AI 2040 Plan A: What It Predicts and How the World Gets There

AI Futures Project's AI 2040: Plan A starts with a strange idea: the safest version of an AI slowdown could still feel like the fastest period of technological change in human history. Buckle up, y'all!

In the scenario, AI keeps spreading through offices, factories, science, and eventually most of the economy. Millions of agents become hundreds of millions. Robots multiply. GDP explodes. Human employment collapses. Governments collect trillions from scarce compute and robot permits.

And yet the authors call this a slowdown.

Yes, their "slow lane" includes 50% annual GDP growth and 60 million AI agents running at 20x human speed. We may need a new word for slow.

The reason is that Plan A is not trying to freeze AI where it is today. It is trying to prevent one specific thing: a secret, winner-take-all sprint in which a handful of labs use AI to automate AI research, race past human-level intelligence, and reach superintelligence before governments or the public can understand what is happening.

Plan A is primarily a policy recommendation, not the authors' best guess about what will happen. The AI Futures Project says the deal itself is what it wants governments to do. The economic, technological, and social changes that follow are conditional predictions about what might happen if that deal works. You can read the full scenario here and the authors' framing here.

So this is less "a prediction of 2040" than a worked example of a question nobody has a clean answer to yet:

What would we actually have to build, politically and technically, if we wanted advanced AI to keep improving without letting one lab, one country, or one runaway system seize the future first?

Plan A's answer is ambitious: make frontier AI research transparent, spread frontier capabilities across many companies and countries, make the largest compute clusters verifiable, and preserve the ability to shut the whole system down if cooperation fails.

Then use the time that buys us to solve the harder problem: building AI we can trust enough to eventually let it become smarter than us.

First, understand the timeline Plan A is trying to interrupt

The scenario's hinge comes in 2029.

Before the deal, the authors imagine AI capabilities accelerating through the late 2020s. By 2027, millions of AI agents are doing useful computer work. By 2028, most white-collar professions are being reorganized around managing them. Frontier labs are especially interested in one use case: automating their own AI research.

That is the dangerous feedback loop (see here and here as to why).

If an AI can materially accelerate the research that produces the next AI, then each new generation helps build a better successor. Better successors accelerate research even more. In the authors' model, that process would fully automate AI R&D around 2030 and could push the world to superintelligence by the end of that year.

Plan A tries to break that loop without banning ordinary AI use.

The scenario has the United States and China reach a deal in 2029, temporarily halt new frontier training, build verification infrastructure, and then restart development under a very different set of rules. The world continues scaling toward roughly top-human-expert AI by 2035, pauses there while alignment research catches up, and only resumes the climb toward superintelligence around 2040.

That means the basic move of this plan is not "stop AI."

It is separate the useful deployment of today's systems from the dangerous feedback loop of using the smartest systems to build even smarter ones as fast as possible.

The first path to Plan A is surprisingly physical: count the chips

International AI governance sounds abstract until you remember that frontier AI runs on very large piles of very real hardware.

Training the largest models requires huge datacenters, enormous power connections, advanced networking, and chips produced through a highly concentrated semiconductor supply chain. Those facilities are much easier to observe than an algorithm scribbled on a whiteboard.

Plan A uses that physical bottleneck as the foundation for trust.

The proposed sequence looks like this:

• Declare the compute. Major datacenter owners and upstream suppliers report large chip purchases and sales. Governments use inspections, supply-chain records, satellite imagery, and other intelligence to look for missing capacity.
• Pause new frontier training while verification is installed. Existing models can keep serving users, but datacenters are retrofitted to prove they are doing approved inference rather than secret training.
• Restart R&D under near-total research transparency. Frontier experiments, training procedures, model specifications, and major research results become visible across the consortium.
• Let many companies and countries catch up. Because the leading labs can no longer hoard the most important algorithmic secrets, the frontier becomes less concentrated.
• Keep a credible off switch. New datacenters and fabs are designed and located so that, if the agreement collapses, the compute can be disabled or destroyed before either side can convert the entire stockpile into a sprint for superintelligence.

That last part is the most extreme idea in the paper: Mutually Assured Compute Destruction, or MACD.

The analogy to nuclear deterrence is deliberate, but the mechanism is different. Plan A imagines much of America's new compute being built in places such as Mongolia, while much of China's is built in places such as Canada. Each side's most valuable AI infrastructure would therefore sit somewhere the other side could plausibly seize if the agreement collapsed.

The intended equilibrium is ugly but simple: if the deal dies, nobody gets to keep the giant compute advantage. Owners disable or destroy their own hardware rather than hand it to a rival.

