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
I think the chips alone are 10B minimum. It'd be way bigger than CERN.
Provided that the systems work, they can at least be repurposed to other things. If the organizations that are to train public LLMs can't do it, we can rent the system out to Mistral or something.
So I think something like 5B, starting with 10B to get started, in public money per year, the chip firms are private, some of the LLM firms will be private, but the system is available to train European LLMs-- that's I think a realistic approach.
The biggest problem with public money funded endeavours is corruption. At the end of the day most of the people would be there for the pay check and even if nothing is delivered after burning billions of euros, nobody would be held accountable.
That’s why I think it’s not going to be successful with public funding. What EU needs is structural unpopular reforms. Reduce taxes so companies and top talent would have an incentive to choose EU over US. Reform employment laws so people can be fired for being pencil pushers and lacking off. Relax privacy and copyright laws so the data can be used to train the model. Completely repel all laws that create a bureaucratic nightmare for startups. Today all these suggestions would be a complete no go in any European country, but that’s the hard reform EU needs like yesterday even to have a chance to compete against US and China.
Yes, of course people would be working there for the paycheck.
I think there's nothing special about public funding though. The field is so competitive that people will be mad if a competitive model is not achieved, making corruption more damaging to the organizations. There would also be some internal competition. There are after all several EU LLM/AI/etc. firms that would probably try to use this infrastructure.
By paycheck I meant that they get the money through corruption. Like fake bills, bloated estimates and huge commissions, but in the surface it all looks legal and they get away stealing millions of euros of public money of hardworking tax payers.
I guarantee you that maybe people would get mad that no competitive model is achieved because the politicians burned the tax payers money. But there would be no repercussions, nobody would be held accountable and would just be marked as a failed project.
I think the bigger issue of public funding and corruption, is that usually these contracts are awarded to "friends" with the purpose of 100% mismanaging and stealing as much of the funds.
Yes I agree, I'd only wish there's some way for do this by the book, like its not so hard, Brasil government has produced almost SOTA models, why cant the UE figure shit out? They could even rent hw like everyone else to do the first trainings...
But no, lets make this a 10B deal who someone will suck out 1B out of it even before its shipped
I think the way to do it is this: let EU chip design firms bid. They say what their systems can do, they give their prices, and then we choose the one that can achieve the requirements (pretrain multiple 10T+ models in a reasonable amount of time and then do RL on them) at the lowest cost.
There has to be a deadline and also a penalty if they don’t deliver. That way the company cannot just take the money and pocket it and just say oops, we failed
this feels like a small question, the money needed for this isn't that much (probably ~20b) and every country needs cyberwarfare abilities, especially automatable ones, so the military can pay for it. And its not like 10T models cant be used for anything else, so people who want ROI in health, chip development, all sorts of stuff
That’s assuming everything would be in time and no budget overflows which is never the case in public funded projects.
I can give you one concrete example. Google for Stuttgart 21. It’s a project to build a main train station in Stuttgart, Germany. The project has been delaying for years. By the time it’s finished it would be 12.5 years overdue its original deadline and so far it has costed 14.5 Billion euros.
One could assume that the biggest economy in Europe can build a train station in no time because how hard could it be? Given the fact that they can’t even do that, I can imagine how it would go with publicly funded cutting edge technological AI endeavour. Even if EU has billions of euros to burn, what they don’t have is time.
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.
Ah, yes. That's good too. Especially VSORA and the interconnect, and I guess the CPU is needed too. I guess what remains is the training chip and OpenChip will be the first.
Yeah, but if its final performance comes from being trained with data from a bigger model one can question whether it's a way to build genuinely new 40B models.
pretraining becomes more expensive actually as you make MoE models sparser (you need more tokens in the pretrain, and if you don't have that then you need to train for longer)
If we end up with a world where only US firms can use the latest LLMs and the latest LLMs are needed to keep up in the software world, or in making prototypes, then that's a whole series of fields which are blocked from us.
So I think we need to make sure that we no only can, but build frontier LLMs from scratch-- not RLed on foreign data, but genuinely from scratch.
Even from an economic point of view, I don't think a continuous outflow of ~200 USD/month for every office job is sustainable, and that's what we'd get if the plausible scenario is borne out. An inter-EU cost of 200 USD/month for every office job though, that's survivable.
A lot actually. Obviously not to write these comments, but a whole lot of Claude.
With regard to the substance: sales of ASML machines are not currently connected to the EU getting chips, but to ASML getting paid for their work. For EU chips for training transformer models we'll need chip design firms, not chipmaking, and as I stated in my comment, there are some promising ones that will probably be able to design the chips we need if we order them.
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.
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.
My understanding is that you can turn gas plants on and off very quickly, and they can run at any time of day. That makes them excellent suppliers for peak demand. For much of the year, the evening peak occurs when the sun is not shining and the winds have died down, so renewables are not viable.
Renewables can be combined with storage technologies such as batteries, so the syllogism "no generation during Dunkelflaute, but need reliablye generation -> renewables not viable" is too simple. My question is about whether renewables with storage technologies are a viable replacement for the gas power plants.
