On July 9 2026 the AI Futures Project — the team behind AI 2027, the scenario that made “superintelligence by the end of the decade” a mainstream conversation and, per press reports, reached the US vice-president's reading list — published its sequel at ai-2040.com. Six authors, among them Daniel Kokotajlo, the former OpenAI governance researcher who quit over safety concerns, and project lead Thomas Larsen. The title: AI 2040: Plan A.

The framing does a lot of work: “It's called Plan A because it's a recommendation, not a prediction. It's what we think should happen, not what will happen, though we think it's plausible enough to aim for.” And it's called AI 2040 because that's when superintelligence arrives in it — which, in their own model, “would have happened much sooner (in 2030, to be precise) if not for decisive action on the part of the US and Chinese governments.”

The deal

The core is an international agreement anchored by Washington and Beijing, built on verification instead of trust. Frontier AI research becomes transparent to the deal's institutions; compute is the handle. Advanced chips are traced through customer records — in the plan's model that locates about 98.5% of them — and concentrated into mutually audited datacenters the authors call whitesites, where each side can see what the other is training. Capability ceilings rise in sync. Nobody has to believe anybody; they just check.

The schedule, in the scenario: a US AI Transparency Act in 2027, after Congress wakes up to how much power is concentrating in the labs. In 2028, automation hits white-collar work hard enough that the presidential election is about AI. In 2029, the US and China sign. Through the early 2030s the frontier keeps moving — slowly, inside monitored datacenters — and pauses in the mid-2030s at the level of top human experts. Years of alignment research at that level, then, in 2040, a deliberate, controlled push to superintelligence. It is not a degrowth plan: along the way it pays out a citizen's dividend (from 2033) and cures diseases. The bet is that a slow ascent captures the upside instead of gambling it.

Plans B through S

The document stress-tests Plan A against its rivals: Plan B (“Fight China” — race, but with safeguards), Plan C (“Burn the Lead” — sprint for unilateral American dominance), Plan D (the status-quo race), and Plan S (“Shut It All Down”). The authors' verdict is blunt: every path except A ends in either extinction risk or an irreversible concentration of power. A is not their prediction — it's the only door they see worth pushing on.

The gap, in reverse

I keep an almanac of AI and the rules; its first issue is a dashboard of the gap between what AI can do and when the rules catch up. Every entry on it has the same shape: capability ships first, governance arrives later, if at all. ChatGPT launched in November 2022; the first binding rules for general-purpose AI took effect in August 2025. Plan A is the first document of this weight drawn in the other direction: the deal comes before the intelligence explosion. Rules first, capability waits. Closing the gap by design instead of measuring it after the fact.

The honest counterweight: everything rests on a verified US–China agreement by 2029 — signed by the same two governments currently fighting over chip exports, to govern an industry the scenario itself sizes at $10 billion a month on frontier models by 2027, with datacenters under construction by 2028 costing about twice the US military budget. The authors know it — and wrote the plan anyway.

• Jul 9: the AI 2027 team publishes AI 2040: Plan A — “a recommendation, not a prediction.”
• Their default trajectory: fully automated AI research and superintelligence around 2030.
• The plan: a verified US–China deal — transparent frontier R&D, traced chips, mutually audited “whitesites,” capability ceilings that rise in sync.
• Pause at top-human-expert level in the mid-2030s; deliberate, controlled superintelligence in 2040.
• Their verdict on every rival plan — race, “burn the lead,” fight China, shut it all down: extinction risk or concentrated power.

• AI 2040: Plan A — the full scenario↗
• AI Futures Project — the announcement↗
• Axios — first look: a call to slow the race to superintelligence↗
• The Washington Post — “He warned AI could lead to extinction. Now he says there's a better path.”↗
• Scott Alexander — Introducing Plan A↗