/Tech8h ago

Seth Lazar argues nations must build sovereign frontier AI models, comparing the scale to the $350 billion AUKUS defense deal

AUKUS costs $350 billion without granting nuclear sovereignty to Australia

312012412621.1K

#501

Original post

Seth Lazar@sethlazar#1093inTech

Really glad Anton wrote this. The complacent assumption that the prospect of an alternative to American AI hegemony is inconceivable has become very widespread. But it's hard to argue with his dual thesis here: that it seems necessary, as well as impossible.

But I think the scale of the impossibility needs to be placed more in context. The AUKUS deal, which doesn't even get nuclear sovereignty for Australia, is forecast to cost over $350bn. We *are* capable of making commitments on this scale. But I suppose the prospect that we will do so without clearer demonstration that ASI is on the horizon is dim. So a lot depends on working out when will be too late to start.

Something else I appreciate in Anton's writing here: he acknowledges a lot of uncertainty. People seem wildly overconfident in their views about geopolitical strategy, which strikes me as a crapshoot at best.

Anton Leicht@anton_d_leicht

http://writing.antonleicht.me/p/the-moonshot

7:30 AM · Jun 17, 2026 · 1.1K Views

Sentiment

Positive users endorse middle powers building independent frontier AI systems and data centers to gain sovereignty as an exhilarating practical step, while negative users dismiss the idea as futile and impossible due to prohibitive costs.

Pos

66.7%

Neg

33.3%

6 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Related links

The Moonshot

ANTONLEICHT.MEVia

Posts from X

Most Activity

VIEWS1.9KBOOKMARKS8LIKES17

Anton Leicht@anton_d_leicht

http://writing.antonleicht.me/p/the-moonshot

10h1.9K178

RETWEETS21

Anton Leicht@anton_d_leicht

AI sovereignty debates dodge their own conclusion. If you really think frontier AI is critical and the US can't be trusted, there is only one path: build it yourself, whatever it takes.

So let's drop the sovereignty theatre and look at how we'd pull off the middle power project.

10h19.1K177120

REPLIES1

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

«We live in the worst case scenario» is one of the hardest things for people to accept.

Anton Leicht@anton_d_leicht

AI sovereignty debates dodge their own conclusion. If you really think frontier AI is critical and the US can't be trusted, there is only one path: build it yourself, whatever it takes.

So let's drop the sovereignty theatre and look at how we'd pull off the middle power project.

1h1.2K152

Anton Leicht@anton_d_leicht

@sethlazar yes, I think your second paragraph is the problem - I could see us mobilising 500b for this in, say, two or three years. but it'll be much, much more expensive (perhaps impossible because of recursive effects and siloing in the labs pulling up the ladder) then.

8h1477

Kevin A. Bryan@Afinetheorem

@anton_d_leicht Great, including conclusion. Beyond cost, coordination is so hard here. Is Germany going to enforce privacy rules that make AI use impossible? Is Australia going to limit sharing with Canada because of high level China links in govt? Etc. Even this is no "sovereign guarantee".

7h2785

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

@XxXxALVESxXxX this isn't really telling us much

1h41

Sunless.jp@jpFromTlon

@anton_d_leicht Excellent. IMO, the no-brainer option is to start spamming data centers in Europe ASAP, as this is a critical part of any plan we decide on.

8h2861

Alves@XxXxALVESxXxX

@teortaxesTex Russia and Iran do not have frontier models and still are winning their own wars, although very slowly in the first case.

The other scenario seems more likely to be the one we currently live with, no?

1h5

Fritz Espenlaub eu/acc@Fritz_Espenlaub

@anton_d_leicht wenige Leute haben einen höheren Nailungs-Quotient als Anton Leicht

10h421

Arun Rao@sudoraohacker

Fun exercise but it shows to me how futile it is for any middle power to try building their own model from scratch-a $500bn investment for a quasi-private entity is a non starter.

A better plan: -Implement Draghi reforms to free up the economy and single market, and also to trim the welfare state and encourage more welfare flexibility (your economy will fail if you penalize work and innovation over state welfare payments) -Tie welfare and education to work, esp in high skilled jobs with labor deficits, and retraining to these jobs and to rebuild all workflows with AI - offer tax credits to private companies to run this (so the welfare payments become credits for companies to make jobs in deficit areas) -Just use open source Chinese and American models and focus on post-training and building agentic systems from scratch - this has some risk, but is much cheaper (5bn to have a few labs, vs 500bn for a single lab making a commodity product). -Rebuild education systems from scratch to focus on creating an AI savvy workforce, with internships and hands on experience at the center. The current K-16 Western schools based on the German gymnasium, Prussian factory, or sybarite undergraduate period are seriously broken - they function more as state-sponsored luxury goods as opposed to a real economic institution. (And I’m fine with higher education as a luxury good, and personally value it - we just shouldn’t kid ourselves about economic value or public funding for it).

7h1031

Oliver Klingefjord@klingefjord

@anton_d_leicht @sethlazar Norway's sovereign wealth fund is already somewhat AGI-pilled, investing heavily in AI-related companies. Could see it being leveraged

7h381

dwayne woods@twittwoods

@anton_d_leicht Because it’s impossible! Damn, I anticipated this discussion with my paper written two weeks ago

Sovereign Ai and its Paradoxes

https://youtu.be/MWhfi7D1cWI?si=vbhfL9alZIZ0_KIg

7h102

The Little Disturber@Victori45792596

@anton_d_leicht What about the risks ?

7h89

Joel Ormala@joel_ormala

@anton_d_leicht Great read. The Airbus anchor is relatable. The massive investments in datacenters would serve both a project like this and a leverage strategy towards US frontier developers. The bottleneck seems to be the pace of change and coordination it would require.

8h71

Jacob Schaal@FutureEconJacob

@anton_d_leicht Exhilarating read! Worth dreaming about to clarify thoughts!

5h4