There are many facets of “viability”: technical, economic, political. There is obviously some reason why new gas plants are being built instead of renewables with storage.
Yes, and I'm still interested in better understanding that reason. My previous response just noted that the explanation you gave seemed overly simplistic to me in that it did not address that renewable energy can be stored, making it less volatile.
I live near one of the world’s largest grid scale battery facilities. It was in the news last year for catching fire and filling the air with lithium. [0] So I’m aware of the technology’s existence. But I also know that it’s still a very young technology, with battery chemistry constantly improving. There are also inherent losses going to and from AC power, which is what the grid operates on and which most plants produce.
I think you overestimate the currently available storage technologies. Batteries are very limited in capacity and super expensive in large scale. Hydrogen or synthetic methane might work, but those are gases, if you were opposed to them.
> My question is about whether renewables with storage technologies are a viable replacement for the gas power plants.
I'm guessing probably not. If it was truly the no-brainer "cheap AND fast AND better" option everyone thinks it is, data centers would be rolling out renewables right now. They're not so dumb that they'd pass up on a superior, cheaper solution just because.
Actually I’m not so sure about that. The turning on and off quickly part.
Because in my country (in Asia) there’s surplus energy from renewables (solar, wind) during the daytime, so much so that they’re having to shut things down because the grid can’t handle it (yet).
However at night we need power from traditional sources (coal, basically. We do have hydroelectric and nuclear power also, but that’s not sufficient).
They cannot shut down the coal plants during the day because apparently turning it back on is a tricky time consuming process.
They cannot run the coal plants at reduced capacity because that reduces the efficiency and increases the wear and tear of the equipment.
And the renewables also brought up another problem in that the power that comes is not steady (unlike the traditional sources).
All that is putting pressure on the grid.
The long term solution is to upgrade the grid to handle all this. And the government is working on it (I think). But that’ll take time to do. At least a decade.
Uhh, I don’t know much more technical details about all this. Maybe more knowledgeable members can contribute.
Coal is different. Natural gas is commonly used for peaker plants that provide electricity only when needed. They’re a better fit for using along with other intermittent sources.
You can build gas power plants on site at your data center and don’t have to wait for the grid to be upgraded. In the long run gas will be replaced by cheaper renewables
> 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.
It won't. Once you actually visit Bruxelles or talk with ex-policymakers you realize they don't actually care to give up sovereign control to the EU.
It's still viewed as a prestige post, and European industries and states will continue to work with partners outside of Europe if it means surviving.
Additonally, capital consolidation within the EU also means subnational capital flight which means layoffs. For example, look at how East Germany shifted hard to AfD after Germany Inc left for Poland. Therefore, when push comes to shove, it becomes politically untenable.
Okay. Lemme ask you a simple question: why does Dario Amodei live in the United States? Why do his parents live here? What happened?
It's not the first time in recent history frontier technology has been developed and commercialized, before any of this "capital" you are talking about really existed. That's the easy part (print money). The hard part is, taking leadership to convince the citizens of a place to be... what? Do you know?
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.
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.
The export control directives on Anthropic are under the authority of the Arms Export Control Act. The International Traffic in Arms Regulations (ITAR) implement the AECA for commercial exports, including the "technical data" under control. This story is about circumventing ITAR directives.
Who needs allies when you can have a digital god to subjugate those allies into proper vassals?
I know this is an insane statement, but I think much of current American foreign policy can be explained by the belief that their technology makes allies unnecessary.
Ask yourself the same question, but replace it with France kicking out NATO or New Zealand kicking out the Americans. What did they gain by pissing off their allies, and what have been the honest long term effects for the country in question?
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
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.
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.
Switzerland could probably extend their hydro if they wanted to, but it's NIMBY rich. Wind and solar has stalled, though there's plenty of both in the Alps.
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.
> Capital markets in the EU are simply unable to operate at that speed and scale.
Yeah that's true, getting investment of few million EUR to your company takes years and investors are looking at you like they are giving you their whole national budget. It is comical to the point that businesses rarely seek investments in general in Europe and you just need to grow naturally.
The AI Act Regulation (EU) 2024/1689 was introduced as a proposal on January 6th, 2021, so it took a bit over three years for that specific piece of legislation until it had the final vote in May 2024.
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.
There's a fundamentally different definition of how laws are supposed to work. EU law isn't a list of checkboxes that you can technically check while going counter to the spirit, it is a philosophical direction, the details of following it are up to you. The spirit matters, not the letter.
> When interpreting EU law, the CJEU pays particular attention to the aim and purpose of EU law (teleological interpretation), rather than focusing exclusively on the wording of the provisions (linguistic interpretation). This is explained by numerous factors, in particular the open-ended and policy-oriented rules of the EU Treaties, as well as by EU legal multilingualism. Under the latter principle, all EU law is equally authentic in all language versions. Hence, the Court cannot rely on the wording of a single version, as a national court can, in order to give an interpretation of the legal provision under consideration. Therefore, in order to decode the meaning of a legal rule, the Court analyses it especially in the light of its purpose (teleological interpretation) as well as its context (systemic interpretation).
