back
139 comments
I think what's actually needed is two things: an EU training infrastructure that allows training of 10T+ models, and an EU inference infrastructure that is sufficient that it's possible to do RL on them.

This effectively reduces the problem to a specialized supercomputing infrastructure problem which I think is relatively easy to solve. I think the chips are coming. I think Euclyd will be able to do the inference chip and I think the training chip won't be harder. It's just a matter of accepting the need to order a huge number of them, being willing to think a little bit like the kind of people who operate corners. So we can be there next year, I think. What we then lack is a training chip-- maybe OpenChip can do it, maybe they can't, but there are reasonable but still unfinished projects. Maybe if Euclyd finishes an inference chip in 2027 we can have the state pay them to make a training version, put in fp32, put in communication tiles. If their design is real and works (which it should, since it's basically a fancier version of Groq, as it's described, and since even Groq works) I think the advantage these chips is likely to have would be enough that a training version would be NVIDIA-beating.

We probably need some solution for the data-- i.e. to allow people to do things that are against copyright law in a limited way, but I think it's a better idea to start EU firms than to try to attract Anthropic.

Because of the need for capital the hardware-software carousel is necessary. We can't pay for NVIDIA chips and then have NVIDIA feed that money into US firms. We have to feed money into EU chips that either carousel the money into EU AI firms or who just offer cheap chips.

The big question is who is going to fund it? Is it tax payers money? If so how can they guarantee it’s not going to be another waste and corrupted disaster. Also EU is already late to the AI race and by the time the lawmakers starts to think about this, it would be game over
VSORA already finished their inference chip if you want to take a look at what they're doing. They got a handful produced at TSMC 5nm node. They are expected to be offered by Scaleway at some point in the near future (alongside with other French components from the AION consortium, like a SiPearl Rhea1 CPU and an Eviden Bull BXIv3 interconnect, all of them are also already produced, only integration is remaining). It will still probably take some more time though, as the AI stack needs to be finalized (ZML), but I am surprised that those developments are not being much discussed.
GLM 5.2 is ~40B active parameters, which is what matters most for training cost.
Europe has asml
To stay near the frontier of AI without being subject to the discretion of foreign countries the EU has to stay near the frontier of R&D themselves. Even if they can get around ITAR now and self-host, they would be stuck with having to repeatedly negotiate permission to use each new advance.

If they do relax regulation (especially on energy generation) sufficiently to unleash the continent's big brained boffins and entrepreneurs on AI, they could quickly develop their own advances that would give them real leverage.

> If they do relax regulation (especially on energy generation) sufficiently to unleash the continent's big brained boffins and entrepreneurs on AI

Capital. Capital, capital, capital.

The EU is still not a single unified economy and capital markets remains semi-sovereign.

Every Euro that goes out of (eg.) a Dutch taxpayer's pocket into an (eg.) German domiciled competitor gets pushed back against by national competitors as well as by the government.

You see this with French and German rivalry against Scaleway+OVHCloud versus Hetzner (edited because of early morning brain snafus) to Dassault versus Airbus.

But the issue is, a single unified capital market that overrides national sovereignty also leaves vast swathes of European voters at risk of unemployment via capital flight. You saw this with East Germany's shift towards the AfD following industry's shift to Poland.

So neither industry nor national governments (who remain the overriding power of the EU) have an incentive for a single unified market, and actually remain incentivized to work with outside partners instead.

What is it that makes gas power plants so much more attractive than renewable energy? From what I heard, it's a bit easier to build them very fast and they reliably produce energy on demand (as long as gas is available of course). But I imaging one could replicate this using solar/wind and storage units.

I could imagine that the challenge is that that having enough solar panels for a few gigawatts of consumption is hard to do on-site, so one needs to connect the data center to the grid, which, in turn, complicates matters and transformators are scarce right now.

Is this about right? I do hope that we find a way to do this more sustainably. AI doesn't solve climate change in the next few years, so clean energy isn't irrelevant.

Or they can just wait for disease and famine, along with Israel and Russia, to destroy America from the inside. Or just get their fusion projects working.
What specific regulations are currently blocking AI entrepreneurs?
or, we could just wait a hot second, get GPU and associated hardware over the 30% utilization mark, develop a fault tolerance strategy that recovers more useful work, and spend a bit more time researching how models actually converge. 50% savings on training time would mean even more energy savings because of the add-on effects of cooling.

this spending of billions just to get a 4 month lead, without even trying to invest in getting this stuff to run properly is wasteful to the point of insanity. I don't think it's at all productive to chide people for not wanting to dump their resources into a black hole.

it seems pretty clear that the investors and the AI companies _like_ to throw around big GW numbers. it gives them a moat, and it fuels the bubble.

I don't think ITAR has anything to do with any of this.
Dario is an American patriot who wants the US to win. Don't see this happening.
What is the US "winning" by pissing off their allies?
They also need compute. The US has more compute than anywhere else, so Anthropic won't be moving until the EU can offer something equivalent.
US can simply ban export of AI tools and wieghts like they did with PGP. Austria should start using Mistral or open models.
Yes, but the highest levels of the US government… are not currently the best and the brightest, so they may well implement a ban after they've been exported or remove a ban in exchange for a shiny golden bauble.
Hosting current gen intelligence is like recruiting rocket scientist with no space program, university system, military purchasing or start up and commercialization ecosystem.

