What Is "AI 2040"? The Five Plans from the "AI 2027" Authors, Explained

Hello, this is Ryuta Hamamoto from TIMEWELL.

In 2025, a future scenario called "AI 2027" became a global talking point. It was a document that told, year by year and in fine detail, how superintelligence might arrive within a few years and how a US-China development race might spiral out of control. Many readers came away with a chill down the spine, but it was, in the end, a forecast: "if things keep going this way, here is what could happen." Now the same team of authors has put out a new work that steps into the next question, "so what should we actually do about it?" That work is "AI 2040 Plan A."

Why is it worth reading this document now? There are two reasons. One is that, over the past year, the debate around superintelligence has been shifting its center of gravity from "when will it arrive" to "if it arrives, how do we steer." The other is that, although this document looks like a story about national security and international politics, it throws up a question with the very same structure as a business decision. Do you choose speed, or do you choose safety? Do you sprint ahead alone, or do you keep pace with those around you? This is at once a decision that states face over semiconductors and a decision that executives face when they wonder how to bring AI into their own companies.

This article walks through what AI 2040 describes, how it differs from its predecessor, what timeline the first-choice Plan A follows, how the verification mechanism that underpins that plan works, how the other four roads it contrasts differ from one another, and how far you should take this document on faith, following both the primary sources and the criticism as we go. I will unpack the jargon as it comes up, so readers new to the AI debate can follow along without worry. If the distance between your own company and AI is on your mind, checking your starting point first with our free AI literacy self-check will make the second half of this article a good deal more concrete.

What "AI 2040 Plan A" actually is

Let me start with what this document is, because getting that straight changes how you read the whole thing.

AI 2040 was produced by a research non-profit called the AI Futures Project. The core members who worked on the earlier AI 2027 are the same people writing here, and there are six authors. The best-known figure among them is Daniel Kokotajlo, a researcher formerly at a major AI developer, alongside Thomas Larsen, Eli Lifland, Romeo Dean, Brendan Halstead, and Ryan Greenblatt. For the AI Futures Project this is the third major release, after AI 2027 and a model for forecasting the future. That "model for forecasting the future" is a computational model the team published in order to estimate, in numbers, how fast AI's capabilities will grow. They are not placing years on the calendar by gut feeling alone; they try to build up a timeline within a consistent framework, from how fast computing resources grow and how much the automation of research contributes. A model is, of course, a model, and its results swing widely depending on the assumptions you feed it, but it is worth keeping in mind that their timeline is at least not pure guesswork. In other words, this is not a one-off idea. It is best understood as the next step in a body of work that the same team has been building up under a consistent set of concerns. The official site is ai-2040.com, and the announcement went out on the team's blog.

On the timing of publication, let me be careful. The first edition was published around the spring of 2026, and it is said to have been edited repeatedly since. The length is often described as running to something like ninety pages, but that figure comes from secondary sources such as summary articles, so here it is safest to take it simply as "a fairly detailed, long-form work." News outlets covered it one after another, but for the exact page count and the fine points of the dates, checking the text of the primary material at ai-2040.com is the sure route.

The heart of it comes down to a single idea: deliberately delaying the arrival of superintelligence. Superintelligence means an intelligence that surpasses the best human experts in every field. A related term is AGI (artificial general intelligence, an AI that can handle a broad range of work at roughly human level); if the differences between such terms make you uneasy, it helps to first glance at our primer on the basic terms of AI and DX, which makes the discussion ahead easier to read. AI 2040 sets down the premise that such superintelligence could otherwise appear all at once around 2030, and then constructs, as a series of year-by-year events, a path along which humanity deliberately delays it until 2040 so that we can keep control. The 2040 in the title points to the destination reached as the result of that intentional slowdown.

And how you grasp the character of this document is the most important thing of all. AI 2040 is not merely a forecast. The authors themselves position it as "a recommendation for forecasting and steering the course of an evolving AI," and some commentators frame it as "a governance plan rather than a forecast." Governance means how a society controls and steers a technology that could otherwise run out of hand. Where a forecast speaks of "what will happen," a governance plan speaks of "how to steer." Even when they treat the same future, the vantage point is entirely different.