If I was a business owner I’d rather operate under laws that don’t have highly ambiguous definitions of terms that introduces extra risk that is unnecessary in other places.
There's really only a risk for those trying to walk a line so they can actually get around the purpose of those laws while seemingly still being in compliance. For those with no round about intent, the "ambiguous definitions" are perfectly fine.
Right! Thanks for the link, I remembered reading that quote but couldn't find it. European regulators don't need hyper-specific definitions, because to them it's entirely normal to tell a company that they must do X or can't do Y even though the rules as written seem to authorize their current course of action Z.
All regulatory systems have some informal edge cases, of course. But Americans expect law to in general work more like a list of checkboxes and rely less on divining the regulator's intent. Indeed, that's one of the reasons why the regulatory environment under Trump is so frustrating to many of us; in the American view, there's supposed to be a strict distinction between what the law is and what the people at suchandsuch agency think the law is supposed to be or meant to achieve.
> But Americans expect law to in general work more like a list of checkboxes
To me as a European, this is a very low-trust view of lawmaking that assumes a hostile relationship between a government and its people.
The European approach is a bit more of a living conversation.
In the implementation period there's workshops where you figure out how to best comply in a way that makes sense for your business. There's a lot of flexibility there since you're just aiming for the spirit of the law, not some formal definition that might not make sense in your case.
If you're found out of compliance theres a bit of a back and forth and if you put in a good-faith effort to fix things, nobody has any issues.
The advantage of this approach is that the government doesn't tell you how to run your business and things stay agile as new use cases and business models come up.
It works out pretty well in general, and allows for a more cooperative approach to reaching policy goals.
Problems usually only arise when American companies try their bad-faith technicalities and find that doesn't fly here, like when Facebook changed their ToS to try to argue that using their services itself constitutes consent under the GDPR and predictably got dinged for it.
I don't want to sound like I'm criticizing any of that. It sounds like a reasonable approach and if European citizens like it more power to you.
To me as an American, this is a very high-touch view of lawmaking that sounds like a big problem for companies trying to do new stuff or challenge incumbents. If the meaning of the law is adjusted to fit each individual business case, doesn't that mean the regulator might not let me have all the same adjustments my competitors got? I wouldn't call this a question of hostility as such; even a kind and friendly regulator might think that some of those adjustments depend on doing business as normal, and thus they don't apply to the new abnormal things I'm doing. (Of course, I'm making the stereotypically American assumption that running around disrupting normal business practices is a valuable thing to do.)
First of all, I wanted to thank you for the constructive and interesting exchange of views - I find that too often conversations on EU regulations on HN devolve into caricatures and trolling, its been nice to have a substantial conversation with you.
Ironically, I have the opposite read on the same situation: "Just do what the law wants from you" seems like a pretty straightforward thing that even the smallest startup can follow.
In the case of GDPR that would be roughly "don't store any personal data you don't strictly need for the feature, ask for consent, delete data when asked".
Checkbox-lawyering on the other hand requires just that, lawyers. Expensive ones. Ideally entire legal departments.
If you're a huge incumbent, you can afford that. Meta finds the neatest little ways to technically comply without actually doing what the lawmakers wanted. The little startup? No chance.
The other concern is addressed by the judicial system. If a competitor got some exception you didn't get, that's your correction mechanism. The CJEU exists precisely to ensure consistent application across member states and cases. The purpose is the same for everyone; the implementation differs based on context, but the purpose doesn't change.
How can you say that when Europe has completely failed at producing any big, successful tech companies in the past couple decades? China and even India have a lot more staetups-turned-bigtech companies.
Past couple of decades, how many exactly? I mean Apple (48), Microsoft (51), Amazon (32), Nvidia (33), Oracle (49), Adobe (44), Cisco (42), Intel (58), Google (28) aren’t exactly young.
With the exception of Tesla (23) and Meta (22), USA is not brimming with large new tech companies from the past decades either.
I mean King, Spotify, and Klarna are not trillion dollar companies. But at least they are younger than Google.
American companies at around Spotify's market cap include Palantir (23), CrowdStrike (15), AppLovin (14), Uber (17), ServiceNow (22), Robinhood (13), AirBNB (18), Snowflake (13).
- 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
Confidential computing is not secure against a potential attacker who has physical access to the hardware. The CC security guarantees explicitly assume the attacker has no physical access.
> is not secure against a potential attacker who has physical access to the hardware.
Well, yes, its the oldest adage in computing that "physical access == game over".
So I would argue it is more about reducing your risk to a more acceptable level.
And in that respect I would say using services such as Tinfoil or Privatemode is an enormous step up from "trust me dude, we won't look at your data".
Remotely verifiable attestation combined with independent audits of the company hosting is a large step up from a Zero Data Retention clause in your contract that you have no way of verifying is actually happening other than "trust me dude".
