mobility support, is already being mechanised, and fastest where the labour shortage is worst rather than where the wages are highest. By 2022, 63% of Japanese nursing homes were using monitoring robots and 26% mobility robots, against a projected shortfall of around 380,000 care workers. But the best evidence on what that did to staffing points the other way from the obvious reading: research on Japanese nursing homes found that facilities adopting robots employed more care workers and nurses rather than fewer, retained them better, and moved staff time towards the parts of care that need a person. That evidence comes from task-specific machines rather than humanoids, so it does not settle what a general-purpose robot will do. Waseda's care humanoid is targeted at facility deployment around 2030 at roughly $65,000, and the home humanoid in chart 9 is taking pre-orders at $20,000 today. Our reading is that this mechanises care well before it thins it, because the sector is short of people rather than oversupplied with them, so the row reaches shared in 2036 and thinning only in the 2040s.

The fifth is that the rows which move last are the ones where the job is to be physically present somewhere awkward and answer for what happens there. Electrical, plumbing and HVAC installation moves once in ten years. Industrial maintenance, last-mile delivery and personal care move once. Astronauts move once, and even that is generous given that the first crew bound for Mars is currently planned to be a cargo of robots. Long-haul freight ends up further along than last-mile by 2047, for the unglamorous reason that motorways are a simpler environment than a terraced street with a locked gate, but over the next decade the two move together. Driverless lorries are hauling paying freight in Texas with nobody on board, and the leading operator aims to have around 200 in service by the end of 2026 and more than 30,000 by 2030. Set against a heavy-goods workforce counted in millions, and with nothing comparable yet running on British motorways, that makes the decade a shared one rather than a thinning one. Warehouse picking moves sooner than industrial fault-finding for a related reason: a warehouse is a controlled environment. Internal documents reported by The New York Times, which Amazon says reflect one team's view, describe doubling what it sells by 2033 without growing its American workforce. That is a halving of people per item over about a decade at the most automated operator in the sector, and slower everywhere else, which is why the warehouse row thins in 2032 and reaches the machine state only in the 2040s.

And then 2047

The last column is ruled off because it is not the same kind of claim as the other four. Twenty-one years is longer than the entire commercial history of the web browser at the time this article's first chart begins, and nothing in the measured evidence reaches anywhere near it. Read it as the direction the rest of the table points if nothing structural intervenes, and assume something structural will.

Two features of it are worth stating plainly. The first is that no row is human-led in 2047. That is not a claim that people stop working. It is a claim about what a default looks like: by then the ordinary path through almost any task we can currently name has no person on it, and a person appears when someone wants one or a statute requires one. The three rows that remain shared are litigation advocacy, clinical nursing judgement and social work, and they have one thing in common that is not technical. Each is a decision a law says a named human must make. The moat at the end is not capability and not even accountability in the commercial sense; it is a legislature that has not changed its mind.

The second is the row we expect to be argued with most, which is control systems and robotics engineering. It is human-led today, it is one of the few genuinely growing engineering disciplines, and we have it reaching the machine state by 2047 — ahead of the plumber, ahead of the carer, ahead of the structural engineer. The logic of the rest of the article does not permit an exemption here. If the capability curve in the first three charts continues at all, the people who build automation are not standing outside it, and the work of specifying and commissioning a machine is closer to software than to unblocking a drain in a Victorian terrace. We would rather write the uncomfortable version than quietly leave our own trade out of the table.

One consequence runs underneath the whole column. The jobs that will be human-led in 2047 are mostly not on this list, because they do not have names yet. Nobody writing a list of occupations in 2004 included prompt engineer, cloud architect or content moderator. The right reading of a column with no human cells in it is not that human work ends; it is that our vocabulary for human work runs out about a decade before the forecast does.

The ten jobs least likely to be replaced

No job in the grid is untouched by AI, so "safe" here means something narrower: machines take no more than 30% of the work by 2036, and a person is still on the ordinary path in 2047. Ranked by the grid's own figures, these are the ten categories that hold out longest. We have left out astronauts, whom the grid protects just as firmly but who number a few hundred worldwide.

