Suppose the US and China decide that AI development is moving too fast.

"Turn it off" sounds simple until the switch also kills coding agents, research tools, and business APIs.

Let the labs continue, and there is no slowdown.

AI 2040 proposes a narrow middle path: keep using the models we already have, but temporarily stop the giant training runs and experiments that produce stronger ones.

The proposal comes from the AI Futures Project, the group behind AI 2027. They call it Plan A.

The authors describe Plan A as a recommendation rather than their expected future. The 2040 date belongs to the scenario. The machinery around it is what deserves attention.

I am Alex, welcome to ShortCu8 by Innov8.

Lets Dive Deep 🐰

AI development has two separate layers:

Plan A puts the brake on the first layer while leaving the second one running.

That sounds easy. Proving that every country and AI lab is following the rule is the hard part.

Under Plan A, consumer chatbots, coding agents, business software, and most robot systems would run from inference datacentres.

Those datacentres could serve approved models. Frontier training and experiments would happen inside separate R&D datacentres with heavier monitoring.

Think of it this way: the finished medicine can remain in pharmacies while the laboratory producing the next formula comes under inspection.

AI 2040 argues that this inference-only system could preserve much of AI's economic use while governments negotiate what research should be allowed next.

The US cannot simply ask China to stop training models. China cannot trust an American AI company saying, "Nothing new is running in that cluster."

Plan A therefore starts with compute.

Large AI training runs need many advanced chips connected inside datacentres. Those chips are bought, shipped, installed, powered, and cooled.

The plan proposes:

The authors want roughly 99% of the world's frontier compute covered. That target comes from their modelling and has not been demonstrated.

Their verification supplement admits that the required system is not ready. If a deal had to happen today, the most reliable option might involve powering datacentres down.

Plan A adds another rule: publish most frontier AI research.

Code, experiments, algorithms, prompts, and research logs would become visible. Model weights, sensitive training data, and some technical details would stay protected.

A secret algorithm is valuable because it gives one lab a lead. If competitors receive the same discovery immediately, that lead shrinks. Independent researchers could also inspect safety claims instead of relying on the company being regulated.

Public research can help safety work and countries outside the current frontier. It can also help competitors or covert projects move faster. AI 2040 chooses transparency because its authors consider concentrated, poorly supervised development the larger danger.

Plan A calls one part of its system "mutually assured compute destruction."

Post-deal datacentres would be built so the US and China could disable or destroy the other's frontier compute if one side restarted the race.

The idea borrows from nuclear deterrence. Neither side gets to leave the deal with a giant untouched computing advantage.

This may be the least politically believable part of Plan A. It also gives cheating a physical consequence.

A 2026 study examined 20 hardware-level governance mechanisms. Cloud records, power monitoring, and physical inspection are available now. Treaty tools such as on-chip compute metering, cryptographic proof of training, and hardware enforcement remain in R&D or speculative.

Plan A also needs the US and China to expose sensitive infrastructure, allow inspections, share frontier research, and keep cooperating while both fear cheating.

It also assumes advanced chips remain a useful control point. Better algorithms, distributed training, model distillation, or smaller systems could weaken that control.

AI 2040 writes down the requirements for a believable slowdown. Government acceptance and treaty-scale verification remain unproven.

Now what do you think about this… reply

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Most platform evaluations start with a demo request and end three weeks later in a conference room. This one takes 15 minutes and puts you directly inside Gladly's interface — navigating it on your own terms.

See how AI surfaces real-time customer context before a conversation starts. Watch how a single conversation thread pulls in purchase history, channel history, and account details without a handoff.

No installation. No commitment. Start the interactive demo and see the platform for yourself.

See it yourself

#AD2

AI is in your development workflow. While the token spend shows it, the throughput doesn't. The human is very much still in the loop, and that's a context problem. Join live (FREE) on Jul 23 to see why teams get stuck and how leading teams use a context engine to fix it.

Register For Webinar