Clearly I absolutely agree, having it on your own infrastructure is best for confidentiality. But even then, what about evil-maid attacks in the datacentre ? Unless you have your own datacentre, you're going to be in a shared colo facility ...
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.
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.
Which makes it even more difficult to limit access to Americans. Deepmind was originally a British company and has R & D in multiple countries to it will be even more awkward with them.
I have always thought the UK was too liberal in allowing foreign takeovers of key businesses. Arm and Deepmind are great examples of why.
Deepmind I can undertand, as it was early days and no guarantee of success? and they needed massive investment. But the decision to allow Arm to be sold was surprising.
But UK sells nearly everything - we need ongoing foreign investment.
>Its companies with no guarantee of success that are the future innovators. Look at how important Deepmind's field has become.
Most don't succeed, and DeepMind needed levels of money that would have taken the bulk of the Uk pot. Also no guarantees that a uk only based DeepMind would have succeeded in the same way as a Deepmind with multiple locations and able to use the Google Halo to attract staff.
Also the paper Attention is all you need was release by Google Brain researchers, not Deepmind.
UK backed Graphene financially stating it would take over the world, but it failed to succeed commercially.
Appears Switzerland as the lead guy there has lived in Switzerland for 10 years prior and probably refused to move (and he was previously at deep mind who also have researchers there so established talent pool)
DMA and DSA are very easy to comply with, there is an easy heuristic function to know if your AI startup is affected:
|_: Company| false
DMA and DSA only applies to companies explicitly designated as gatekeepers according to EU law. There are criteria for that, about scale and market control, but generally you really don't have to worry, you will see that coming from very far.
Not all of those regulations apply to AI training or inference. And having models coming from and provided by Mistral should be good enough of a signal that it is possible. Isn‘t this enough of an answer that it is possible to do it?
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.
I couldn't dislike the Trump administration more, but I think even the kindest and least vindictive of countries would not allow a company to relocate overseas to avoid export controls. (I had understood the proposal to be just that Anthropic should set up an EU office.)
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?
I don't think that this is a loophole. Anthropic would need to copy the model to the new company in Europe, which would be akin to exporting the model, which is exactly what export restrictions are restricting.
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)
So is recognizing the fact that one of the most luddite countries in Europe who can't ask ChatGPT for directions is incapable of thinking forward with AI, is "fascism" now? Cool
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.
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.
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
I think the chips alone are 10B minimum. It'd be way bigger than CERN.
Provided that the systems work, they can at least be repurposed to other things. If the organizations that are to train public LLMs can't do it, we can rent the system out to Mistral or something.
So I think something like 5B, starting with 10B to get started, in public money per year, the chip firms are private, some of the LLM firms will be private, but the system is available to train European LLMs-- that's I think a realistic approach.
The biggest problem with public money funded endeavours is corruption. At the end of the day most of the people would be there for the pay check and even if nothing is delivered after burning billions of euros, nobody would be held accountable.
That’s why I think it’s not going to be successful with public funding. What EU needs is structural unpopular reforms. Reduce taxes so companies and top talent would have an incentive to choose EU over US. Reform employment laws so people can be fired for being pencil pushers and lacking off. Relax privacy and copyright laws so the data can be used to train the model. Completely repel all laws that create a bureaucratic nightmare for startups. Today all these suggestions would be a complete no go in any European country, but that’s the hard reform EU needs like yesterday even to have a chance to compete against US and China.
Yes, of course people would be working there for the paycheck.
I think there's nothing special about public funding though. The field is so competitive that people will be mad if a competitive model is not achieved, making corruption more damaging to the organizations. There would also be some internal competition. There are after all several EU LLM/AI/etc. firms that would probably try to use this infrastructure.
By paycheck I meant that they get the money through corruption. Like fake bills, bloated estimates and huge commissions, but in the surface it all looks legal and they get away stealing millions of euros of public money of hardworking tax payers.
I guarantee you that maybe people would get mad that no competitive model is achieved because the politicians burned the tax payers money. But there would be no repercussions, nobody would be held accountable and would just be marked as a failed project.
I think the bigger issue of public funding and corruption, is that usually these contracts are awarded to "friends" with the purpose of 100% mismanaging and stealing as much of the funds.
Well, some risk must be taken, even the risk of such things.
Some sort of alternative must be created, after all, since models are being restricted to the US only.
Yes I agree, I'd only wish there's some way for do this by the book, like its not so hard, Brasil government has produced almost SOTA models, why cant the UE figure shit out? They could even rent hw like everyone else to do the first trainings...
But no, lets make this a 10B deal who someone will suck out 1B out of it even before its shipped
I think the way to do it is this: let EU chip design firms bid. They say what their systems can do, they give their prices, and then we choose the one that can achieve the requirements (pretrain multiple 10T+ models in a reasonable amount of time and then do RL on them) at the lowest cost.