That difference shows plainly in the document's form. AI 2040 takes the shape of presenting five plans to the candidates in the 2028 US presidential election, in effect saying, "if you become president, why not proceed like this?" Why address it to presidential candidates? Because the authors see the decisive steering over superintelligence as resting, in the end, not with companies or researchers but with a nation's highest leader. However refined a technical idea may be, the authority to put it into practice lies with the political leadership. So they write it in a form that reaches the decision-makers directly. This choice is what sets the document apart, as a real policy proposal rather than a mere thought experiment.

The story's starting point is 2029. In that year, the document says, the US president must answer a single question. Do you charge into a race to hand data centers, factories, and even the control of weapons over to AI faster than China does, or do you stop? If nothing is done, by 2030 fully automated AI research and development is achieved, and within that same year the road to superintelligence opens. So 2029 is placed as the last fork in the road, where you decide whether to hit the brake or the accelerator. How you handle this single point branches the next decade and determines which of the five plans you follow. That is precisely why the story begins not at the distant 2040 but at the imminent 2029.

Here is one thing I want to stress at the outset. The authors are not saying that if things proceed as this plan describes, a good future is assured. On the contrary, as their own estimate, they are reported to have put the probability of reaching a good outcome at around 42 percent. That figure is their subjective read, not an objectively calculated probability, but the fact that even their first-choice plan is, by their own account, less than an even bet says a great deal about both the document's honesty and the sheer difficulty of the problem. You do not need to swallow the number itself, but the recognition that "even this leaves a hard road ahead" is worth sharing.

How it differs from "AI 2027"

What most people want to know is how this compares with the talked-about first work. Pin this down and the new work's place becomes clear.

AI 2027 was, above all, a forecast. In English, a forecast is a view of "how things might turn out if they keep moving at the current pace," and here it was told in story form. Superintelligence arrived rapidly across 2027 and 2028, the development race between the United States and China overheated, and from there two endings were laid out: one branch where control of AI is lost and heads toward catastrophe, and another where the two countries cooperate to avoid the crisis. It was a document heavy with the color of warning, meant to make readers ask, "isn't it dangerous to keep charging ahead like this?" The very AGI timeline that the first work suggested is something we also take up in our piece organizing how realistic the arrival of AGI toward 2027 really is, and reading the two together helps you follow the thread.

AI 2040, by contrast, is a proposal. It describes not "how things might turn out" but "what should be done." The authors named this plan "Plan A" to convey that it is "the least-bad plan we currently know of." It helps to understand the new work as a timeline that answers the alarm the first work sounded: concretely, what moves would let us land relatively safely? If the first work is a "map of the crisis," the new one is an "itinerary for avoiding that crisis," and one that recommends a single route after comparing several.

There is another important difference. The first work was close to a two-way choice between "catastrophe or avoidance," whereas the new one widens the options to five and lays out the gains and losses of each side by side. The road to avoiding the crisis is not a single lane; it branches into several depending on how thoroughly you slow down and how you posture toward China. Within that spread, the option the authors place first is Plan A. This stance of "lining the options up and comparing them" is the backbone of the document, and it also ties into the criticism we will come to later.

That said, the authors themselves note that reality is likely to move somewhat faster than this scenario. In other words, the year 2040 is not a forecast that "this will certainly happen" but something closer to a target, "it would be ideal if we could proceed with roughly this much care." Layered on top is the recognition that many AI developers estimate a high likelihood of building AI smarter than humans within the next one to ten years. The sense of urgency the first work raised, that "there may not be much time left," carries over into the new work unchanged.

The timeline the first-choice Plan A draws

So let me trace, carefully, exactly what flow of time the authors' first-choice Plan A moves through. This is the document's main reading. Because it is a story divided into year-by-year sections, I will explain it by following the major turning points. Let me note in advance that everything below is an assumption within the scenario, not a settled fact or a confirmed forecast.

The starting point is the case where nothing is done. In the scenario, around 2030 fully automated AI research and development is achieved. This part deserves a few extra words. AI development today is a human endeavor: researchers come up with ideas, design experiments, interpret results, and decide the next move. Fully automated AI research and development refers to a state in which AI itself can run this whole loop with almost no human hand. Even without the human researchers moving much, AI designs AI, improves it, and produces still smarter AI, and that cycle begins to turn.