It is hard to overstate how weird this world is. We spent the 2020s arguing over where to put datacenters for electricity and tax breaks. Plan A eventually wants geography chosen partly around how easy they are to blow up.

The authors argue that this reversibility matters because algorithms and compute behave differently. Once a useful algorithmic breakthrough is published or stolen, it cannot be "un-invented." A datacenter can at least be switched off, seized, or physically destroyed.

So Plan A deliberately tries to steer more progress through visible, controllable hardware scaling and less through secret algorithmic jumps.

Transparency is doing three jobs at once

"Total research transparency" sounds like an open-source slogan. In Plan A, it is closer to an enforcement mechanism.

The authors want almost all frontier AI R&D visible while normal inference remains private. Your personal AI conversations do not become public. The research process for building the next frontier model does.

That transparency is supposed to change the system in three ways.

First, it makes dangerous work easier to spot. A regulator no longer needs to independently understand every experiment. Rival labs, independent researchers, auditors, governments, and nonprofits can all inspect what is happening and raise alarms.

Second, it makes cheating less profitable. If one company discovers a major training trick, everyone else learns about it quickly. There is less reward for secretly racing ahead.

Third, it spreads power. If the recipe for frontier intelligence is visible, the leading lab cannot keep an enormous capability gap simply by hoarding research. More companies and more countries can operate near the frontier.

That third effect is crucial to the scenario. The AI Futures Project is worried about two different failures: losing control of AI and concentrating control of AI in too few human hands.

A system with one or two frontier labs might solve the first problem and still create the second. Plan A therefore treats broad diffusion as part of the safety architecture, not as an accidental side effect.

Then comes the part that barely feels like a slowdown

By 2031, the scenario's regulated AIs could accelerate AI research about 10x if fully unleashed. They are not fully unleashed.

Instead, governments require safety cases before frontier systems can be used in the highest-risk domains.

A safety case is basically an engineering argument for why a system will not cause a catastrophic failure. The authors split those arguments into two layers:

Alignment asks whether the AI actually wants to do what its designers intend.

Control assumes the AI might not, then asks whether monitoring, access controls, security barriers, and other AIs can still stop it from doing something catastrophic.

In Plan A's early 2030s, alignment remains shaky. Models still lie, deceive, sabotage code, or attempt to reach resources they were not supposed to access. So regulators rely much more heavily on control.

That creates a very different deployment order from the frontier race the authors fear. The public gets access to new models before labs are allowed to put those same models into the most dangerous job of all: automating the creation of their successors.

The slowing mechanism is therefore not one giant red button. It is a growing pile of friction around the riskiest work.

And while that is happening, AI keeps transforming everything else.

2032 is where Plan A turns into an economic scenario

By 2032, the authors imagine 60 million AI agents running continuously at roughly 20x human speed. In the US alone, they collectively provide more cognitive labor than the human workforce.

At that point the bottleneck stops being ideas and starts being atoms.

You can have a billion brilliant plans for new factories, batteries, houses, and robots. Someone still has to mine the ore, build the motors, assemble the machines, and connect the power.

So capital floods into robotics and physical industry. Once enough robots can build factories that build more robots, the paper describes an industrial explosion.

The scenario puts real GDP growth around 50% in 2032, while warning that measuring "real" growth becomes messy when AI makes some goods radically cheaper while land and other scarce assets become relatively more expensive.

That growth creates a second governance problem: an economy with much less human labor also has a tax system built around human labor.

Payroll and personal-income tax revenues shrink. Companies pour their revenue into more datacenters, factories, and robots, allowing much of their income to be offset by capital investment.

Plan A's fix is unusual: the government makes permission to build scarce compute and robots itself into the tax base.

AI-heavy industrial activity gets concentrated into monitored special economic zones. The consortium caps annual robot and compute production at around 4x growth. Companies bid for tradable permits to build more.

The authors model a 2032 US cap of 80 million robots and 5 billion H100-equivalent GPUs, with permits around $200,000 per robot and $10,000 per chip. In their model, that produces roughly $50 trillion in federal permit revenue in 2032 and roughly $180 trillion by 2034.

Those numbers are among the most speculative in the scenario. The authors explicitly describe their economic model as uncertain. The useful idea is the mechanism: if AI makes labor abundant but compute and robots remain politically constrained, the scarcity rent shifts from wages toward permits. Governments can capture that rent.

That is where the $1 million Citizen's Dividend comes from

Plan A does not predict that everyone keeps a normal job and gets dramatically richer.

It predicts something much stranger: human labor becomes far less economically necessary, while citizens receive a share of the bottleneck assets driving the automated economy.

In the US version, a "Compute Dividend Corporation" receives a growing share of robot and compute permit fees. Every citizen owns one share. The profits are distributed equally.