There has to be a deadline and also a penalty if they don’t deliver. That way the company cannot just take the money and pocket it and just say oops, we failed
> The big question is who is going to fund it?
this feels like a small question, the money needed for this isn't that much (probably ~20b) and every country needs cyberwarfare abilities, especially automatable ones, so the military can pay for it. And its not like 10T models cant be used for anything else, so people who want ROI in health, chip development, all sorts of stuff
That’s assuming everything would be in time and no budget overflows which is never the case in public funded projects.
I can give you one concrete example. Google for Stuttgart 21. It’s a project to build a main train station in Stuttgart, Germany. The project has been delaying for years. By the time it’s finished it would be 12.5 years overdue its original deadline and so far it has costed 14.5 Billion euros.
One could assume that the biggest economy in Europe can build a train station in no time because how hard could it be? Given the fact that they can’t even do that, I can imagine how it would go with publicly funded cutting edge technological AI endeavour. Even if EU has billions of euros to burn, what they don’t have is time.
train stations are boring and no military spends time thinking about train stations. ASI hacker nonsense, that gets their attn.
Now you're just rambling
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.
Ah, yes. That's good too. Especially VSORA and the interconnect, and I guess the CPU is needed too. I guess what remains is the training chip and OpenChip will be the first.
GLM 5.2 is ~40B active parameters, which is what matters most for training cost.
Yeah, but if its final performance comes from being trained with data from a bigger model one can question whether it's a way to build genuinely new 40B models.
for RL cost*
pretraining becomes more expensive actually as you make MoE models sparser (you need more tokens in the pretrain, and if you don't have that then you need to train for longer)
Europe has asml
Yes, but what does that affect?
If we end up with a world where only US firms can use the latest LLMs and the latest LLMs are needed to keep up in the software world, or in making prototypes, then that's a whole series of fields which are blocked from us.
So I think we need to make sure that we no only can, but build frontier LLMs from scratch-- not RLed on foreign data, but genuinely from scratch.
Even from an economic point of view, I don't think a continuous outflow of ~200 USD/month for every office job is sustainable, and that's what we'd get if the plausible scenario is borne out. An inter-EU cost of 200 USD/month for every office job though, that's survivable.
asml is part of the chip supply chain; amsl has chip making talent, and EU needs more better chips to make more better ai
also an em-dash and genuinely in the same sentence? cmon bro how much claude have you been using?
A lot actually. Obviously not to write these comments, but a whole lot of Claude.
With regard to the substance: sales of ASML machines are not currently connected to the EU getting chips, but to ASML getting paid for their work. For EU chips for training transformer models we'll need chip design firms, not chipmaking, and as I stated in my comment, there are some promising ones that will probably be able to design the chips we need if we order them.
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.
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.
My understanding is that you can turn gas plants on and off very quickly, and they can run at any time of day. That makes them excellent suppliers for peak demand. For much of the year, the evening peak occurs when the sun is not shining and the winds have died down, so renewables are not viable.
Renewables can be combined with storage technologies such as batteries, so the syllogism "no generation during Dunkelflaute, but need reliablye generation -> renewables not viable" is too simple. My question is about whether renewables with storage technologies are a viable replacement for the gas power plants.
There are many facets of “viability”: technical, economic, political. There is obviously some reason why new gas plants are being built instead of renewables with storage.
Yes, and I'm still interested in better understanding that reason. My previous response just noted that the explanation you gave seemed overly simplistic to me in that it did not address that renewable energy can be stored, making it less volatile.
I live near one of the world’s largest grid scale battery facilities. It was in the news last year for catching fire and filling the air with lithium. [0] So I’m aware of the technology’s existence. But I also know that it’s still a very young technology, with battery chemistry constantly improving. There are also inherent losses going to and from AC power, which is what the grid operates on and which most plants produce.
[0] https://www.readymontereycounty.org/emergency/2025-moss-land...
I think you overestimate the currently available storage technologies. Batteries are very limited in capacity and super expensive in large scale. Hydrogen or synthetic methane might work, but those are gases, if you were opposed to them.
> My question is about whether renewables with storage technologies are a viable replacement for the gas power plants.
I'm guessing probably not. If it was truly the no-brainer "cheap AND fast AND better" option everyone thinks it is, data centers would be rolling out renewables right now. They're not so dumb that they'd pass up on a superior, cheaper solution just because.
Actually I’m not so sure about that. The turning on and off quickly part.
Because in my country (in Asia) there’s surplus energy from renewables (solar, wind) during the daytime, so much so that they’re having to shut things down because the grid can’t handle it (yet).
However at night we need power from traditional sources (coal, basically. We do have hydroelectric and nuclear power also, but that’s not sufficient).
They cannot shut down the coal plants during the day because apparently turning it back on is a tricky time consuming process.
They cannot run the coal plants at reduced capacity because that reduces the efficiency and increases the wear and tear of the equipment.
And the renewables also brought up another problem in that the power that comes is not steady (unlike the traditional sources).
All that is putting pressure on the grid.
The long term solution is to upgrade the grid to handle all this. And the government is working on it (I think). But that’ll take time to do. At least a decade.