Once this cycle starts turning, what happens? A smarter AI can build the next AI faster and better. That next AI builds the one after it faster still. Progress accelerates itself and swells exponentially. This is called an "intelligence explosion." Picture a snowball rolling down a slope, growing larger and faster the more it rolls. Left alone, this explosion could bring superintelligence within that same year, the scenario supposes. Progress that ought to take decades gets compressed into a short span. That is exactly why you need the reins ready before the acceleration begins, and that concern is the starting point of Plan A. Because the development is so fast, you cannot afford to be caught flat-footed.

Plan A deliberately slows this runaway acceleration. Tracing the main milestones on the timeline: first, in 2029, the United States and China agree to "avoid a reckless race to superintelligence." This is the linchpin of the story. Then in 2030, where an absence of intervention would drive fully automated AI research and development straight to superintelligence, this agreement holds it back. From 2030 to 2035, AI is strengthened step by step within a range humans can understand and control. The destination of this period is AI at a level rivaling the top human experts. What is worth noting is that it stops just short of surpassing humans. Then in 2035, development is deliberately paused at that top-expert level. The idea is to take the time here to make human control secure and to firm up safety. Finally in 2040, development resumes in earnest toward superintelligence. That is the storyline.

Within this timeline, I want to draw attention to two dividing points. One is that from 2030 to 2035 it holds AI to a "range humans can understand and control." The ceiling for AI in this period is set at rivaling the top human experts, no further. Stopping just short of surpassing humans is meant to keep AI in a state where humans can verify its inner workings and step in if something is off. Make it so smart that human understanding can no longer reach it, and humans lose any way to confirm whether that AI is truly safe. So the care built in here is to make it smarter only within the range you can still verify.

The other is the 2035 pause. Having come all the way up to top-expert level, it deliberately halts there. It seems wasteful, but this is by design. The idea is to stop development here and take time to confirm whether that AI truly acts in line with human intent and whether it will avoid dangerous behavior. Only after firming up the footing of safety does it take on the final climb toward superintelligence. It does not try to raise capability and confirm safety at full speed at the same time. This is the symbol of Plan A's caution.

Rather than racing up the hill, this is a plan that climbs while stopping at the key points, and holding that image makes the whole thing easier to grasp. In 2029 you promise the one running beside you, "let us slow down together"; in 2030 you brake at the mouth of the runaway; through the first half of the 2030s you strengthen AI only within human reach; in 2035 you pause to firm up the footing; and in 2040 you climb the final slope at last, fully prepared. This idea of "climbing while managing your speed" is the spine of Plan A.

At this point a natural question arises. If states merely make a verbal promise to "slow down," where is the guarantee that the other side is not secretly sprinting ahead in the background? Why would great powers competing against each other keep a promise that ties their own hands? In fact the answer to exactly this question is the most original part of Plan A, and it leads into the idea of "verifying without relying on trust" that we will look at closely in the next section. Now that you have the big picture of AI in mind once, keep a corner of your thoughts on where your own company stands within this flow.

How to verify "without trusting," and what happens to the economy

What sets Plan A apart from other proposals is that it does not rest the promises between states on "trusting the other side." The idea, put in English roughly as verifying without relying on trust, does not count on the other side's goodwill; it sets out to build a mechanism in which the promise cannot be broken. In the world of international politics, there is a long history of failures where one side disarmed on faith and was betrayed. So this plan places verification, not trust, at its foundation.

The core of that verification is an idea called compute governance. Compute means computing power, that is, the computing resources needed to run AI, and governance means to rule or to manage. Put together, the idea is not to police AI itself directly but to keep the reins on AI's advance by managing the computing resources that AI cannot run without. Why is this effective? Because training a frontier AI takes an enormous amount of computing power. Software can be copied across borders in an instant, but the physical chips that produce that computing power, and the vast facilities that run them, cannot be hidden so easily. So instead of chasing the hard-to-see innards of AI, you take hold of the visible chips and facilities. That is the crux of compute governance. AI 2040's verification mechanism extends this idea to agreements between states, and it is built in roughly three layers.

The first is controlling the supply of chips itself. Running advanced AI takes large quantities of specialized semiconductors, so-called AI chips. For these chips, the companies that design them and the factories that actually manufacture them, called foundries, number only a very few in the world. On the design side are firms such as NVIDIA and Huawei; on the manufacturing side, firms such as TSMC, Intel, and SMIC. That the numbers are limited means that by taking hold of this point, you can grasp, at the upstream stage, who could obtain how much computing power. If there are few taps, the flow of water is easier to manage.