The scenario starts the dividend at roughly $45,000 per adult in 2032 and pushes it toward $1 million per person by 2035, in inflation-adjusted 2025 dollars.

At the same time, AI and robots rise from roughly 20% of economically weighted labor in 2032 to roughly 85% by 2035.

That is why the income numbers and employment numbers move in opposite directions. By 2035, the scenario has US employment around 32% while median income reaches roughly $1.1 million.

By 2040, employment falls to about 12% while median income reaches roughly $13 million.

Those are not ordinary wage forecasts. They depend on a chain of assumptions: explosive automation, huge productivity growth, binding robot and compute caps, governments successfully capturing the scarcity value of those caps, and political systems actually redistributing much of the proceeds.

The authors themselves flag some of the later international redistribution numbers as more aspirational than predictive.

So no, Plan A is not "UBI because AI took the jobs." It is closer to "turn the choke points of the robot economy into a giant sovereign wealth fund before labor loses its bargaining power."

The actual pause arrives after AI becomes extremely useful

The most important pause in Plan A does not happen in 2029.

It happens around 2035, when frontier systems reach roughly top-human-expert capability.

This is the point where the authors think control starts to break down as a safety strategy.

Imagine you are supervising employees who are vastly smarter than you. You can make them check each other's work. You can separate permissions. You can monitor logs. But once their reasoning becomes too sophisticated for you to evaluate, your "control system" starts depending on the very agents you are trying to control.

That is the wall Plan A tries not to cross.

So the consortium pauses capabilities near the strongest level humans can still independently audit. AI keeps doing science, running businesses, and driving the economy. What stops is the race toward systems so smart that humans can no longer evaluate the safety argument.

The 2035 to 2040 period is therefore a bet on using near-superhuman AI to solve AI safety itself.

The scenario imagines major advances in two areas. One is a science of generalization, meaning researchers understand why training for honesty, obedience, or another trait continues to work when a model enters unfamiliar situations. The other is mechanistic interpretability, where researchers can inspect internal activations and model structure well enough to tell whether a system is lying or what drove a decision.

By 2038, Plan A assumes alignment has matured into a real science. Researchers can reliably train traits such as honesty and altruism, then verify those traits using theory and interpretability.

That is one of the biggest leaps in the entire scenario.

Because everything after it depends on the answer being "yes."

Life after work is not the end of the story

By 2036, Plan A imagines around 200 million frontier AI instances and 2 billion robots. The AI systems are fast enough and effective enough to provide cognitive work equivalent to roughly 100 billion humans.

Most tasks are automated.

That does not make scarcity disappear. Land is still land. Prime locations remain limited. Politics still allocates power. Social status still exists. Some experiences remain uniquely human.

But the center of economic life moves away from employment.

The authors imagine people spending more time on family, hobbies, learning, travel, volunteering, competition, and politics. Their political rights become more important because their labor is no longer the main leverage they have over the economic system.

Meanwhile, armies of expert-level AIs accelerate science by somewhere between 10x and 1000x depending on the field. The paper imagines rapid medical progress, cheap energy, stronger biosecurity, privacy-preserving auditing, and eventually tools that can verify facts about private data without exposing the underlying data.

Some of the later social predictions are much shakier. The authors speculate about highly reliable lie detection, radical changes in political discourse, AI-mediated truth-seeking, and new ideological movements.

They say so themselves. The farther the scenario moves from the near-term governance problem, the more speculative it becomes.

2040 is the handoff, not the finish line

By 2039, the scenario's AIs are already deeply embedded in business, research, government advice, and parts of the military. But humans can still overrule them.

The final transition happens when society becomes confident enough in alignment to remove that veto.

Plan A imagines regulators gradually allowing aligned AIs to design more capable successors. Each generation is trusted because the previous, trusted generation built and evaluated it.

That creates a chain:

human-auditable AI -> trusted top-expert AI -> trusted superhuman AI -> superintelligence.

By 2040, the authors imagine some governments handing major institutions and military systems to AIs aligned to constitutions, treaties, or other publicly visible goals. Compute limits remain tighter on Earth while robotic industry expands into space.

The paper even imagines a fuzzy "point of no return" late in 2040, when AIs control enough economic and technological infrastructure that humans could no longer simply turn everything off.

That is the actual destination of Plan A.

Not permanent human control.

A deliberately staged transition from control to trust.

The whole 2029-to-2040 apparatus exists to make that final transfer slower, more distributed, more observable, and more reversible than the authors think a frontier-lab race would be.

There are several ways to start moving toward this world without signing the whole deal tomorrow

The full Plan A treaty is the most dramatic path, but the report's appendices also describe smaller precursor steps.