Uhh, I don’t know much more technical details about all this. Maybe more knowledgeable members can contribute.
Coal is different. Natural gas is commonly used for peaker plants that provide electricity only when needed. They’re a better fit for using along with other intermittent sources.
Yes, coal takes very long to start up. Natural gas is very quick.
You can build gas power plants on site at your data center and don’t have to wait for the grid to be upgraded. In the long run gas will be replaced by cheaper renewables
As far as I understand, building a large solar farm is still considerably _faster_ than gas turbine plant.
It is hard to operate renewables 24/7. At night and winter, you need a big array of generators and LED lights to shine on solars. Very inefficient.
Or for every solar you need a generator as a backup!
I asked whether solar/wind in combination with storage technologies, including but not limited to batteries, was able to reliably power a data center.
Really, no! You were going to dump peak solar/wimd energy into grid, and play stupid why "backward" public utilities need gas to do stabilisation!
Batteries
Very fast and reliable? Sold!
What specific regulations are currently blocking AI entrepreneurs?
There may be an advantage to be able to use all available data.
Money, data, electricity, at least? In addition to the AI-specific laws.
> 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.
I really hope that will push for completion of the single market. I wouldn’t bet on it, but that’s something we should have done a decade ago
It won't. Once you actually visit Bruxelles or talk with ex-policymakers you realize they don't actually care to give up sovereign control to the EU.
It's still viewed as a prestige post, and European industries and states will continue to work with partners outside of Europe if it means surviving.
Additonally, capital consolidation within the EU also means subnational capital flight which means layoffs. For example, look at how East Germany shifted hard to AfD after Germany Inc left for Poland. Therefore, when push comes to shove, it becomes politically untenable.
Both Scaleway and OVH are French and partly made possible by France's inexpensive energy mix.
> Both Scaleway and OVH are French
Doh, I meant Hetzner
> partly made possible by France's inexpensive energy mix
Not really. Both developed well before the energy crisis when energy wasn't a significant input for DC construction.
The reason both succeeded is because French conglomerates continue to invest within France and only buy French.
It also didn't hurt that Xavier Niel and Klaba were able to leverage their preexisting telecom business and network.
Okay. Lemme ask you a simple question: why does Dario Amodei live in the United States? Why do his parents live here? What happened?
It's not the first time in recent history frontier technology has been developed and commercialized, before any of this "capital" you are talking about really existed. That's the easy part (print money). The hard part is, taking leadership to convince the citizens of a place to be... what? Do you know?
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.
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.
The export control directives on Anthropic are under the authority of the Arms Export Control Act. The International Traffic in Arms Regulations (ITAR) implement the AECA for commercial exports, including the "technical data" under control. This story is about circumventing ITAR directives.
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?
Who needs allies when you can have a digital god to subjugate those allies into proper vassals?
I know this is an insane statement, but I think much of current American foreign policy can be explained by the belief that their technology makes allies unnecessary.
Ask yourself the same question, but replace it with France kicking out NATO or New Zealand kicking out the Americans. What did they gain by pissing off their allies, and what have been the honest long term effects for the country in question?
Broadly negative, I'd assume.
Allies? What 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.
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
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.
Or start a war for developing own weights /s
I wouldn't put it past them, and I'm saying that without an /s
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.
Is it only Norway that has abundant, cheap and renewable hydro-power?
North of Sweden too, from the same mountains.
Switzerland could probably extend their hydro if they wanted to, but it's NIMBY rich. Wind and solar has stalled, though there's plenty of both in the Alps.
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.
> Capital markets in the EU are simply unable to operate at that speed and scale.
Yeah that's true, getting investment of few million EUR to your company takes years and investors are looking at you like they are giving you their whole national budget. It is comical to the point that businesses rarely seek investments in general in Europe and you just need to grow naturally.
> 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?
The AI Act Regulation (EU) 2024/1689 was introduced as a proposal on January 6th, 2021, so it took a bit over three years for that specific piece of legislation until it had the final vote in May 2024.
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.
There's a fundamentally different definition of how laws are supposed to work. EU law isn't a list of checkboxes that you can technically check while going counter to the spirit, it is a philosophical direction, the details of following it are up to you. The spirit matters, not the letter.
> When interpreting EU law, the CJEU pays particular attention to the aim and purpose of EU law (teleological interpretation), rather than focusing exclusively on the wording of the provisions (linguistic interpretation). This is explained by numerous factors, in particular the open-ended and policy-oriented rules of the EU Treaties, as well as by EU legal multilingualism. Under the latter principle, all EU law is equally authentic in all language versions. Hence, the Court cannot rely on the wording of a single version, as a national court can, in order to give an interpretation of the legal provision under consideration. Therefore, in order to decode the meaning of a legal rule, the Court analyses it especially in the light of its purpose (teleological interpretation) as well as its context (systemic interpretation).
https://www.europarl.europa.eu/RegData/etudes/BRIE/2017/5993...