Let me add a little more on why so much can be staked on the single point of chips. Making the most advanced AI chips requires manufacturing equipment called extreme ultraviolet (EUV) lithography systems, which etch circuits at the nanometer scale, and the firms that can supply these number, in practice, only a handful worldwide. The most advanced factories that install such equipment take years to build and investment on the order of trillions of yen, and only a countable few exist in the world. In other words, AI's computing power is not something anyone can replicate overnight the way software can; it is a deeply physical resource that can only be increased with limited facilities and long stretches of time. This "hardness to increase" and "hardness to hide" is exactly the ground for choosing chips as the single point of control. Rather than counting weapons one by one, knowing that the factories that can make them exist in only a few places worldwide makes them far easier to manage. Compute governance is discussed as a realistic option rather than a fantasy precisely because of this lopsidedness in the industrial structure.

The second is tracking chips that are already out in the world. What comes into play here is the physical nature of AI chips. To run the most advanced AI, you need to gather chips into vast data centers on the scale of a small city fitting inside. Because they demand enormous power and cooling, it is hard to quietly scatter and hide them; they can only be operated as large clusters. As a result, by following sales and customer records, the author estimates that the whereabouts of up to about 98.5 percent of existing chips can be tracked. Put the other way around, the claim is that only a very small remainder cannot be traced.

The third is aggregating the tracked chips into specific data centers that regulators can monitor. The document sketches a concept in which all chips are gathered into monitored, auditable data centers called white sites, where the US and China verify each other's operations. When chips are scattered here and there, monitoring is hard to reach, so you gather them into designated places and put them in a state where each side can look in. It resembles how a bank keeps money together in a vault under strict control. By gathering them, management and verification actually become easier.

Backing these three layers is an idea with a slightly menacing name, "mutual assured compute destruction." The name is hard, so let me build it up step by step. What underlies it is the Cold War idea of nuclear deterrence. In the era when the US and the USSR each held vast numbers of nuclear weapons, if either fired first, retaliation meant it too would surely perish. So neither could fire in the first place. This state of "no one fires because firing means mutual ruin" is, ironically, said to have prevented war. Mutual assured compute destruction transposes this to AI's computing resources. Even if one country breaks the agreement and tries to steal a march, the aggregated computing resources are arranged in advance so that they can be mutually neutralized, so the moment a side defects it loses its own computing power too, and no single side can seize a decisive advantage. If you set up a state where the other side's betrayal breaks both sides' computing power, the motive to betray disappears in the first place. What is clever about this idea, unlike nuclear weapons, is that it does not cost human lives; what is lost is only computing resources.

Alongside it is the idea of a "moving capability ceiling." This is a mechanism in which the US and China agree to raise the ceiling on AI capabilities little by little at the same pace, confirming each other's step under monitoring as they go. Picture raising the roof together, gradually, so that neither side alone breaks ahead. It resembles matching your stride when running a three-legged race so that one runner does not bolt off. Make things transparent and watch each other, hold each other's computing resources hostage, and raise the ceiling together. With this three-part setup, Plan A tries to make a slowdown promise hold even with a party you cannot trust. That is the whole picture of Plan A's verification.

So, at the end of proceeding this carefully, what does the document say happens to the economy? This is one of the boldest parts of the document, and also the part that has drawn the most criticism. The document introduces a "citizen's dividend," a mechanism for distributing to the public the enormous wealth AI generates. Per person, it says the figure could reach roughly 25,000 dollars a year in 2033, and, after adjusting for extreme deflation (a sharp fall in prices) in 2035, roughly 1.6 million dollars. The digits leap up all at once because the premise is an explosive rise in productivity from AI. And in 2040, a development is drawn in which the layer corresponding to the sovereignty of governance is entrusted to a so-called aligned AI, one adjusted to follow human intent completely, with an intelligence explosion following soon after. That is the close of the story.

