What comes after artificial intelligence

What could change over the next ten years, and where new businesses might emerge

What we can already ask for today

Imagine the office has run out of coffee. You tell your phone, “Remind me to order some at six.” The reminder goes off. Talking and taking notes already work.

Now imagine asking for something else: “Place the order at the usual shop, don't spend more than usual, and pay with the company card.” That is no longer a reminder. The program would have to find the product, confirm it and use real money. That is still not the norm.

And even if the payment went through, the boxes do not appear in the kitchen by themselves. They arrive at a warehouse. There might be a machine there that lifts them, stacks them and returns to its spot. That machine does not chat. It moves things that are heavy.

When people talk about artificial intelligence, they sometimes lump these three tasks together as though they were the same. A program that writes, one that might pay and a machine that carries boxes do different jobs.

For a long time we have described big changes as stops on a train: the computer, the internet, the mobile phone, programs that talk, factory machines and, more recently, ones that run digital errands for you, like Grok Bot.

This time there is not just one stop. What comes next is not a brand or a gadget to give as a gift. It is being able to finish a real errand: someone—or something—pays, and something moves what was bought. Those two things can get in each other's way, because both want electricity, and electricity does not always have a clear path.

If this reading is right, what will be worth paying for is not copying a simple program, which can take minutes. It is permission, trust and the ability to change something that has weight.

Why the order still cannot make it all the way through

For the office coffee to replenish itself, several pieces have to fit. The shop has to accept an order that was not placed by a person tapping a screen. It has to be clear that the company gave permission to spend up to a certain limit. And if the program mistakenly orders hundreds of extra bags, someone has to answer for it.

If any of those pieces is missing, the program can talk beautifully and still fail to finish the errand.

Until the middle of the last century, loading a ship was rather like clearing an attic in a hurry: sacks, crates of different sizes and loose bundles, each lifted aboard separately. At the other port they were unloaded, sometimes mixed up, and loaded onto a truck again.

In 1956, Malcolm McLean changed the way loading worked. An old converted oil tanker, the Ideal X, sailed from Newark to Houston with sealed metal boxes of the same size on its deck.

You did not move the contents. You moved the box, from ship to truck and, if necessary, to train, without opening it.

That only works if the next port has a crane that can grab that box, and if the next truck is also made to carry it. In 1956 this was not yet a completed worldwide network. If every dock invented a different size, the box became a loose bundle again.

The office order has the same problem. Without a shop that accepts the request, a spending limit that can be proven and someone who answers if things go wrong, the box stays on the dock.

In September 2026, a governor of the United States Federal Reserve, Christopher Waller, described this commerce conducted by programs as an early stage. He said it could change how people pay and shop, if it truly grows. He distinguished a program that helps you search while you still pull out your card from one you entrust with the purchase. The big obstacle, he said, is not writing another program. It is trust.

Three specific things are needed. How to prove that the program is authorized to spend. Who is responsible if it buys badly. And how to prevent fraud, because many of today's systems are designed for people, not for a helper that never gets tired.

There is also an open question that shapes industries: will we have common rules, like a box that fits in any port, or will every big company invent its own dock?

People are working on rules so these programs can identify themselves and communicate with one another. The United States standards institute has an initiative underway. It is in progress, not finished. That does not amount to an already settled common payment system.

They can chat all day and, if that is missing, the coffee still will not arrive.

From the order on the screen to work in the warehouse

Suppose the order has been paid for. The boxes still have to move. That is where other helpers come in: not the machine that waves at a trade show so we applaud, but the arms already welding, stacking and driving screws in real factories.

In 2025, more industrial robots were installed worldwide than ever before, after some fairly stable years. Asia took the largest share. The Americas set their own record. Europe, by contrast, contracted.

The organization that keeps those counts says it plainly. A machine that can improvise in almost any job, like a person, is still not reliable and safe enough for everyday use in a factory. Most human-shaped ones remain prototypes or trials. On the plant floor, today's “boring” robot beats the trade-show one on what matters: working, being safe and not wrecking the accounts.

The next step would be for a robot to work even when things are not perfectly organized. Today these machines usually stay inside cages where every movement is measured; the hard part is working in a warehouse aisle like the one receiving the office order.

But neither the program nor the robot runs by magic. The program runs on computers; the robot has motors. Both need electricity, and multiplying them requires that electricity to arrive when it is needed.

A panel turns sunlight into electricity. Clouds can reduce its output; at night it produces nothing. Installing panels is not enough to have power at any hour.

A battery can store some of that electricity for later. How long it lasts depends on how big it is and how much is used. Wires, towers and substations carry that electricity to the office, factory or warehouse. If that grid cannot cope, or if you are not allowed to connect to it, it does not matter how cheap the panel is to manufacture. The electricity gets stuck.

Connecting a new plant is not like plugging in a lamp. You have to ask permission, study whether the grid can handle it and sometimes wait years. In the United States, at the end of 2025, there was an enormous queue of projects—panels, batteries and gas too—waiting for that permission. That queue covers only the United States, not the entire world.