Three things protect every one of them, usually in combination. The work happens in a physical place that is different every time: a loft, a ward, a kitchen, a flat behind a locked gate. Somebody has to be legally or professionally answerable for it: the registered nurse, the social worker making a statutory decision, the barrister with rights of audience, the engineer certifying a gas or electrical installation, the named incident commander. And a large part of what is being bought is the person: care, teaching, hospitality and advocacy are partly purchased for the human on the other side. Outside the grid, the same three tests point to paramedics, firefighters, clergy, psychotherapists and most skilled trades, which is why the list reads more like a hospital, a school and a building site than an office.

The uncomfortable pattern is pay. Anthropic's March 2026 analysis found that workers in the most AI-exposed occupations earn 47% more on average than those with no measurable exposure, so the safest jobs are, for now, mostly the lower-paid ones. If the grid is right, that gap narrows from both ends.

Will AI hire humans?

It already does, in two different senses. The first is indirect and far larger: AI now sits between most applicants and the person who hires them. The popular claim that three-quarters of CVs are rejected by software before a human sees them traces back to a 2012 press release with no published method. The better evidence is more modest and more interesting. A 2026 survey of a thousand US hiring managers found that 19% use AI to screen applicants out before a person reviews them, and only 6% let it advance or reject candidates with limited human review. The machine is choosing the shortlist far more often than it is choosing the hire.

The second is direct, and it began this year. RentAHuman, launched on 1 February 2026, is a marketplace on which AI agents search for people by location and skill, give them physical tasks and pay them on completion through an API, often in stablecoins. The tasks are what the grid would predict: collecting parcels, photographing a shop shelf, confirming that a place exists, attending an event, recording household chores so that robots can learn them. A security study of 303 bounties posted on the platform found them spread across 46 countries, and warned of the obvious problem: an automated hiring loop with no person in the middle is also an automated way to recruit people for harmful tasks.

So the answer is yes, and increasingly. By the mid-2030s we expect it to be routine for software agents to commission human work, and for that work to be precisely the parts of the grid that stay human: turning up somewhere, checking something with your own eyes, and putting your name to it. Two things will not change. An AI cannot be an employer of record: a company or a person stands behind every agent that pays someone, and carries the legal duties that come with it. And recruitment is already classed as high-risk under the EU AI Act, with its human-oversight obligations applying from December 2027, so the more consequential the hiring decision, the more certain it is that a person signs it.

The work most exposed to AI is the work we used to call safe because it needed a degree.

15. Which businesses stop existing

One thing this section will not do is name companies. A published forecast that a particular named firm will not exist in 2032 is an accusation rather than an analysis, and it is the kind of claim that gets an article taken down rather than read. Where a business has already closed, we name it, because that is a fact. Everything forward-looking here is about categories and formats, which is where the analysis lives anyway: almost no business model dies because one firm ran it badly.

The measured evidence starts with the format nobody expected to be the story. British retail is usually discussed as a single collapsing thing. It is not. It is two formats moving in opposite directions at the same time.

Retail parks now have the lowest vacancy rate of any format and the strongest rental growth, having been the format written off as a relic twenty years ago. Shopping centres, the format that replaced the high street in the 1980s, have the highest. The building type that wins is the one with parking, a supermarket anchor and a loading bay big enough to serve as a click-and-collect depot, which is to say the one that turned out to be useful to online retail rather than opposed to it. That is the single most useful generalisation in this section, and it applies well beyond shops.

Fifty-three categories, and where each one is going

The states below are deliberately blunt. Growing means more of these businesses, or bigger ones, in that year than today. Holding means the category persists at roughly its present size, though individual firms churn. Consolidating means the model still works but supports far fewer firms, usually as the survivors absorb the rest. Gone means it has stopped being a standalone business worth naming: the function may still exist, but it is performed inside something else. As with the jobs grid, this is our forecast and not a measurement.

• Growing
• Holding
• Consolidating
• Gone as a standalone business

Four patterns run through that table, and they are more useful than any individual row.