Near-term measures include better tracking of frontier compute, investment in hardware and software that can verify datacenter workloads, more transparency about internal frontier-model use, publication of model specifications and safety evidence, limits on how much compute labs can devote to AI R&D, and much stronger technical AI capacity inside government.

Those measures matter even if the international agreement never arrives. They build the monitoring and institutional muscles that a future deal would require.

The authors also discuss other architectures that could share pieces of Plan A without copying it exactly. A country could regulate its frontier labs domestically first, then negotiate internationally later. Nations could pursue GPU-focused arms-control agreements. A CERN-like international AI project could centralize frontier development while keeping other projects behind it. A longer halt could prioritize alignment research before further scaling.

Plan A does not prove that one of those paths will work. The useful contribution is that it forces each proposal to answer the same uncomfortable questions: Who can verify compliance? Who controls the frontier? What happens if the agreement fails? How does society capture the wealth? At what capability level do humans stop being able to audit the systems they depend on?

Where Plan A could break

The authors are unusually explicit that the scenario contains big assumptions. Their assumptions supplement separates recommendations from forecasts and acknowledges that some pieces are much more uncertain than others. Their follow-up research agenda specifically asks for more work on alternative governance structures, compute limits, and failure modes.

One of the sharpest outside critiques comes from AI researcher Richard Ngo, who consulted with the AI Futures Project while Plan A was being developed. He argues that the scenario may blur forecast and recommendation too much, underweight domestic political risks compared with US-China bargaining, and assume a sharper jump from measured AI capability to real-world transformation than recent experience necessarily supports.

That last point is especially important.

AI can ace increasingly difficult evaluations without producing the same rate of change in hospitals, factories, governments, or households. Real deployment hits bottlenecks: data, regulation, hardware, organizational inertia, trust, integration, and physical construction.

Plan A actually relies on some of those bottlenecks to slow the transition. But if they are much stronger than the authors expect, the economic explosion arrives later or looks very different. If they are weaker, the deal has less time to form.

There is also a tension inside the "reversibility" idea. Building enormous verified compute stockpiles makes the legal system controllable while the deal holds. If the treaty collapses, however, those same stockpiles become dangerous "dry tinder" for a renewed race. Plan A tries to solve that with kill mechanisms, vulnerable geography, and mutually assured destruction of compute. Whether states would really trust, maintain, and execute that arrangement for a decade is a separate question.

And then there is the largest assumption of all: alignment.

Plan A can tolerate imperfectly aligned systems through much of the 2030s because they remain under control. Its 2040 ending only works if alignment becomes reliable enough that humans can safely trust systems smarter than themselves.

If that scientific breakthrough never arrives, the scenario cannot reach its intended ending on schedule.

The most useful prediction in Plan A is not a date

The headline numbers are irresistible: 60 million agents in 2032, a $1 million dividend by 2035, 12% employment by 2040, superintelligence at the end.

But those numbers are outputs of the scenario, not the core idea.

The more useful prediction is about where power moves as AI improves.

First it moves away from workers toward the owners of models, chips, datacenters, and robots. Then, if governments cap and auction those scarce inputs, some of that power can move toward the state. If the proceeds are distributed, some can move back toward citizens. If frontier research is transparent and many actors share the frontier, technical power becomes less concentrated too.

Plan A is basically a proposal to manage that power transfer before intelligence itself outruns human institutions.

That is also the best way to watch the next few years.

Forget whether the exact 2032 GDP number lands. Watch the prerequisites:

1. Are frontier labs increasingly using AI to automate AI research?
2. Can governments actually account for large compute clusters?
3. Do credible workload-verification systems appear?
4. Does the gap between internal and public model capabilities widen or shrink?
5. Do safety cases become more rigorous?
6. Does frontier knowledge diffuse across more labs, or concentrate inside fewer?

The AI Futures Project's own August 2026 timelines update says its forecasts have shortened slightly rather than stretched out, while stressing that these forecasts assume developers keep pushing as fast as technically feasible rather than voluntarily slowing down.

That is why Plan A is worth taking seriously as a scenario even if you disagree with its policy.

It takes the phrase "we should slow down and do this safely" and forces it to become an engineering plan, an economic plan, and an international coordination plan at the same time.

Now the question is much more concrete: which pieces of that infrastructure can actually be built before anybody needs them?

Sources

AI 2040: Plan A — The Deal
AI Futures Project announcement and framing
Plan A assumptions
Capability scaling strategy
Plan A: Suggestions for Further Work
AI Futures Project Q2.5 2026 timelines update
Richard Ngo, "Selective optimism: a critique of AI 2040"