The word alignment is central to this debate, so let me add to it. Alignment means making an AI's goals and behavior properly follow what humans truly want. If a very smart AI starts pursuing goals that diverge from human intent on its own, its very smartness makes it unmanageable. A common analogy is the tale of a genie who grants wishes. It grants your wish to the letter, but if it does not read your true intent, it invites disastrous results. A powerful AI is the same: it is dangerous unless it follows the true wish behind the wording rather than the literal instruction. That is exactly why, before raising capability, aligning intent first becomes decisively important. That Plan A pauses in 2035 also makes sense once you understand it as using that time for the work of aligning intent.

That said, these depictions of economic figures and the handover of governance are assumptions placed there to make the story hold together; they are neither measured values nor promises. The figure of 1.6 million dollars per person for the citizen's dividend rests on the premise of "extreme deflation," meaning that as AI raises productivity explosively, the prices of all goods and services fall dramatically. Pulling out only the face-value digits and reading it as "each person gets over 200 million yen" is hasty; it makes sense only paired with the premise about prices. Rather than being surprised by the size of the digits, it is healthy to receive it as a problem statement: "if AI really does grow this smart, the very premises of money, work, and even governance could change at the root." As we will see in the next section, this economic depiction is also where the criticism has concentrated most.

For executives wrestling with where to bring AI into their own companies, before dismissing such extreme scenarios as a distant story, you may find that the question we grapple with day to day in our AI consulting service WARP, "how do you balance speed and safety," shares the same root. Whether the subject is a nation or a company, the design philosophy is common: not all-out acceleration, not pure wait-and-see, but steering in a planned way somewhere between the two.

Contrasting the four roads other than Plan A

What makes AI 2040 distinctive is that it does not simply push Plan A; it lines up responses of differing character and holds them in contrast. The options are named from the first-choice Plan A through Plan B, Plan C, Plan D, and Plan S, with Plan A positioned as the recommended one. Let me follow the gist of each. First, a word of caution: for the labels and points of emphasis of the plans, there are places where the wording differs between the official site and secondary summaries. Here I will mainly introduce the broad character each option holds, and avoid asserting the fine labels. If you want the precise divisions, do go to the primary material, the text at ai-2040.com.

The first-choice Plan A is the verified slowdown I have explained so far. In English it corresponds to Verified Slowdown. Through transparency and international agreement it deliberately slows development; the US and China show each other their hands, back it up with mechanisms for verification and deterrence, and move while stopping at the key points. Of the five it is the most labor-intensive, and the one that cannot hold together without passing through the hard terrain of international cooperation. That the authors place it first while estimating its success probability modestly is an honest reflection of exactly this character: ideal, but with a high bar to realize.

Next, Plan B is the line of containing China. It corresponds to a positioning of Fight China, and it is framed as an option in which the US forms a coalition it leads and buys itself time by, among other things, obstructing the other side's efforts. Where Plan A is a "let us slow down together" idea, Plan B is closer to a "secure safety by holding the other side down" idea. In placing its weight on securing an advantage rather than on cooperation, it is also an option that takes on the risk of the confrontation growing sharper.

The third, Plan C, is a line that uses an early lead for the sake of safety. It corresponds to a positioning of Burn the Lead, the idea being that the AI project running at the front deliberately turns part of that lead toward safety measures. It may include imposing strict domestic safety regulation and, in coordination with other frontier firms, aiming for a degree of slowdown. Without stepping into as thorough an international agreement as Plan A, it can be understood as an intermediate stance that tries to use the advantage it has gained for "safety" rather than "speed."

The fourth, Plan D, is the option of a race that aims for superintelligence as fast as possible. In English it corresponds to Race to ASI. ASI stands for artificial superintelligence; in short, it keeps regulation to a minimum, has the frontier firms charge into an intelligence explosion at nearly maximum speed, and devotes only a tiny fraction of resources to safety measures. This sits at just the opposite pole from Plan A. In prioritizing speed above all, it is an option that takes on a large risk of control failing to keep up, and it is also the option closest to the situation in which countries and firms are currently continuing to compete, in effect the "if we proceed as we are" case.

The fifth, Plan S, is the option of halting development itself entirely, worldwide. It corresponds to Shut It All Down, and the initial S stands for shutdown, that is, a halt. If the danger of superintelligence is simply too great, the idea goes, we might as well stop development across the whole world. It is the option at the other pole, cutting off the risk at its root in exchange for giving up the benefits AI could bring. Including the political difficulty of realizing it, it is positioned as an extreme choice.

