Hey Pieter! It's been a long time since we've last spoken (high-school CGU times...), but I recently stumbled upon your substack because of the interest/worry in rapid AI development that we apparently share.
I'd like to refer you to this article; https://europe2031.ai/ (there is a TLDR here: https://europe2031.ai/summary/) which was published this week. It relates to some of the points you list in the post, and I find it to be quite a convincing case for how the future might turn out if we don't take drastic action.
Ha Stan! Leuk om van je te horen. Ja, ik ben ook een groot fan van hun stuk (en heb op een paar plaatsen kunnen meehelpen). Cool om te horen dat jij je ook bezighoudt met AI vraagstukken. Ben je bekend met het Delta Instituut? Onno en Stan van Baarsen doen super goed werk voor Nederland daar.
Wat gaaf dat je daar aan hebt meegewerkt! En ja, die ken ik! Ga toevallig komende vrijdag op gesprek bij het Delta Instituut voor een mogelijk afstudeerproject :)
Europe should tell ASML to build kill switches into chips, so they can trip the kill switches if reckless American AI companies are about to destroy humanity.
While I understand your concern, I think what you're suggesting is not very realistic. We are lucky that, thanks to ASML, we are not completely excluded from the semiconductor supply chain, but we cannot (and should not) expect them to single-handedly regulate AI.
ASML is unique as of today, but in the model of the world you are describing (with superintelligent AI), I would not be surprised if other companies would slowly start to close that gap. It will be epecially hard to maintain that technological advantage if European companies (like ASML) get denied access to frontier AI models. The pullback of Mythos was the first real example of that.
I think we have the talent, capital and infrastructure to develop a wider portfolio of EU firms that can become essential to the global AI ecosystem. That would create a much broader co-dependent relationship with the US/China, which in turn puts us in a better position to have a say on the regulation of A(G)I.
>ASML is unique as of today, but in the model of the world you are describing (with superintelligent AI),
My view is that once superintelligent AI is achieved, it is likely to be unfriendly/misaligned with human values, and humans won't have much control anymore. It won't much matter which nation develops an unfriendly superintelligence. (Or a friendly superintelligence, for that matter, see https://www.astralcodexten.com/p/you-have-only-x-years-to-escape-permanent )
I'm focused on intervention points in the pre-superintelligence period. Presumably ASML will retain substantial competitive leverage in the pre-superintelligence period.
The message of the 2031 scenario was that Europe should think and act boldly. From my point of view, getting ASML to add kill switches is actually a much lighter lift than creating a hyperscaler/frontier model/etc.
Since the kill switches won't be activated at first, things will appear to be business as usual for ASML's customers. The Trump administration might not be super happy about kill switches, but public opinion in the US is skeptical of AI, and public opinion might matter more (side note: from my perspective as an American, Europeans have a recurring blind spot of focusing overmuch on government-to-government communications, and neglecting the role of US public opinion in transatlantic relations).
The kill switch thing will take time to design and set up, and it will only apply to newly manufactured chips. So it might be too late already, but I would go ahead and get started now anyways.
Simplistic analysis. Every point is valid, but it omits the most important one: if you don’t have vibrant internet companies, you don’t have the talent base needed to build internet giants. If you don’t have internet giants, you don’t have the talent base needed to build AI giants. So every time Europe—or anyone else, for that matter—delays the development of giants in the current leading technology, it also dramatically reduces its chances of playing a meaningful role in the next technological breakthrough.
Re. 2: The qualitative difference is that the Google ecosystem (and the internet and e-commerce more broadly) was an addition to the previously existing economy, whereas AI is partly a replacement. Europe’s lag in digitalisation was a missed opportunity; being left behind in AI, by contrast, means lasting economic decline.
Re. 4: Some of the other key technologies of the 2030s and 2040s will be space launch, semiconductors, quantum computing, humanoid robots, and BCIs—in all of which Europe is behind the U.S. and China (with the sole exception of EUV technology).