If I was a business owner I’d rather operate under laws that don’t have highly ambiguous definitions of terms that introduces extra risk that is unnecessary in other places.
There's really only a risk for those trying to walk a line so they can actually get around the purpose of those laws while seemingly still being in compliance. For those with no round about intent, the "ambiguous definitions" are perfectly fine.
Unfortunately individual courts in some EU countries don’t care about the spirit but are fixated on the letter.
Right! Thanks for the link, I remembered reading that quote but couldn't find it. European regulators don't need hyper-specific definitions, because to them it's entirely normal to tell a company that they must do X or can't do Y even though the rules as written seem to authorize their current course of action Z.
All regulatory systems have some informal edge cases, of course. But Americans expect law to in general work more like a list of checkboxes and rely less on divining the regulator's intent. Indeed, that's one of the reasons why the regulatory environment under Trump is so frustrating to many of us; in the American view, there's supposed to be a strict distinction between what the law is and what the people at suchandsuch agency think the law is supposed to be or meant to achieve.
The EU has pretty good documentation for the various regulations. For example for GDPR they do provide checklists:
- https://www.edpb.europa.eu/sme/be-compliant/respect-individu...
- https://www.edpb.europa.eu/sme/be-compliant/secure-personal-...
And guidance: https://www.edpb.europa.eu/system/files/2026-04/edpb-summary...
> But Americans expect law to in general work more like a list of checkboxes
To me as a European, this is a very low-trust view of lawmaking that assumes a hostile relationship between a government and its people.
The European approach is a bit more of a living conversation.
In the implementation period there's workshops where you figure out how to best comply in a way that makes sense for your business. There's a lot of flexibility there since you're just aiming for the spirit of the law, not some formal definition that might not make sense in your case.
If you're found out of compliance theres a bit of a back and forth and if you put in a good-faith effort to fix things, nobody has any issues.
The advantage of this approach is that the government doesn't tell you how to run your business and things stay agile as new use cases and business models come up.
It works out pretty well in general, and allows for a more cooperative approach to reaching policy goals.
Problems usually only arise when American companies try their bad-faith technicalities and find that doesn't fly here, like when Facebook changed their ToS to try to argue that using their services itself constitutes consent under the GDPR and predictably got dinged for it.
I don't want to sound like I'm criticizing any of that. It sounds like a reasonable approach and if European citizens like it more power to you.
To me as an American, this is a very high-touch view of lawmaking that sounds like a big problem for companies trying to do new stuff or challenge incumbents. If the meaning of the law is adjusted to fit each individual business case, doesn't that mean the regulator might not let me have all the same adjustments my competitors got? I wouldn't call this a question of hostility as such; even a kind and friendly regulator might think that some of those adjustments depend on doing business as normal, and thus they don't apply to the new abnormal things I'm doing. (Of course, I'm making the stereotypically American assumption that running around disrupting normal business practices is a valuable thing to do.)
First of all, I wanted to thank you for the constructive and interesting exchange of views - I find that too often conversations on EU regulations on HN devolve into caricatures and trolling, its been nice to have a substantial conversation with you.
Ironically, I have the opposite read on the same situation: "Just do what the law wants from you" seems like a pretty straightforward thing that even the smallest startup can follow.
In the case of GDPR that would be roughly "don't store any personal data you don't strictly need for the feature, ask for consent, delete data when asked".
Checkbox-lawyering on the other hand requires just that, lawyers. Expensive ones. Ideally entire legal departments.
If you're a huge incumbent, you can afford that. Meta finds the neatest little ways to technically comply without actually doing what the lawmakers wanted. The little startup? No chance.
The other concern is addressed by the judicial system. If a competitor got some exception you didn't get, that's your correction mechanism. The CJEU exists precisely to ensure consistent application across member states and cases. The purpose is the same for everyone; the implementation differs based on context, but the purpose doesn't change.
>It works out pretty well in general
How can you say that when Europe has completely failed at producing any big, successful tech companies in the past couple decades? China and even India have a lot more staetups-turned-bigtech companies.
Past couple of decades, how many exactly? I mean Apple (48), Microsoft (51), Amazon (32), Nvidia (33), Oracle (49), Adobe (44), Cisco (42), Intel (58), Google (28) aren’t exactly young.
With the exception of Tesla (23) and Meta (22), USA is not brimming with large new tech companies from the past decades either.
I mean King, Spotify, and Klarna are not trillion dollar companies. But at least they are younger than Google.
American companies at around Spotify's market cap include Palantir (23), CrowdStrike (15), AppLovin (14), Uber (17), ServiceNow (22), Robinhood (13), AirBNB (18), Snowflake (13).
CrowdStrike (15)
EU equal of similar age and employment: SAP (31) or Darktrace (13)
AppLovin (14)
EU equal of similar age and employment: Delivery Hero (15), Infineon (27) or Evolution (20)
Uber (17)
EU equal of similar age and employment: Klarna (20) or Bolt (12)
ServiceNow (22)
EU equal of similar age and employment: Mistral (2)
Robinhood (13)
EU equal of similar age and employment: Revolut (10) or eToro (19)
AirBNB (18)
EU equal of similar age and employment: Booking.com (30), Amadeus (29)
Snowflake (13)
EU equal of similar age and employment: OVHcloud (26), Arm (29), or ASML (38)
Edit, okay those last ones are a bit old. But they are truly big and high technology.