The typical project that actually switched on that year had waited around five years after applying to connect. Of those that requested a connection between 2000 and 2020, only a small share was operating by the end of 2025; most had dropped out of the queue. And it is not a list of factories already promised: it is a queue of competing proposals. The researchers themselves warn that not everything listed will be built.

A gas plant can be switched on when needed, even without sunshine, if it has fuel and its equipment works. That ability to produce without depending on sunshine is also why gas plants are proposed. They do not solve everything: they need fuel, produce emissions and must also connect to the grid.

Two things can be true at once. Manufacturing panels has become much cheaper. And electricity still needs a path. Anyone who believes more panel announcements automatically mean timely, cheap power at the outlet is telling themselves only half the story.

What could change over the next ten years

What follows is not an appointment. It is a weather forecast: useful for taking an umbrella, not for committing yourself to an exact time.

To avoid inventing the future from hunches alone, we used a tool called TimesFM. It does not guess the next invention. It looks at numbers that have already happened, arranged over time, and tests what might happen if that pattern continued more or less unchanged. Imagine, as a made-up example, a building's air-conditioning bills, month by month: the program can try out the idea that next July will resemble previous Julys. It cannot tell you that someone will invent a window that no longer needs cold air.

We compared it with two simple rules: repeating the last value and following the recent pace. It did not always win. Sometimes it did better, especially over short horizons and in series with a familiar rise and fall, such as monthly solar electricity generation in the United States. In other series it lost. One thing is enough: this program helps test an expectation. It does not announce inventions or guarantee a business.

Around 2027, the everyday development worth watching is not a new gadget. It is whether we see more real tests of an office order placed by a program, with a spending limit. A signal to revise this picture would be shops starting to block these requests, or real solar projects—not just announced ones—no longer getting connected to United States grids.

Around 2029, if these programs begin to pay, they will collide with the question of who answers when they buy badly and with a broader battle: common rules or each company with its own dock. Disputes and products to cover those risks could emerge. If in three years there are neither serious trials of payments by programs nor discussion of those rules, this expectation would need revisiting.

Around 2031, in places where almost everyone is already online, some competition could shift from connecting people to obtaining permission to act on their behalf. Bringing connectivity to those who still lack it will remain a separate task.

In electricity, five years is now, in the United States, almost the time a project that does get built takes to go from application to switch: what matters is watching what gets connected, not just what gets announced.

If by then a program can pay an approved supplier within a ceiling, with a trail and someone accountable, we will be closer to making that normal, like paying by card. If all we have are friendlier conversationalists and the same long wait on the grid, we will have advanced less than the headlines promised.

Around 2034 there is no finer number, but two paths to compare. In one, paying through a program and having a machine work outside a very orderly cage become boring and useful. In the other, everything remains stuck on permissions and the unforgiving nature of the physical world. The signal is not a decimal. It is whether those two things show up in daily life or remain exceptions.

Around 2036 that fog does not lift simply by adding two years. The question remains whether, in some places, accountability after an error and machines that truly work have become ordinary. Progress and stagnation in purchases made by programs and in machines can coexist in different countries. There is no basis for claiming the same path will prevail everywhere.

Where new kinds of work could emerge

This is not advice to buy shares or a list of companies. These are three possible kinds of work, presented as hypotheses.

One would be helping a company buy within clear limits. A possible customer orders screws—or coffee—from approved suppliers every month.

The service would check who authorized the purchase, prevent the budget from being exceeded and keep a record for reviewing mistakes. It could charge for running and maintaining those controls. Responsibility if the purchase goes wrong would have to be agreed with those involved in the payment; the program does not settle that by itself.

To start, I would test a recurring order with a small limit and see whether the customer and its suppliers accept the process. The risk is that a big platform closes access, or that rules and fraud make this way of buying unviable.

Another would be installing and maintaining a machine for a specific task. A possible customer is the warehouse receiving those boxes, or a plant that needs an arm to keep working in the small hours.

The service would install the equipment and return when it fails; it could charge for setup and maintenance. I would start with one task and one site, agreeing on how quickly someone must respond.

If nobody buys maintenance separately, we would have to ask why: perhaps the manufacturer or the plant already covers it. That would call this offer into question, not necessarily the use of robots.

A third would be preparing electrical projects that can get connected. A possible customer has panels or batteries awaiting permission in the United States.

The service could review requirements, prepare the necessary studies and compare design changes. It would charge for technical work and for support through the process, not for promising a better place in the queue.

I would start with a specific obstacle: is a study missing, does the grid need reinforcement, or must the design change? The grid operator still decides on connection. Adding batteries does not guarantee acceptance. If procedures are simplified, the offer would also have to change.

In the end there is a reasoned answer, not a promise of income. The phone can already talk. The warehouse machine can already move boxes in very orderly places. What remains difficult is getting a program to pay without throwing the house into disarray, getting a machine to work in a real aisle and getting electricity to arrive when needed. If this reading is right, that—what cannot be copied in an instant—is where the work of the coming years will be.