The first is that the businesses which disappear are almost all intermediaries. High-street travel agents, contingency recruiters, volume conveyancers, freight brokers, stock photo libraries, single-feature software products: what every one of them sold was access, either to information the customer could not otherwise reach or to a counterparty the customer could not otherwise find. That is precisely the product a general-purpose model and a well-indexed marketplace supply at zero marginal cost. Businesses that sell a thing, hold an asset, or carry a risk are much harder to disintermediate, which is why the data centre, the launch provider and the robotics integrator are all in the growing column and the broker is not. The timing is slower than the logic, though, and the table reflects that. One bank forecast that Purplebricks alone would hold 15% of British property listings by 2022; the whole online agency sector stalled at under a tenth. Getty Images reported the highest revenue in its 30-year history in 2025, with editorial coverage that no image generator can supply making up around two-fifths of it, and its merger with Shutterstock collapsed in July 2026 after the UK competition regulator demanded a divestment. India's largest IT services firms are still broadly growing revenue while AI deflates their legacy work. So the intermediary rows consolidate for most of the decade before they go, and none of them is gone by 2029.

The second is that in manufacturing the factory usually survives and changes hands. European volume car manufacturing is the clearest live case. European vehicle sales fell from 15.3 million in 2019 to below 13 million in 2025, Chinese brands took about 6.3% of Western Europe in 2025 and are forecast at 10.3% for 2026 and 14% by 2030, and several Stellantis plants have been running between 8% and 54% of peak capacity. But Volkswagen ended production at its Dresden plant in December, the first German plant closure in the company's 88-year history, and by May BYD was reported to be in talks to take half of it, which Volkswagen denied; Nissan has consolidated Sunderland onto one line and is working through a plan for Chery to build there. One analysis puts the displacement at around ten legacy European car factories. The buildings, the tooling and a good share of the workforce carry on. The company on the sign changes. That is consolidation, not disappearance, and it is why the row says consolidating rather than gone.

The third is that a category can grow enormously while the number of firms in it collapses, and the two must not be confused. Air taxis are the clearest case, and an earlier version of this table got them wrong: eVTOL manufacturing was marked as merely holding through the 2030s on the reasoning that at least six manufacturers have already failed since 2023, Lilium and Volocopter among them, and that certification economics tend to leave two or three survivors. That is a statement about the number of companies, not about the business. The definition above says growing means more of these businesses or bigger ones, and a firm going from zero passenger revenue to billions is growing however few of its rivals remain. Type certification is expected no earlier than mid-2027, and section 8 sets out how close the approvals now are. The manufacturers row says growing from 2029, and it is split from a new operations row, because the two behave differently: aircraft manufacturing consolidates towards an oligopoly the way it always has, while the business of flying people between vertiports is a service industry that barely exists yet. The grocery delivery row shows the same distinction from the other side: the service keeps growing, but the standalone dark-store operators that pioneered it have largely left Europe, with Getir pulling out of the UK, Europe and the US in 2024, so the growth now happens inside supermarkets and delivery platforms rather than in new firms.

One row still goes backwards and we have left it that way on purpose. Suborbital space tourism operators are marked consolidating now and holding later, because the operator count has gone from two to zero: nobody has flown a paying customer since January 2026, Blue Origin is parked into at least 2028 and Virgin Galactic's next flight is planned for February 2027. The technology works and the seats are sold. A category can pause, shed most of its firms and come back smaller, and a grid that only moves one way would be a grid that had stopped thinking.

The fourth is the one to be most suspicious of, which is how well the growing column matches what is currently fashionable. Data centres, launch providers, AI drug discovery, robotics integrators and battery assembly are all in it, and all five are also where the capital is. Reread section 8 before treating that as a finding. The air taxi sector had exactly this profile in 2023 and has since consumed more than $13bn against effectively zero passenger revenue. Being in the growing column here means we expect more of these businesses to exist in 2036, not that the ones alive today will be the ones that do. Battery assembly already shows how that plays out. Europe's own champion, Northvolt, went bankrupt in 2025 and most European-led gigafactory projects have been delayed or cancelled, while plants built by Chinese and Korean manufacturers have largely gone ahead. The row still says growing, and it is right; the businesses doing the growing are not the ones Europe set out to build.