Re. 11: The graph shows that Mistral’s revenue was 2.9–4.4% of Anthropic’s at the same point in time (the beginning of this year). It also shows that Anthropic’s revenue was exactly 4.4% of OpenAI’s at the beginning of 2024. So should Anthropic have given up then?
Clarifying piece — the convergence history and the second-mover problem make the anti-sovereignty case on frontier models simply and crisply. Two things I'd press, and one omission.
The pushbacks first. Point 2 — that model dependence is qualitatively no different from the Maps/Gmail/Search dependence Europe already lives with — is where I diagree. Search and Drive were tools one reached for; frontier models increasingly shape the judgment itself, what gets framed, weighed and eventually believed. Chor Pharn calls this the gap between economic and epistemic dependency: importing productivity is one thing, importing cognition another. The same entry on the balance sheet, a very different effect inside the institutions that have to metabolise it.
Then the harder question the piece sets aside: not "should Europe build a frontier model" — you answer that really well, no — but whether Europe can absorb the ones it already has. Access is not adoption. Europe has had frontier access throughout and still trails US adoption by a mile, which suggests sovereignty is not the binding constraint. #19 could become the whole essay.
The omission is China. You frame the choices as build-your-own-frontier or integrate-and-use, but skip the path China is actually running: modest frontier ambition, heavy investment in application and deployment, state coordination to nudge adoption. If returns accrue to deployment rather than frontier supremacy, and state nudging closes the very adoption gap you'd otherwise worry about, that combination answers both your concerns at once. Whether it works turns on the same unresolved question sitting in #9–14 — whether this is truly an increasing-returns world. If it is, China's frontier thinness eventually bites; if it isn't, it may prove the shrewder bet. Either way it should be on the options to consider.
I agree that the idea of an European homegrown frontier lab competitor is basically copium. But turning a blind eye to to the downsides of just accepting to always be dependent on others is also dangerous. We already saw how this played out for natural gas - the foregone revenue from European consumers didn't end up being much of a deterrent at the end
America being a democracy makes businesses and investors a pressure group in the way they are not in Russia. On your first point: once we accept that a homegrow frontier lab is copium, what choice does Europe have beyond using the American model?
Hard to say, it's a Scylla and Charybdis type situation. Maybe something akin to the Nasa-Esa model of cooperation for Artemis is possible- the American side calling the shots and pulling most of the weight, but European partners responsible for smaller yet still indispensable parts of the overall effort. Create a little bit of interdependence that goes both ways so that you at least get a seat at the table. Not sure how exactly this may look like in practice or if it is at all possible to get US labs and govt to agree on such a partnership.
> It seems unlikely that AI will be the only technology, or Anthropic the only company. Superhuman intelligence may make progress in plasma physics, but that will create a need for companies that build fusion reactors. Those could be German as much as they could be American.
I find this a confusing take. Nobody is saying it will be "the only technology". What does that even mean? Superhuman intelligence presumably can build fusion reactors, or at least manage all the cognitive work involved in it, much better than people. At least, eventually.
1) Europe doesn't need to be at the frontier, it just needs to have something good enough in case the US does slap export controls. Similar to fighter jets, Europe doesn't need to have jets that are as good as America's but ones that are good enough to fight Russia in case America's policy changes.
2) While at this stage it may look like the US labs are running away, the gap, as measured in time between milestones, between US-China-Mistral has remained pretty constant over time (also, there are other players).
3) Your argument that the US labs will be able to leverage their revenues to widen their lead is a bit questionable when none of them are profitable and when they are burning tens of billions on compute
Really enjoyed this piece. Everyone wants to know what the ideal outcome is and how to get there. Reality - as in many policy challenges - is there aren't ideal outcomes. Just paths to incrementally better places from which Europe (or other countries) may have more optionality and autonomy.
I'm generally pro building more capacity and... love the hard truths embedded in point 15 and 16. If were going to spend public dollars, let's at least try to reshape the market towards utility and commoditization to gain leverage and lower rent extraction.