A note on how we examined the numbers

TimesFM extends lists of numbers already published. We compared it with two simple rules: repeating the last value and following the recent pace. It did not always win. It lost when estimating how much new solar capacity is added each year and in a published tally of the computing used to train large models; in both cases the rule of following the recent pace beat it. For monthly solar generation in the United States, it often did better than those rules over short horizons. The result depends on the series and horizon; it does not always beat a simple rule.

We used public data on internet access, mobile subscriptions, installed solar capacity, monthly solar generation in the United States and published computing for training large models. Capacity indicates the power all those installations can deliver together under reference conditions. It does not count how many panels there are or how much energy they produce over a month or year. In other words: how large the collection of installations is, not how much it actually produced.

The latest official data on internet use and worldwide solar capacity are for 2025. The latest month of United States solar generation we used is July 2026.

In 2005, using the internet was something a minority did. In 2025, roughly three out of four people worldwide were already online; about a quarter remain offline. If you add the same amount every year as was added since 2005, you end up inventing more connected people than people. Mobile phones, counting the entire world, already have more subscriptions than inhabitants, because many people have two lines or a little card in another device. That does not mean every human being carries a phone, or that there are no poorly served regions and people. Anyone building a business whose only thesis is that the whole world “still has no internet” or “still has no mobile phones” is confusing worldwide expansion with the gaps that remain.

We have not checked whether any of these series were among the data used to train the model. That was neither demonstrated nor ruled out. It is a limitation of the evaluation, not a reason to trust its forecasts more.

We requested total solar capacity and the amount added each year separately, and the results do not fit together. It is like calculating an account balance and its transactions separately: the two should reconcile. That is why we do not use those lines as though they told the same story.

The computing series is a cumulative maximum: the highest record published up to that date. Such a maximum cannot fall. If the program draws a late decline in that record, it is a flaw in the line, not a finding. Installed solar capacity could fall slightly if old panels are retired; a late decline in the model is not well justified, and we do not use it as a conclusion.

For eight or ten years, there is not enough history to examine the program with the same rigor. Those long lines are scenarios.

---

References

1. Engineering and Technology History Wiki (IEEE), “Milestones: Pearl Street Station, 1882”. https://ethw.org/Milestones:Pearl_Street_Station,_1882
2. Frank Gallant, “Flashbacks: Rural Electrification by the Numbers”, Rural Electric Magazine (NRECA), December 7, 2016. Secondary source; the original Rural Electrification Administration table was not opened. https://www.cooperative.com/remagazine/articles/Pages/flashbacks-rural-electrification-cartogram-numbers.aspx
3. Guinness World Records, “First container ship” (Ideal X, April 26, 1956). Accessed November 7, 2026. https://www.guinnessworldrecords.com/world-records/62889-first-container-ship.html
4. World Bank / ITU, “Individuals using the Internet (% of population)”, indicator IT.NET.USER.ZS, world, latest official year 2025. Extracted November 7, 2026. https://data.worldbank.org/indicator/IT.NET.USER.ZS
5. World Bank / ITU, “Mobile cellular subscriptions (per 100 people)”, indicator IT.CEL.SETS.P2, world, latest official year 2025. Extracted November 7, 2026. https://data.worldbank.org/indicator/IT.CEL.SETS.P2
6. IRENA, Renewable Energy Statistics (2026), processed by Our World in Data, “Solar photovoltaic capacity” (grid-connected photovoltaic capacity). IRENA.org returned a block from the working environment; the OWID processing is opened here, not the primary Excel file. Metadata updated July 28, 2026. https://ourworldindata.org/grapher/solar-pv-cumulative-capacity
7. U.S. Energy Information Administration, Monthly Energy Review, table 10.6, series SOTEPUS (United States net solar generation). Latest month used: July 2026. https://www.eia.gov/totalenergy/data/monthly/
8. International Federation of Robotics, World Robotics 2026 – Industrial Robots, public executive summary. IFR's own forecasts for 2026 and 2029 were not treated as observations. https://ifr.org/img/worldrobotics/Executive_Summary_WR_2026_-_Industrial_Robots.pdf
9. Joseph Rand and others, Queued Up: 2026 Edition, Lawrence Berkeley National Laboratory, June 2026 (data through the end of 2025). https://emp.lbl.gov/queues
10. Christopher J. Waller, “Payments in the Age of AI Agents”, speech, Board of Governors of the Federal Reserve, September 29, 2026 (file waller20260928a.htm). One governor's opinion, not a regulation. https://www.federalreserve.gov/newsevents/speech/waller20260928a.htm
11. National Institute of Standards and Technology, “AI Agent Standards Initiative”. An ongoing initiative, not a finalized standard. https://www.nist.gov/artificial-intelligence/ai-agent-standards-initiative
12. Google Research, TimesFM (Apache-2.0 code; version 2.5 weights Apache-2.0). The google/timesfm-2.5-200m-pytorch model was used. TimesFM 3.0 weights were not loaded. https://github.com/google-research/timesfm and https://huggingface.co/google/timesfm-2.5-200m-pytorch
13. Epoch AI, “Data on Notable AI Models”, accessed November 7, 2026. Research compilation, not official statistics; specific model names are not used here as launch facts. https://epoch.ai/data/ai-models