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...
> I believe Nvidia chips have a secure way to run your model on other infra.
Yes. And its already on offer today.
See Tinfoil(US)[1] and Privatemode(Germany)[2]
Tinfoil have not been independently audited, it is somewhere on their long-term radar.
Privatemode have been thoroughly independently audited with documentation available on request.
[1]https://tinfoil.sh/ [2] https://www.privatemode.ai/
Confidential computing is not secure against a potential attacker who has physical access to the hardware. The CC security guarantees explicitly assume the attacker has no physical access.
> is not secure against a potential attacker who has physical access to the hardware.
Well, yes, its the oldest adage in computing that "physical access == game over".
So I would argue it is more about reducing your risk to a more acceptable level.
And in that respect I would say using services such as Tinfoil or Privatemode is an enormous step up from "trust me dude, we won't look at your data".
Remotely verifiable attestation combined with independent audits of the company hosting is a large step up from a Zero Data Retention clause in your contract that you have no way of verifying is actually happening other than "trust me dude".
Clearly I absolutely agree, having it on your own infrastructure is best for confidentiality. But even then, what about evil-maid attacks in the datacentre ? Unless you have your own datacentre, you're going to be in a shared colo facility ...
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.
AWS already hosts a few Anthropic models in EU datacenters.
Reminds me of some of the scenarios from https://europe2031.ai.
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.
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
Zurich is also Engineering (any many staff in Zurich and London will be from other countries temporarily working there).
Which makes it even more difficult to limit access to Americans. Deepmind was originally a British company and has R & D in multiple countries to it will be even more awkward with them.
I have always thought the UK was too liberal in allowing foreign takeovers of key businesses. Arm and Deepmind are great examples of why.
Deepmind I can undertand, as it was early days and no guarantee of success? and they needed massive investment. But the decision to allow Arm to be sold was surprising.
But UK sells nearly everything - we need ongoing foreign investment.
> Deepmind I can undertand, as it was early days and no guarantee of success?
Its companies with no guarantee of success that are the future innovators. Look at how important Deepmind's field has become.
> But UK sells nearly everything - we need ongoing foreign investment.
and then we complain about the resulting trade deficit! You cannot have both - it literally does not add up.
>Its companies with no guarantee of success that are the future innovators. Look at how important Deepmind's field has become.
Most don't succeed, and DeepMind needed levels of money that would have taken the bulk of the Uk pot. Also no guarantees that a uk only based DeepMind would have succeeded in the same way as a Deepmind with multiple locations and able to use the Google Halo to attract staff.
Also the paper Attention is all you need was release by Google Brain researchers, not Deepmind.
UK backed Graphene financially stating it would take over the world, but it failed to succeed commercially.
Do they put offices in only non-EU European countries purpose?
Appears Switzerland as the lead guy there has lived in Switzerland for 10 years prior and probably refused to move (and he was previously at deep mind who also have researchers there so established talent pool)
https://www.greaterzuricharea.com/en/success-stories/anthrop...
London as the UK rolled out the carpet inviting them, plus deepmind created a talent pool and centre
Curious how you can develop ai in Europe and still be compliant with eu ai regulations , dma, dsa and gdpr at the same time
DMA and DSA are very easy to comply with, there is an easy heuristic function to know if your AI startup is affected:
DMA and DSA only applies to companies explicitly designated as gatekeepers according to EU law. There are criteria for that, about scale and market control, but generally you really don't have to worry, you will see that coming from very far.Not all of those regulations apply to AI training or inference. And having models coming from and provided by Mistral should be good enough of a signal that it is possible. Isn‘t this enough of an answer that it is possible to do it?
I support
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.
Indeed, not for the foreseeable future.
Today and in this context, I consider "foreseeable future" to be "up to 2 years".
I couldn't dislike the Trump administration more, but I think even the kindest and least vindictive of countries would not allow a company to relocate overseas to avoid export controls. (I had understood the proposal to be just that Anthropic should set up an EU office.)
> Anthropic should set up an EU office
I read that from it too, but not sure how it would help as that doesn't stop export controls.
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.
But anthropic is. Hosting the model outside the US doesn’t do anything.
I think the idea is to host the whole company, physically.
I don't think that this is a loophole. Anthropic would need to copy the model to the new company in Europe, which would be akin to exporting the model, which is exactly what export restrictions are restricting.
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 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 don't think the Austrians were smart enough of think that for themselves
Leave your fascist remarks out of here
So is recognizing the fact that one of the most luddite countries in Europe who can't ask ChatGPT for directions is incapable of thinking forward with AI, is "fascism" now? Cool
Ignorant, yes, but not fascist.
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.
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.