Why on earth would someone write an article like this? I don’t understand the point of it. The lack of critical analysis makes it seem like a PR piece for the leading U.S. AI companies, whose main problem is, of course, well known worldwide: poor return on investment. They urgently need many more paying users or contacts that can be used for advertising so that their upfront capital expenditures might eventually pay off.
You should take your own advice—that Europe should ask itself how all this effort will ever pay off—and apply it to yourselves: How will your efforts ever pay off if Europe builds its own network of agents, AI agents for public administration, industry, security and law enforcement, and scientific purposes? This is, after all, the trend that’s actually emerging. (A quick tip: While US papers in medicine and bioinformatics don’t contain as many errors in the code as Chinese ones, the coding isn’t as good as in papers from Europe.)
It’s possible that non-paying internet users and small businesses will continue to use U.S. services indefinitely, but the real value created by AI might largely remain in Europe. There may also be more collaboration with non-EU countries, that like the idea of collaboration.
The problem with U.S. solutions has—unfortunately—arisen from the total politicization of technological progress: Today, politics is ideologically gripped by “America First and Only” and nationalist protectionism. If you’re going to fire up the big think tank, you mustn’t overlook this: There are deep-seated reasons why Europe can no longer believe in fair, constructive cooperation with the U.S.
Anyone who sometimes acts like an imperialist enemy, even almost fraternizes with Russia, and behaves in a non-cooperative manner in a thousand different ways—currently in their support for Ukraine—shouldn't be surprised if Europe finds itself forced to do everything to preserve key aspects of their sovereignty so as not to be completely vulnerable to blackmail. Or not?
And then the Europeans see the entire CEO team of AI and Big Tech constantly nodding in agreement with the Trump administration’s radical policies (Anthropic recently refused to bow to pressure, which was a smart move).
If I want to visit my brother in the U.S. again soon (as I have been doing since 1975), am I supposed to show my social media accounts at the border? Is there a chance that the Palantir software might mistake me for someone else, and I won't be able to enter the country at all, even though I have a visa? U.S.-European relations are in a deep crisis, and it’s not Europe’s fault.
"Europe has signaled that governments will tolerate destroying industries as important as the German car and chemical industries. It is not clear how you rebuild this trust." Precisely, it will be the failure to accelerate the transition, along with weak regulation, that ends up destroying the German automotive industry.
this singles out the Baltics for a property that applies to nearly every member state. France has its own AI champion to protect, Germany has its industrial base and energy politics, the Netherlands has ASML, Ireland has its tax relationship with US tech
also, the causal story is backwards. Baltic security dependence on the US isn't a "conflicted interest" in the sense of a preference that makes them unreliable, but an existential response
Indeed, it applies to every member state! I did not single them out as a particular weak point, but as an illiustration of the fact that member states can have wildly different priorities in their relationships with third countries.
Extremely valid arguments! I can only hope they will be internalized very quickly by the decision-makers.
I find your point 11 the most powerful. It says: "In a world of highly skewed returns to capabilities, the second mover earns nothing. The revenue that Mistral reported earlier this year is less than one percent of Anthropic’s today."
Another but related argument is that productivity in the American frontier companies is extremely high and would be very hard to match by the European late-starting "me-too" companies: SpaceX has about 22,000 employees, OpenAI has around 4,500–8,000 (planning to nearly double by the end of 2026), and Anthropic has roughly 2,500–5,000. Source: Grok.
That implies an urgent need for a very smart non-"brute-force" play, just as you recommend, Luis.
Hi Cristina! Thanks you very much for reading. My view is that sovereignty is a valuable goal, but that, in the long run, the clearest way to be sovereign is to be prosperous and rich. Japan is able to conduct its politics independently of the United States despite being massively integrated in the American economy; a place like Iraq is wholly independent economically but politically far more subjugated.
My fear is that a near-term effort at substitution will drain resources and political capital without delivering long term economic gains. There are basically two versions of the world: one where models keep improving exponentially and one where they do not.