Lining the five up this way, the design behind the document comes into view. These five roughly line up on a scale of how much you prioritize "speed" versus "safety." All-out speed is Plan D; casting speed aside entirely for safety is Plan S. Between those two poles sit Plan B, which goes to secure an advantage through containment of China, Plan C, which turns the advantage it gains toward safety, and Plan A, which keeps step through international cooperation and slows down. Whichever you choose, you give something up in exchange for what you gain. Take speed, and the risk of losing control rises; take safety, and you fall behind in the competition or defer the benefits. This tension of "raise one and the other falls" is, I think, what the document most wanted to convey.

So where does today's reality stand among these five? Given today's situation, in which countries and firms keep developing as they compete and the resources turned toward safety are relatively small, it is closest, of the five, to Plan D. Put the other way around, if you want to move to a verified slowdown like Plan A, you have to steer sharply away from the current line of extension. Agreements between states, the tracking and aggregation of chips, mutual monitoring, none of these can be put in place overnight. That the authors set Plan A's success probability modestly can be read as reflecting exactly this gulf between where we stand and the goal. The contrast of the five points to the desirable endpoint while at the same time driving home how far the road to it runs.

There is one more point worth holding, in fairness. This document does not depict any particular country as the villain. Both the United States and China are treated as parties with the rational motive of wanting to avoid catastrophic risk, and it is on that basis that the 2029 mutual-slowdown agreement can hold. Rather than stoking conflict, it tries to draw a blueprint for cooperation: if neither can open a decisive gap, then rather than forcing the competition onward, keeping pace through transparency and verification is safer for both. Whether the practice of international politics really moves that way is a separate matter, and that is where the criticism in the next section presses. But as the scenario's intent, it is built neutrally, and that is worth not misreading.

What we work on together with executives through our AI consulting service WARP is nothing other than this part: designing the moves without swinging to extremes, keeping both risk and benefit in view. Whether the subject is a nation or a company, the pattern of thought is the same: line the options up from both ends and weigh the middle.

How far you should take this plan on faith

I have traced mainly the attractive parts of Plan A so far, but reading this document as prophecy is dangerous. On top of the authors' own reserved posture, a fair amount of criticism has come from outside. Looking fairly at both sides leads to a healthy way of receiving it.

First, the authors' own reservations. As noted, even for the first-choice Plan A, they present a self-estimate that the probability of reaching a good future is around 42 percent. It is not optimism all the way down. On the verification mechanism too, the estimate that about 98.5 percent of existing chips can be tracked means, put the other way around, that the remaining 1.5 percent or so could slip past tracking. The room for some actor to secretly advance development using those slipped chips is not zero. This point is raised even by commentators favorable to Plan A.

Outside criticism, too, comes in several forms of differing character. Here I will introduce both the reservations from commentators broadly favorable to Plan A and the stances that doubt it more fundamentally. Placing both side by side brings the document's position into sharp relief.

First, reservations from a commentator who shows understanding for Plan A. One well-known writer, while taking a stance of defending Plan A, frankly lists a number of weaknesses. One is the problem of the chips that slip past tracking, touched on above. About 1.5 percent of existing chips are estimated to be untraceable, and this gap leaves room for some actor to secretly advance development. The reason 1.5 percent is not to be taken lightly is that, in the development of superintelligence, the slightest march stolen can create a decisive gap. The second is the premise of alignment. This plan relies, to a considerable degree, on a genius-level AI following human intent, that is, on alignment succeeding. But no one can guarantee, at present, that an AI that smart will truly be safe. The third is doubt over the durability of international cooperation. Can the US and China keep an agreement to slow down for a dozen-odd years? Look back over the history of arms control and there are no few cases where promises between great powers were torn up partway or hollowed out. Recalling how frameworks for nuclear disarmament and international agreements over nuclear development wavered with changes of government or shifts in the situation, this writer suggests that counting on long-term cooperation may be too optimistic. Having laid out these weaknesses, the writer calls Plan A itself an "insanely bold undertaking" and notes that every part of it is provisional. Even the side that defends it holds this careful a posture.

