# Wild theory: the AI “slowdown” is because they’re running out of GPUs, power and datacenters 😂, not fear of IA — retained record

> Retained reddit record from reddit:singularity.

Canonical: https://highsignal.app/data/records/3f38ce880cc18b15

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:singularity)  published 2026-09-12 17:40 ingested 2026-09-13 02:38

[original source ↗](https://www.reddit.com/r/singularity/comments/1weixv8/wild_theory_the_ai_slowdown_is_because_theyre/)

## retained text



```
Ok hear me out before murdering me

In the space of a few days we suddenly have Dario Amodei talking about slowing frontier AI, OpenAI talking about “pacing” model development, Sam apparently open to coordinating that with other labs, and now Musk backing the general idea too.

Very responsible. Very mature. Bla bla bla 

But I have a stupid alternative theory that I can't completely get out of my head:

**Maybe these guys are also starting to run out of planet. 😂**

Because while everyone is publicly discovering the brake pedal, look at what is happening on the infrastructure side.

Anthropic already says it uses **over 1 million AWS Trainium2 chips**, and this year signed an agreement with Amazon for **up to 5 GW of additional compute**, worth more than **$100 billion over 10 years**.

Then they signed for **multiple gigawatts of next-gen Google/Broadcom TPU capacity** starting in 2027.

And apparently that still wasn't enough, because Anthropic also grabbed the entire compute capacity of Colossus 1: **300+ MW and more than 220,000 Nvidia GPUs**.

Meanwhile OpenAI is the customer behind that ridiculous Ohio AI campus where Nvidia/SB Energy secured about **4.25 GW of initial IT load**, with the broader project potentially going much higher.

4.25 GIGAWATTS.

For reference, at some point you stop describing this as "a datacenter" and start describing it as "a small industrial civilization with a chatbot attached to it."

And Musk's own Colossus went from 100k to roughly **200k H100s**, with a stated roadmap towards **1 million GPUs**.

And the funniest detail is that OpenAI itself recently wrote that coding agents are making researchers run more experiments, that experiments per researcher hit an all-time high in August, and that **compute may become an increasingly important bottleneck as other bottlenecks disappear**.

That sentence made my  brain light up.

To be clear: **I'm NOT saying the safety concerns are fake. But all the incidents for sure where provoked by people, not AI per say, by sentencie.** 

AI capability is obviously moving insanely fast and there are real reasons to worry about autonomous agents, cybersecurity, recursive research acceleration, alignment etc.

I'm saying something slightly different:

Maybe the "responsible slowdown" narrative is arriving at an extremely convenient moment where the frontier labs are simultaneously discovering that scaling AI from here requires absurd quantities of GPUs, electricity, substations, cooling, networking and physical infrastructure that simply cannot be built as fast as the models are improving.

You can invent a better training method in six months.

You can't `pip install 5 gigawatts`.

So perhaps the real situation is:

**Safety concerns are  real, sure.**

**Desire to slow competitors down: probably real.**

**Physical compute bottleneck: also very fucking real.**

And meanwhile, behind the nice speeches about slowing down, I strongly suspect every one of them has engineers working 24/7 trying to make the next generation cheaper, more efficient and less compute-hungry.

Basically:

*"We have collectively decided to slow the AI race."*

*\[sound of 400,000 GPUs being installed in the background\]*

Poker face level: frontier AI CEO.
```



document fetched 2026-09-13 00:29 · canonical https://reddit.com/r/singularity/comments/1weixv8/wild_theory_the_ai_slowdown_is_because_theyre