In the former case, which it seems likely we are on, the model layer is not competitive and the American companeis do get to extract rents. In this world, they don't have their dominant position because of network effects and lock-ins, but because they simply have far better models (which are a function of their previous models), and users prefer these better models. In this world, I do not see any scenario for how Europe can provide a useful alternative. To put this concretely: in a world where Mythos plays a major role in accelerating AI research, Anthropic has an enormous advantage over Mistral.
In the second world, where the exponential ends, we are in much better shape. In this world, it is extremely like that there wil be commercial European providers at the frontier, or at least open source Chinese and American providers. In either case, it is not clear to me what a full-scale government effort at the model layer adds. I otherwise agree with you on many of your points, including that Europe is not a middle power.
Not necessarily that the models will stop improving, but that our usage of it will stop needing the extra power.
If you look at audio equipment, headphones specifically, we got better and better at them until the 70 and 80 where we reach the limits of normal human audition. To this day that audiophile community keep comparing everything to some references headphone that came out years ago and is still sold.
There still is a niche need for jetter quality equipement but that is not where most of the market is.
Likewise, there will be a point where AI model get good enough and cheap enough that getting the newest model is not worth the hastle when you can't spot the mistake in the previous generation
Hey Pieter! It's been a long time since we've last spoken (high-school CGU times...), but I recently stumbled upon your substack because of the interest/worry in rapid AI development that we apparently share.
I'd like to refer you to this article; https://europe2031.ai/ (there is a TLDR here: https://europe2031.ai/summary/) which was published this week. It relates to some of the points you list in the post, and I find it to be quite a convincing case for how the future might turn out if we don't take drastic action.
Ha Stan! Leuk om van je te horen. Ja, ik ben ook een groot fan van hun stuk (en heb op een paar plaatsen kunnen meehelpen). Cool om te horen dat jij je ook bezighoudt met AI vraagstukken. Ben je bekend met het Delta Instituut? Onno en Stan van Baarsen doen super goed werk voor Nederland daar.
Wat gaaf dat je daar aan hebt meegewerkt! En ja, die ken ik! Ga toevallig komende vrijdag op gesprek bij het Delta Instituut voor een mogelijk afstudeerproject :)
The 2031 scenario assumes that humans will be able to control a superintelligent AI. That's a big assumption. See https://www.youtube.com/watch?v=7wy3xyoXYt8
Europe should tell ASML to build kill switches into chips, so they can trip the kill switches if reckless American AI companies are about to destroy humanity.
While I understand your concern, I think what you're suggesting is not very realistic. We are lucky that, thanks to ASML, we are not completely excluded from the semiconductor supply chain, but we cannot (and should not) expect them to single-handedly regulate AI.
ASML is unique as of today, but in the model of the world you are describing (with superintelligent AI), I would not be surprised if other companies would slowly start to close that gap. It will be epecially hard to maintain that technological advantage if European companies (like ASML) get denied access to frontier AI models. The pullback of Mythos was the first real example of that.
I think we have the talent, capital and infrastructure to develop a wider portfolio of EU firms that can become essential to the global AI ecosystem. That would create a much broader co-dependent relationship with the US/China, which in turn puts us in a better position to have a say on the regulation of A(G)I.
>ASML is unique as of today, but in the model of the world you are describing (with superintelligent AI),
My view is that once superintelligent AI is achieved, it is likely to be unfriendly/misaligned with human values, and humans won't have much control anymore. It won't much matter which nation develops an unfriendly superintelligence. (Or a friendly superintelligence, for that matter, see https://www.astralcodexten.com/p/you-have-only-x-years-to-escape-permanent )
I'm focused on intervention points in the pre-superintelligence period. Presumably ASML will retain substantial competitive leverage in the pre-superintelligence period.
The message of the 2031 scenario was that Europe should think and act boldly. From my point of view, getting ASML to add kill switches is actually a much lighter lift than creating a hyperscaler/frontier model/etc.