One of the sharpest criticisms is aimed at the economic depiction. A commentator at a conservative think tank takes up the document's premise that GDP grows by 50 percent in 2032, and judges it harshly as closer to "science fantasy" than to sincere economic analysis. He also points out that a depiction of unemployment reaching 12 percent at the end of the 2030s is not politically sustainable at all. The criticism is that the unreality of the economic part undermines the credibility of the whole scenario. If one number is outrageous, even the other claims come to look doubtful. I find this a valid caution too, that figures on the order of the citizen's dividend should not be taken at face value.

There is also criticism that focuses on political realism. A commentator well-versed in marketing and AI said Plan A is coherent as theory but extremely unrealistic politically. On that basis he dismisses anything beyond 2028 as "pure guesswork." What makes this view instructive is the conclusion drawn from it: rather than debating the distant endpoint of 2040 this way and that, the near-term choice, which plan you steer toward now, is far more important. The point that the step you take at this very moment is decisive, more than a refined timeline of the distant future, carries straight over to business decisions as well.

Organizing the criticism this far, the points of contention narrow to roughly three. Can it really be verified technically (the problems of chips slipping past tracking and of alignment)? Can it really hold up as international politics (the durability of US-China cooperation)? And is the economic depiction not detached from reality (50 percent GDP growth and an outsized citizen's dividend)? Put the other way around, that is also an assessment that, so long as these premises do not collapse, the logic of Plan A itself holds together. Much of the criticism takes the form of "the logic is clear, but the premises are soft," and it is worth noting that it does not wholly reject the plan's frame. When you read, the wise way to engage is to set the authors' claims against the critics' points and examine, one at a time, how far you yourself can believe each premise.

Even laying out this criticism, I think this document is worth reading. More than whether the forecast comes true, the frame of thinking itself is instructive: confronting the extreme prospect of superintelligence head-on with concrete means such as transparency, verification, and international cooperation. Rather than swallowing the figures and years, distinguish premises from reservations and take in the whole, criticism included. Read this way, AI 2040 is not a document that stokes fear but a solid piece of material for thinking about an uncertain future in a structured way.

What Japan and the rest of us should prepare now

Finally, let me write about how to take this scenario as a story about your own company. Hearing "an inter-state race to superintelligence" may make it feel like a distant world, but the message at its root connects directly to management.

First, from a Japanese vantage. The story of AI 2040 sets the US and China at center stage, and Japan's name does not come to the fore. But that is exactly why there is something to think through. The premise is that an era is drawing near in which the supply of chips, the location of data centers, and AI's capability itself connect directly to a nation's bargaining power. National policy, in fact, moves on the same horizon. That Japan's medium-to-long-term economic and fiscal plan looks ahead to fiscal 2040 is something I touched on in our explainer on the Basic Policies 2026, and the moves by which AI seeps out of the screen into the physical world, such as robotics and autonomous driving, are covered in our piece organizing Japan's strategy on physical AI. That the same number, 2040, comes up both in the context of national security and in the context of domestic economic and fiscal policy is, I think, no coincidence. How we use this next decade or so is being asked of both the state and companies.

Why does a story that puts the US and China at center stage reach all the way to a single company in Japan? Let me fill this in too. In the world AI 2040 draws, AI's capability itself becomes the source of a nation's bargaining power, and what governs that capability is chips and computing resources. Scale this structure down and it applies to companies as well. What computing resources you can reach, what AI you have managed to build into your operations, and whether you have the people who can put it to use: this difference will sort the competitiveness of companies from here on. It sits well once you think of it as a miniature, within an industry, of what happens between nations. And in Japan's case, a flow in which the country steers toward investment and growth and backs AI and semiconductors is moving on the same horizon. In this phase where the policy tailwind is blowing, what is being asked is whether you are ready to receive it.

On top of that, the single biggest implication executives should draw is the premise that AI could rapidly approach a level rivaling experts over a horizon of a few years to a decade. If even half of that premise holds, the way many kinds of work are done can no longer be spoken of as a simple extension of the present. What matters here is not swallowing the scenario's years as prophecy. Rather, it is accepting the uncertain premise that "AI could become considerably smarter from here" and beginning, now, to think about what your own company should have ready.

