AI Futures Project — the team behind the widely discussed AI 2027 — recently released its counter-image: Plan A. Where AI 2027 ended in either extinction or total power concentration, Plan A is a deliberately positive scenario. At its core is a brake. The United States and China agree in 2029 not to race toward superintelligence, all AI research is made public, dozens of companies in many countries are allowed to catch up to the frontier, and the world enters what the authors call mutually assured compute destruction. Development scales within human capability until 2035, pauses at top-expert level, and is only released toward superintelligence in 2040. Hence the title.
It is an unusually honest document. The authors say straight out that it is a recommendation dressed as a scenario, not a forecast, and they subject their own plan to precisely the scrutiny they demand of others. That is why it deserves to be taken seriously. And my objection is not that the plan is too optimistic about intelligence. It is that it slows the wrong curve.
The brake is not a forecast
It is easy to read 2030–2040 as a claim that intelligence develops slowly. It does not. The slow trajectory is a forced brake. Without the agreement, the authors already have fully automated AI research by 2030 and superintelligence by the end of the same year; one of them believes reality will move even faster than the scenario. On that point, they and the exponential curve we described in Every seventh month are broadly in agreement.
The load-bearing — and least examined — assumption sits somewhere else. Plan A takes for granted that the physical, industrial rollout will track the cognition curve closely: the robot fleet grows fourfold a year, two billion robots are in place by 2036, an “industrial explosion” in which robots build factories that build robots. That is where I hesitate. Cognition and atoms do not obey the same law.
The overhang
Cognition is cheap, copyable and has zero marginal cost. Atoms do not. So the two curves drift apart — and the asymmetry is domain-dependent. What is cognition-heavy and atom-light accelerates, perhaps faster than Plan A assumes: cyber, influence, finance, research, and above all AI development itself. What requires atoms lags: robotics, manufacturing, power grids, construction.
Our own coverage shows both halves in the same week. JadePuffer — the first fully autonomous ransomware chain, from intrusion to encryption without a human in the loop — is already here. At the same time, the leading humanoid companies still count their robots in the tens of thousands: Norway's 1X reports capacity for ten thousand hands per year, China's Zeroth thirty thousand orders. Getting from there to Plan A's two billion robots by 2036 requires the global fleet to grow a hundred thousand-fold in a decade. You can write that into a model. It is another thing entirely to cast it in concrete, copper and permits.
That is the industry-adjacent objection in a nutshell: anyone who has ever built a facility knows that lead times, supplier chains and permitting do not follow Moore's law. An army of genius designers does not solve the transformer shortage.
Physics constrains intelligence too
“Intelligence fast, atoms slow” is still too clean. The cognition curve is itself throttled by physics. The illustration ran in our own feed this week: SK Hynix's largest overseas U.S. listing ever, built on the fact that the company controls most of the world's HBM memory — the bottleneck the entire AI build-out is currently stuck in. And HBM is only the first link: power, cooling and grid connections come next in the same chain. Compute, energy and fabs tie the curves back together at the infrastructure level even when they decouple at the rollout level. You cannot think your way past a power deficit.
Regulatory friction — cuts both ways
Plan A also assumes a nearly frictionless institutional adaptation: new tax codes, markets for emission rights on robots and compute, a citizens' dividend, special economic zones — all built in two or three years and coordinated by a global consortium. Hold that up against reality. The EU AI Act took years to negotiate. Norway's AI law is still in preparation. The UN recently gathered 193 countries for its very first global AI dialogue. Regulation moves on a different time scale than the scenario's institutional sprint.
And the friction cuts both ways. If regulation is as slow as it usually is, then the agreement itself — the largest regulatory mega-project in the entire plan — becomes even less likely. You cannot invoke institutional inertia against the robot rollout while assuming it away for the deal.
Why the asymmetry changes the risk picture
Here is where the point bites. If cognition comes first and atoms last, Plan A's reassuring parts arrive slowly — the orderly, broad, distributed build-out that is supposed to spread power — while the frightening parts arrive early. Concentration of power, cyberattacks, large-scale influence operations, the design of new weapons: none of that requires robots. It only requires cheap, superhuman cognition. The overhang thus makes the risk picture worse at exactly the end where the scenario wants to reassure us. It flips the entire dramaturgy: the dangerous arrives before the reassuring, not after.
In our own scenarios
Translated into our own framework, Plan A is in practice a staged Sharing — a deliberate attempt to spread AI broadly — that assumes away Friction. But Friction, both physical and regulatory, is not a marginal disturbance. It is the variable that determines the outcome. A scenario that wants to slow the right curve has to begin there, not assume it away.
The Nordic chair
One last thing, for us. Plan A is a story about two powers. The United States and China negotiate; “the rest of the world” gets 45 percent of the robot and compute quotas as an afterthought, distributed among nuclear powers and countries with chip manufacturing. Sweden is not mentioned. Neither is Norway.
The question then becomes: where does a small, high-trust, EU-regulated, energy-rich European periphery sit inside a compute duopoly? Probably not at the negotiating table — but in the middle of the value chain. The data centres in Narvik, Luleå and Boden. The power. The HBM memory we own a slice of through the sovereign wealth fund. The kind of European physical AI Norway's 1X demonstrates can be built to world class. This is not the power to hand out the universe. But it is a position, and one best managed with open eyes and proportion — not with the belief that we either win the race or stand entirely outside it.
The temptation, again
Plan A's greatest merit is its stance: daring to write down a detailed plan and let it be scrutinised. Its greatest weakness is the assumption that atoms and institutions obey the same exponential as the bits. The temptation — to borrow from The forecaster's temptation — does not lie in guessing the wrong year. It lies in believing that the physical world and its rules scale like software. They do not. And it is in the gap between them, in the overhang, that the coming decade will actually be decided.