Kill switches could be sold as part of Anthropic's effort to "build a pause button" for use in an emergency: https://www.anthropic.com/institute/recursive-self-improvement
Since the kill switches won't be activated at first, things will appear to be business as usual for ASML's customers. The Trump administration might not be super happy about kill switches, but public opinion in the US is skeptical of AI, and public opinion might matter more (side note: from my perspective as an American, Europeans have a recurring blind spot of focusing overmuch on government-to-government communications, and neglecting the role of US public opinion in transatlantic relations).
The kill switch thing will take time to design and set up, and it will only apply to newly manufactured chips. So it might be too late already, but I would go ahead and get started now anyways.
Simplistic analysis. Every point is valid, but it omits the most important one: if you don’t have vibrant internet companies, you don’t have the talent base needed to build internet giants. If you don’t have internet giants, you don’t have the talent base needed to build AI giants. So every time Europe—or anyone else, for that matter—delays the development of giants in the current leading technology, it also dramatically reduces its chances of playing a meaningful role in the next technological breakthrough.
A thought-provoking read. A few reactions:
Re. 2: The qualitative difference is that the Google ecosystem (and the internet and e-commerce more broadly) was an addition to the previously existing economy, whereas AI is partly a replacement. Europe’s lag in digitalisation was a missed opportunity; being left behind in AI, by contrast, means lasting economic decline.
Re. 4: Some of the other key technologies of the 2030s and 2040s will be space launch, semiconductors, quantum computing, humanoid robots, and BCIs—in all of which Europe is behind the U.S. and China (with the sole exception of EUV technology).
Re. 11: The graph shows that Mistral’s revenue was 2.9–4.4% of Anthropic’s at the same point in time (the beginning of this year). It also shows that Anthropic’s revenue was exactly 4.4% of OpenAI’s at the beginning of 2024. So should Anthropic have given up then?
Fast permissions to get on the grid are much more important than energy prices, as electricity is only a small share of the costs, see: https://carnegieendowment.org/research/2026/06/the-compute-coalition-how-to-build-the-future-of-ai-in-the-free-world Makes the value of permitting reform go up!
Clarifying piece — the convergence history and the second-mover problem make the anti-sovereignty case on frontier models simply and crisply. Two things I'd press, and one omission.
The pushbacks first. Point 2 — that model dependence is qualitatively no different from the Maps/Gmail/Search dependence Europe already lives with — is where I diagree. Search and Drive were tools one reached for; frontier models increasingly shape the judgment itself, what gets framed, weighed and eventually believed. Chor Pharn calls this the gap between economic and epistemic dependency: importing productivity is one thing, importing cognition another. The same entry on the balance sheet, a very different effect inside the institutions that have to metabolise it.
Then the harder question the piece sets aside: not "should Europe build a frontier model" — you answer that really well, no — but whether Europe can absorb the ones it already has. Access is not adoption. Europe has had frontier access throughout and still trails US adoption by a mile, which suggests sovereignty is not the binding constraint. #19 could become the whole essay.
The omission is China. You frame the choices as build-your-own-frontier or integrate-and-use, but skip the path China is actually running: modest frontier ambition, heavy investment in application and deployment, state coordination to nudge adoption. If returns accrue to deployment rather than frontier supremacy, and state nudging closes the very adoption gap you'd otherwise worry about, that combination answers both your concerns at once. Whether it works turns on the same unresolved question sitting in #9–14 — whether this is truly an increasing-returns world. If it is, China's frontier thinness eventually bites; if it isn't, it may prove the shrewder bet. Either way it should be on the options to consider.
I agree that the idea of an European homegrown frontier lab competitor is basically copium. But turning a blind eye to to the downsides of just accepting to always be dependent on others is also dangerous. We already saw how this played out for natural gas - the foregone revenue from European consumers didn't end up being much of a deterrent at the end
America being a democracy makes businesses and investors a pressure group in the way they are not in Russia. On your first point: once we accept that a homegrow frontier lab is copium, what choice does Europe have beyond using the American model?