And then 2047

The last column carries the same warning as the one in the jobs table: twenty-one years out, nothing in the measured evidence reaches it, and it should be read as a direction rather than a destination. It has one property worth dwelling on, though, because it was not designed in and we noticed it only after the table was built.

Nothing is growing in 2047. Not one of fifty-three categories. Data centres, launch providers and AI drug discovery, the three clearest growth stories in the table today, are all merely holding by then, because a market that has finished consolidating into three or four global operators is not a growing category however large it is. The correct conclusion is not that growth stops. It is that every business category capable of growing in 2047 is one we cannot name in 2026, exactly as nobody drawing up a list of business types in 2005 would have written down satellite broadband operator, dark store, or AI drug discovery firm. A forecast table is a list of things that already exist, and the further out you push it the more the interesting answer sits in the space the table has no row for.

The same applies to the thing underneath all of this. By 2047, on the trajectory in the first two charts, the human share of web traffic is somewhere in the low single digits and most retrieval is machine-initiated. The internet stops being a place people visit and becomes a substrate that agents read on their behalf, which quietly removes the shop window that half the categories in this table have spent thirty years learning to compete in. Search-led customer acquisition is not a row here because it is not a business category, but it is the assumption sitting under a great many of them, and it is the one we would expect to break first.

The business that fails is rarely the one doing the work. It is the one standing between two parties who can now find each other.

16. When most work is done by machines: money, income, living and health in the mid-2040s

The 2047 columns in the two grids describe a world in which the ordinary path through most tasks has no person on it. This is the most speculative section in the article, and it is deliberately not a forecast of which world we get. It sets out how such an economy would have to work, what the evidence so far says about each option, and where the choices are political rather than technical.

The problem in one sentence

Wages are how most households get their share of what the economy produces, and taxes on wages are how most governments pay for things. Across OECD countries roughly half of all tax revenue comes from personal income tax and social security contributions, most of it levied on earnings. If machines do most of the work, both halves of that arrangement need replacing at the same time: people need an income that does not come from a job, and the state needs a tax base that does not come from payroll. Nothing in the technology decides how that is done. Every route below is a choice.

Four ways to share the output

There is a fifth route, which is what happens if nobody chooses. The returns flow to whoever owns the compute, the energy and the land, and everyone else competes for the shrinking set of jobs in the list above. That is not a prediction; it is simply what a market does on its own when the scarce input stops being labour. Every previous shift out of a dominant form of work, from the land to the factory and from the factory to the office, eventually came with a settlement about who shared in the gains. The argument about which settlement this one needs is already under way, and supporters and critics of each route above are debating it with more evidence than they had a decade ago.

What gets cheap, and what does not

In a machine-led economy, anything that software and robots can make or deliver gets steadily cheaper: advice, analysis, design, most manufactured goods, a great deal of transport and logistics. What stays scarce gets relatively more expensive, because the price of everything around it is falling. Three things stay scarce: land in places people want to live, energy, and human time. Economists call the last one Baumol's cost disease, after the observation that a string quartet needs four players however productive the rest of the economy becomes, so live music keeps getting relatively dearer. A world where the grids have run to 2047 is one where a legal opinion, a diagnosis and a set of architectural drawings cost very little, while a home in a good location, a reliable electricity supply and an hour of a skilled person's attention cost a great deal. Living standards could rise a long way on the first list while household budgets are still squeezed by the second.

Health

Health is where the two halves of this article meet. The analytical side of medicine is among the most exposed work in the grid: medical coding and billing reach the machine state by 2032, diagnostic imaging by 2047, and AI-first drug discovery is one of the few business categories still growing in the mid-2030s. Diagnosis, triage, monitoring and the discovery of new treatments should become far cheaper and more widely available, including in places that have never had enough doctors. The physical and relational side is the opposite. Nursing, social work and personal care are three of the five jobs least likely to be replaced, and they sit on top of an ageing population. The likely shape of a health system in the mid-2040s is therefore the inverse of today's: far fewer people processing information about patients and far more people looking after them, with care among the largest remaining employers. That is also where the choice of route matters most, because care is labour-intensive by nature and someone has to pay the carers.