They will get capacity, and that is important, but it will be frozen in time. Without an independent ecosystem the engine will stall.

That being said all of this came out of a single community: YC. OpenAI and scale.com created a training data set collection, annotating and training fly wheel. Tiny little startups did this. But the EU is so afraid of failure that it’s hard to get anyone to try for moonshots

Is this the way forward for the EU? Genuinely curious: Are there any EU providers that host the very good recent Chinese open weight models? Like, an EU-based alternative to DeepSeek's own API offering? To me, that sounds like an easier business to turn profitable, albeit maybe less impressive.
nextbit in spain is hosting deepseek but its the most expensive provider on openrouter. inceptron in sweden for glm and nebius in the netherlands hosting smaller models like nemotron. mistral hosting their own models obviously. its not a big market.

the problem is big chunks of europe have expensive power and insane red tape for building data centers (or anything else really). i think france is the best option with cheap nuclear but its years before any large buildout can even start.

So how do these people propose to protect all involved after they defy CFIUS?
To be honest, as much as people complain about EU regulations and bureaucracy, at least they are highly predictable. Every relevant piece of regulation, like the GDPR and the AI Act, was probably more than five years in the making and then added another year or two to take effect.

If I were a frontier lab with a billion-dollar investment under my belt, I wouldn't want to operate in a regulatory environment with the same prediction horizon as the weather.

EU regulation is not the problem, it's the national governments. They are the ones who create red tape and high taxation almost every step of the way, and create years of delays when even one crazy person complains.
Anthropic have raised roughly $100 billion just in the first half of this year. Capital markets in the EU are simply unable to operate at that speed and scale.
> as much as people complain about EU regulations and bureaucracy, at least they are highly predictable.

Hasn't the EU basically already regulated any potential "unsafe" AI in Europe out of existence?

5 years in the making AI act?
They don't go on random personalist whims (so far!), but they also tend to be much less specific in a way that can frustrate US businesses. The GDPR definition of "personal data" is just a couple of lines long; the California definition of "personal information" lists out twelve categories, one of which is "sensitive personal information" with eight more categories.
Non-paywalled alternative: https://www.reuters.com/business/austria-lobbies-eu-host-ant...
Actually I am been asked to pay for Reuters

Edited : this looks like it is in paywalled https://uk.news.yahoo.com/austria-lobbies-eu-host-anthropic-...

Have OpenAI or Anthropic ever had a model hacked/leaked? Is there any good reads on their cultures of preventing it from happening?
- afaik no openai or anth model weight have been leaked to date

- confidential computing and trusted execution environments (TEEs) are the strongest primitive yet invented to prevent model weight exfiltration. there are other techniques. like bandwidth limiting of GPU workers and key sealing. but none as strong as those afforded by confidential computing and TEEs

- see https://confidential.ai/docs/confidential-computing-primer for a primer on confidential computing

- using confidential computing and TEEs to protect model weights in use is an area of active research

note: i work on confidential.ai

I believe Nvidia chips have a secure way to run your model on other infra.

https://www.nvidia.com/en-us/data-center/solutions/confident...

surely the weights for the model & the equipment to run them make it logistically challenging enough to deter that… also I’m sure models have leaked in their APIs before but those would be pretty easy and quick to catch/fix.
I don't think it would resolve anything. Mythos and similar models are under export protections. So even if you get hardware in EU, how are you going to get past the export protections?
Once you're no longer in the US, you're no longer bound by their laws.
AWS already hosts a few Anthropic models in EU datacenters.
Reminds me of some of the scenarios from https://europe2031.ai.
Anthropic already have offices across the world, including Europe, but unless they moved their registered address would be subject to the curbs.

If Anthropic quit the USA, Trump administration would likely make an example of them.

Wouldn't be pretty.

Anthropic engineering is in London and the US. The other offices in Europe are sales oriented
Here is a solution:

Trump forced all European countries to increase their defence budgets, And as AI is both a offensive and defensive tool, it can be argued that a chunk of this defence budget can be spent on AI R&D.

While I love your eye for malicious compliance I believe that Putin also forced European countries to increase their defence budgets.
I support
Curious how you can develop ai in Europe and still be compliant with eu ai regulations , dma, dsa and gdpr at the same time
Typical corruption in Austria, coming from the ÖVP.

Alexander Pröll is like Sebastian Kurz here. The ÖVP always wants to have financial interests leak into politics.

Interesting thought. But the Trump administration is absolutely vindictive enough that they’d put some kind of import restriction on Anthropic as punishment if they left the US and they won’t want to lose the US market, as much as the current situation works against them.
Good luck with that after they effectively "appropriated" the pirated IP to teach their models and admitted to it. They'd be drowning in lawsuits. At least what I think would happen, IANAL.
Interesting country lobbying this. So this is either the OPEC effect or some active measures "another country" to sow division (because I don't think the Austrians were smart enough of think that for themselves)
I was getting heat for proposing companies do this if they truly care about their mission.

Even if nothing comes of it, it’s a healthy consideration to anyone operating in the US to really think about their goals and what best sets them up for success.

Many other parts of the world do not operate under the same capitalistic mindset that American companies are forced into by pressure of the systems they are beholden to.