Hard to say, it's a Scylla and Charybdis type situation. Maybe something akin to the Nasa-Esa model of cooperation for Artemis is possible- the American side calling the shots and pulling most of the weight, but European partners responsible for smaller yet still indispensable parts of the overall effort. Create a little bit of interdependence that goes both ways so that you at least get a seat at the table. Not sure how exactly this may look like in practice or if it is at all possible to get US labs and govt to agree on such a partnership.
> It seems unlikely that AI will be the only technology, or Anthropic the only company. Superhuman intelligence may make progress in plasma physics, but that will create a need for companies that build fusion reactors. Those could be German as much as they could be American.
I find this a confusing take. Nobody is saying it will be "the only technology". What does that even mean? Superhuman intelligence presumably can build fusion reactors, or at least manage all the cognitive work involved in it, much better than people. At least, eventually.
Some thoughts:
1) Europe doesn't need to be at the frontier, it just needs to have something good enough in case the US does slap export controls. Similar to fighter jets, Europe doesn't need to have jets that are as good as America's but ones that are good enough to fight Russia in case America's policy changes.
2) While at this stage it may look like the US labs are running away, the gap, as measured in time between milestones, between US-China-Mistral has remained pretty constant over time (also, there are other players).
3) Your argument that the US labs will be able to leverage their revenues to widen their lead is a bit questionable when none of them are profitable and when they are burning tens of billions on compute
Really enjoyed this piece. Everyone wants to know what the ideal outcome is and how to get there. Reality - as in many policy challenges - is there aren't ideal outcomes. Just paths to incrementally better places from which Europe (or other countries) may have more optionality and autonomy.
I'm generally pro building more capacity and... love the hard truths embedded in point 15 and 16. If were going to spend public dollars, let's at least try to reshape the market towards utility and commoditization to gain leverage and lower rent extraction.
Would love to talk to you about our Parting Clouds report (https://antimonopoly.ca/parting-clouds-creating-a-competitive-marketplace-for-compute/), grounded in part The Path to a Sovereign Tech Stack is Via a Commodified Tech Stack piece(https://www.techpolicy.press/the-path-to-a-sovereign-tech-stack-is-via-a-commodified-tech-stack/) or this Foreign Policy magazine piece on cloud and NATO (https://foreignpolicy.com/2026/06/09/nato-cloud-computer-ai-iran-war/). Shameless link sharing, but really enjoyed this and would love to see if there are overlaps in thinking.
Why on earth would someone write an article like this? I don’t understand the point of it. The lack of critical analysis makes it seem like a PR piece for the leading U.S. AI companies, whose main problem is, of course, well known worldwide: poor return on investment. They urgently need many more paying users or contacts that can be used for advertising so that their upfront capital expenditures might eventually pay off.
You should take your own advice—that Europe should ask itself how all this effort will ever pay off—and apply it to yourselves: How will your efforts ever pay off if Europe builds its own network of agents, AI agents for public administration, industry, security and law enforcement, and scientific purposes? This is, after all, the trend that’s actually emerging. (A quick tip: While US papers in medicine and bioinformatics don’t contain as many errors in the code as Chinese ones, the coding isn’t as good as in papers from Europe.)
It’s possible that non-paying internet users and small businesses will continue to use U.S. services indefinitely, but the real value created by AI might largely remain in Europe. There may also be more collaboration with non-EU countries, that like the idea of collaboration.
The problem with U.S. solutions has—unfortunately—arisen from the total politicization of technological progress: Today, politics is ideologically gripped by “America First and Only” and nationalist protectionism. If you’re going to fire up the big think tank, you mustn’t overlook this: There are deep-seated reasons why Europe can no longer believe in fair, constructive cooperation with the U.S.
Anyone who sometimes acts like an imperialist enemy, even almost fraternizes with Russia, and behaves in a non-cooperative manner in a thousand different ways—currently in their support for Ukraine—shouldn't be surprised if Europe finds itself forced to do everything to preserve key aspects of their sovereignty so as not to be completely vulnerable to blackmail. Or not?