What people do with the time

The long historical pattern is that part of every productivity gain is taken as shorter working hours. The evidence on unconditional income points the same way: given money without conditions, people work somewhat less and mostly spend the time on leisure rather than on retraining or starting businesses. Whether that is a problem depends on what work is for. Paid work currently supplies income, structure, status and a large share of most people's social contact, and only the first of those is easy to replace with a transfer. The societies that handle the transition best are likely to be the ones that deliberately rebuild the other three, through care, education, sport, civic life and the kinds of work the grid says stay human, rather than assuming that income alone will do it.

The technology decides what machines can do. It does not decide who owns them.

17. Climate: the one forecast here with physics behind it

Every other chart in this article is driven by adoption, investment or a company's plans, which is why so many of them come with wide bands. Climate is different. The warming trend is set by physics and by emissions that have largely already happened, so the next decade can be projected with far more confidence than next year's humanoid shipments. The uncertainty that remains is almost entirely about human choices, and it mostly affects the second half of the century rather than the next ten years.

The measured position is stark. 2024 was the warmest year on record, at about 1.55°C above the pre-industrial average and the first calendar year above 1.5°C. 2025 was the second or third warmest, at about 1.43°C, slightly cooler only because La Niña conditions replaced El Niño, and the eleven years from 2015 to 2025 are the eleven warmest in 176 years of records. A single year above 1.5°C is not the same as breaching the Paris limit, which refers to the long-term average. That long-term figure, the warming attributable to human activity, reached 1.37°C in 2025 and is rising at about 0.27°C a decade.

From there, the near-term forecast is close to arithmetic. At the current rate, long-term warming passes 1.5°C around 2030. The remaining carbon budget for an even chance of staying below 1.5°C is estimated at 130 billion tonnes of CO₂ from the start of 2026, a little over three years of current emissions. The WMO's latest five-year outlook gives a 91% chance that at least one year between 2026 and 2030 exceeds 1.5°C, a 75% chance that the five-year average does, and an 86% chance that 2024's record falls; with an El Niño forecast for the end of 2026, 2027 is the likeliest candidate. It puts the chance of any single year exceeding 2°C before 2030 at less than 1%.

The longer view depends on policy. UNEP's 2025 assessment puts warming this century at 2.3–2.5°C if every country's pledges are delivered in full, and up to 2.8°C on current policies alone. Current-policy projections were close to 4°C when the Paris Agreement was signed, which is real progress and not enough. Carbon dioxide reached 425.6 parts per million in 2025, and sea level is now 23 cm higher than in 1901.

For the forecasts elsewhere in this article, climate is less a separate topic than a constraint on all of them. About 1.2 billion people, more than a third of the world's workforce, already face workplace heat risk at some point each year, concentrated in exactly the outdoor and physical jobs the grid says stay human longest. Every data centre in the growing column has to be cooled in a warmer world. And long-term warming is likely to cross 1.5°C in the same years that the jobs grid's middle columns begin to bite, so the decade of fastest change in how work is done is also the decade in which the world passes its most-cited climate limit.

Climate is the one chart in this article where almost all of the uncertainty is about us.

What actually follows from this

Measure your own split now the defaults have changed

Find what proportion of your traffic your CDN or edge classifies as automated, and how much of that is AI-related rather than legacy scraping. Since 15 September 2026, new domains onboarding to Cloudflare have training and agent bots blocked by default on pages carrying advertising, while search crawlers stay allowed, so this is a live setting rather than a research exercise. Check which defaults your own zones are running under before you read anything into this month's analytics.

Start the cryptographic inventory this year

The quantum chart above is reassuring about capability and irrelevant to your deadline. NIST moved FIPS 140-2 validations to historical status on 21 September 2026, the EU's national strategy milestone falls on 31 December, and the NSA's acquisition gate for national security systems opens on 1 January 2027, with phase-out by the end of 2030 and mandatory use by the end of 2031. Begin with anything that must stay confidential past 2030, because that material is already exposed to harvesting.

Do not buy a strategy from a chart with one data source

The Linux line is the cautionary tale. An entire news cycle formed around a measurement artefact, and businesses made noise about migrations on the strength of it. Any single-source chart showing a mature market moving fifteen points in a quarter has broken before the market has.