The contrast of the five plans teaches, in fact, the same thing. Both the fastest race and the full shutdown carry extreme risk, and in reality you can only steer carefully between the two poles. A company's decisions, too, call for neither betting everything on AI nor standing still in wait-and-see, but proceeding in a planned way somewhere in between. And as the earlier criticism showed, the step you take now, which way you move, matters more than a refined picture of the distant future. If a nation's step is which plan it chooses, then a company's step is nothing other than the decision of where in its own operations to start using AI.

Here is something I want to pause on for a moment. The superintelligence AI 2040 depicts has not yet arrived. That is exactly why we can reframe now as time for preparation. Just as Plan A tries to firm up the footing of safety with its 2035 pause, the next few years are, for a company, the period to firm up the footing of how it uses AI before AI becomes smart in earnest. The difference is stark between scrambling to think about how to use AI only after it reaches expert level, and having built up trial and error inside the company beforehand to grasp the knack of it. Prepare while there is still time to prepare. This, I think, is the most practical lesson executives can take from this document.

So, concretely, where should you begin? Here is how I see it. First, gauge AI's true ability, neither over- nor under-estimating it. Neither fearing the extreme scenarios nor, conversely, dismissing AI as "not usable yet," but confirming with your own eyes what today's AI can and cannot do. Without this reckoning, you tend to fall one way or the other: swept along by the scenario's flashy figures, or so wary that you cannot move at all. Next, separate out which of your operations AI works well in and which should stay with people. Neither trying to hand everything to AI nor having people shoulder everything, but judging the fit operation by operation and allocating accordingly. This is, in fact, the same work as the five plans choosing where to stand on the scale of speed and safety. The more a company can do these two things, the more it can turn the advance of AI from a threat into a tailwind. Conversely, drifting with the trend and simply adopting AI tends to leave the investment spinning its wheels.

If you are unsure how to organize that starting point, where AI should be put to work in your own company, talk to the WARP team. Specialists who led DX and data strategy at major companies work alongside you month by month, helping you land AI in management. That said, dropping in AI does not solve everything. Let us start from working out where it works and where human judgment is needed, and go from there together.

Summary

This ran long, so let me organize the key points.

• AI 2040 is a new future scenario the same AI Futures Project behind the talked-about "AI 2027" released in 2026, formally titled "AI 2040 Plan A." It has six authors and is the team's major release following AI 2027
• Where the first work was a forecast of "how things might turn out," the new one is a proposal for "what should be done." Some commentators frame it as "a governance plan rather than a forecast," and it takes the form of five plans presented to the candidates in the 2028 US presidential election, with 2029 as the fork in the road
• The heart of the first-choice Plan A (verified slowdown) is deliberately delaying, until 2040, a superintelligence that could otherwise arrive around 2030. It is drawn as a timeline: the 2029 US-China agreement, avoiding the runaway in 2030, scaling within human range in the early 2030s, the 2035 pause, and resumption in 2040
• The plan's core is a mechanism of verifying without trusting the other side. It is supported by three layers, controlling chip supply, tracking existing chips (about 98.5 percent), and aggregating them into monitored data centers, plus the ideas of mutual assured compute destruction and a moving capability ceiling. On the economy, it also depicts a citizen's dividend that could reach roughly 25,000 dollars per person in 2033 and roughly 1.6 million dollars in 2035 (after adjusting for extreme deflation)
• The document contrasts five plans, from Plan A to Plan S. Around the axis of the verified slowdown (A) sit Fight China (B), which contains China, Burn the Lead (C), which turns a lead toward safety, Race to ASI (D), the fastest race for superintelligence, and Shut It All Down (S), a full halt to development
• How far to believe it calls for caution. The authors themselves modestly estimate the success probability at around 42 percent, and hold reservations on the chips that slip past tracking, the premise of alignment, and the durability of US-China cooperation. From outside come criticism calling the economic depiction "science fantasy," and criticism that it is politically unrealistic and that anything beyond 2028 is guesswork
• Rather than swallowing the years, executives can prepare realistically by beginning, on the premise that AI could grow rapidly smarter, to avoid swinging to the extremes and to gauge where AI works in their own company

A future scenario is not a tool for competing over hits and misses; it is a tool for assuming the range of what could happen, lining the options up, and thinking through the moves. Read it not only as a story about nations but by drawing it close as a story about your own company. Then begin, wherever you can, to get ready to turn AI into a weapon of management.

References and primary sources