And then the Europeans see the entire CEO team of AI and Big Tech constantly nodding in agreement with the Trump administration’s radical policies (Anthropic recently refused to bow to pressure, which was a smart move).
If I want to visit my brother in the U.S. again soon (as I have been doing since 1975), am I supposed to show my social media accounts at the border? Is there a chance that the Palantir software might mistake me for someone else, and I won't be able to enter the country at all, even though I have a visa? U.S.-European relations are in a deep crisis, and it’s not Europe’s fault.
"Europe has signaled that governments will tolerate destroying industries as important as the German car and chemical industries. It is not clear how you rebuild this trust." Precisely, it will be the failure to accelerate the transition, along with weak regulation, that ends up destroying the German automotive industry.
Pieter and Simon, I sincerely apologize for attributing your very informative piece to Luis. No excuses.
No problem!
what framing is this regarding Baltics?
this singles out the Baltics for a property that applies to nearly every member state. France has its own AI champion to protect, Germany has its industrial base and energy politics, the Netherlands has ASML, Ireland has its tax relationship with US tech
also, the causal story is backwards. Baltic security dependence on the US isn't a "conflicted interest" in the sense of a preference that makes them unreliable, but an existential response
Indeed, it applies to every member state! I did not single them out as a particular weak point, but as an illiustration of the fact that member states can have wildly different priorities in their relationships with third countries.
Extremely valid arguments! I can only hope they will be internalized very quickly by the decision-makers.
I find your point 11 the most powerful. It says: "In a world of highly skewed returns to capabilities, the second mover earns nothing. The revenue that Mistral reported earlier this year is less than one percent of Anthropic’s today."
Another but related argument is that productivity in the American frontier companies is extremely high and would be very hard to match by the European late-starting "me-too" companies: SpaceX has about 22,000 employees, OpenAI has around 4,500–8,000 (planning to nearly double by the end of 2026), and Anthropic has roughly 2,500–5,000. Source: Grok.
That implies an urgent need for a very smart non-"brute-force" play, just as you recommend, Luis.
Hi Cristina! Thanks you very much for reading. My view is that sovereignty is a valuable goal, but that, in the long run, the clearest way to be sovereign is to be prosperous and rich. Japan is able to conduct its politics independently of the United States despite being massively integrated in the American economy; a place like Iraq is wholly independent economically but politically far more subjugated.
My fear is that a near-term effort at substitution will drain resources and political capital without delivering long term economic gains. There are basically two versions of the world: one where models keep improving exponentially and one where they do not.
In the former case, which it seems likely we are on, the model layer is not competitive and the American companeis do get to extract rents. In this world, they don't have their dominant position because of network effects and lock-ins, but because they simply have far better models (which are a function of their previous models), and users prefer these better models. In this world, I do not see any scenario for how Europe can provide a useful alternative. To put this concretely: in a world where Mythos plays a major role in accelerating AI research, Anthropic has an enormous advantage over Mistral.
In the second world, where the exponential ends, we are in much better shape. In this world, it is extremely like that there wil be commercial European providers at the frontier, or at least open source Chinese and American providers. In either case, it is not clear to me what a full-scale government effort at the model layer adds. I otherwise agree with you on many of your points, including that Europe is not a middle power.
I want to make an argument for the second case.
Not necessarily that the models will stop improving, but that our usage of it will stop needing the extra power.
If you look at audio equipment, headphones specifically, we got better and better at them until the 70 and 80 where we reach the limits of normal human audition. To this day that audiophile community keep comparing everything to some references headphone that came out years ago and is still sold.
There still is a niche need for jetter quality equipement but that is not where most of the market is.
Likewise, there will be a point where AI model get good enough and cheap enough that getting the newest model is not worth the hastle when you can't spot the mistake in the previous generation
Europe is a SPINO, super power in name only. It’s headed towards irrelevance.