As one of the co-authors of the CERN for AI piece you cite, I've definitely moved away from a publicly-funded effort to directly compete with labs at the frontier, but I'm also sceptical of strategies that rely on open models.
It looks like Chinese models are misreporting their evaluation results and have inflated performance as a result -> when independently evaluated, the gap is increasing between the frontier and open models.
Unclear how much of this is because of intense compute shortages. Regardless, seems like Europe should be building more compute.
I agree with your take on why the model layer won't capture value—and it maps directly onto something I've been working on about why different countries are optimizing for different AI finish lines.
Your three theses explain why the pattern observed empirically makes economic sense. The US is racing toward engagement and platform value (model layer), but as you point out, that layer is sandwiched between monopolist hardware suppliers and customers with zero switching costs. Meanwhile, China has been optimizing for exactly what you're prescribing for Europe: deployment in the implementation layer where org design, not IQ, is the binding constraint.
In my piece, I explored why America builds AI for attention (girlfriends, viral moments) while China wires it into factories, hospitals, and power grids. Your thesis 3—"intelligence is not the bottleneck"—is the theoretical foundation for why China's deployment track will capture more value than America's model-layer race. The marginal gain of moving from GPT-4o to GPT-5.2 is small compared to redesigning a hospital workflow or integrating AI into injection-molding quality control.
So, China is way ahead in executing the "smart second mover" strategy you recommend for Europe. Their progress confirms you don't need frontier models to deploy effectively—you need good-enough models plus org capacity and a willingness to experiment.
Can Europe imitate the China path? Yes but only if Europe can build the org and regulatory capacity to execute on implementation when they don't have China's state coordination or America's VC density.
As always, very thoughtful—thank you. A few points:
1. After the initial euphoria, most consumers—whose budget constraints are binding, with budget shares adding up to 100% and not fully adjustable—will choose to use the older, less expensive versions, much like they stick to old iPhones and cars.
2. This trend will intensify competition.
3. Additionally, market participants know that tacit collusion is very difficult when n > 2;
4. So the war of attrition will continue, but not forever.
5. In this war, the announcements of future spending signal a determination to outspend rivals. However, if the previous arguments hold, these signals may carry limited credibility.
Thesis 3 is the one that should reshape the investment conversation. If intelligence itself isnt the bottleneck, then the hundreds of billions being poured into producing more of it are being invested in the wrong constraint. The binding constraint is organisational, how companies restructure workflows to actually absorb what the models can already do. That makes the AI value chain look like a barbell: durable margin at the hardware chokepoints above and at the implementation layer below, with the model layer squeezed between customers who can switch and suppliers who can't be replaced.
The EMI parallel is the sharpest diagnosis in the piece. The labs are building in the competitive middle of their own value chain, compelled to reinvest every dollar into staying ahead of a frontier that the open-weight fringe compresses from below. Thats a structurally unprofitable position regardless of how impressive the technology is. The market is pricing the labs as if they occupy the chokepoint. The IO economics say they occupy the squeeze zone.
(Copying this from X!) I'm generally in favour of adoption-focused strategy, but a little worried about fully betting on this thesis for two reasons:
first, it seems like the divergence between labs and fast followers is widening due to structural forces that will get much stronger. fast-following was decently easy when no one had big AI datacenters online, flywheels hadn't started, and distillation and research was still a wild west. that means you could build a fairly efficient fast follower, and neither the capability gap nor the efficiency gap matters much.
But all of that is changing - the asymmetries in compute between leading US developers and the world are just starting to kick in, and they are baked into at least the next four years of global compute trajectories because of the lead times on buildouts and chip procurement. crackdowns on distillation, as well as the fact that leading developers get much more revenue to reinvest into talent and compute and much more data to further widen the gap with makes it much worse. I wouldn't be surprised if the gap to fast followers extends to 1.5 years in an otherwise accelerating environment, and I also wouldn't be surprised if that gap also meant that the frontier labs will be able to offer much more efficient sub-frontier models than any pure fast follower competitor.
And second, I'm not sure that 'how many tasks can you do with non-frontier models' is the relevant question, it's 'how prohibitively important are the tasks for which you do need frontier models because you're engaged in some directly competitive dynamic, e.g. between attackers and defenders or directly competing wrappers around frontier systems'.
To the extent that the latter become economically vital and the ability to run them remains scarce, you could just imagine a lot of value (and especially leverage) concentration with those that can provide comparatively leading models even in the world where most tasks are carried out by sub-frontier models, because value asymmetrically accrues around the tasks that do require access to comparatively-frontier instead of absolutely-good models. that's definitely a contentious prediction, but it seems devastating to this strategy if true.
Thanks Anton. On the competitive dynamics that produce a winner-take-all market: agreed. If you are at war, or even just playing deterrence, you need the frontier model and nothing less. The question is how much of the economy works that way. War is the limiting case where the marginal capability is worth almost anything. But medicine, education, manufacturing, law, tax administration: in those, getting it right ninety percent of the time is the whole game, and the gap between the frontier and the model one tier back buys little.
On the gaps expanding, I agree they may. But look at what is actually happening in the frontier market, because I believe it points the other way. OpenAI is some distance ahead of Anthropic on raw model performance --see e.g. math-- ( ChatGPT 5.5 Pro Extended vs. Claude Opus 4.7, leaving Mythos aside). Yet Anthropic has taken the lead in the market, because its harness is far superior. The better model is not winning. The better implementation is. TEvem where intelligence is most directly the product, it is the wrapping around the model, the workflow that decides who captures the value.
I understand the proposal for the ‘middle powers’ to build and maintain a ‘second tier’ model is about ‘insurance’, ie, against the case that model development becomes a monopoly. If this understanding is correct, it would imply that if the monopolisation scenario does not materialise, the enterprise of building and maintaining a second-tier model would be loss-making, would not it (just like you suffer.a net loss if you buy insurance against events that do not materialise)? In particular, as it has already been suggested, frontier model developers should always be able to offer second-tier models at a lower price than non-subsidised independent developers, should not they?
Excellent analysis. Working in applied AI for enterprises, I'd add that the binding constraint in the implementation layer is reliability of agentic deployment.
Large models still hallucinate and, on hard tasks, attempt shortcuts and "cheat" — these are core reasons for the high failure rates reported in enterprise adoption. LLMs working as agents need to reason about the next step or action in a business process, but the involved systems — ERPs, mainframes, legacy software — are rarely fully known. The LLM only has access to an API or UI; the inner workings and state transitions of these systems are not fully documented or discoverable. So the prediction of which action produces which state change is often incomplete, and at times wrong.
Any successful implementation layer needs to overcome this lack of full knowledge of the enterprise IT landscape. Agents need to become a combination of LLM and prediction system, capable of processing at very high reliability — say 99.9% — because otherwise errors accumulate too fast and business process outcomes degrade. The harness around the model must include a layer that explicitly represents the probabilistic nature of its knowledge about involved systems.
This connects directly to Garicano's Thesis 3: if intelligence isn't the bottleneck, reliability is. And if the reliability problem isn't solved quickly, enterprise demand for agentic AI may stall well before it justifies the current investment in frontier models and data centers. Enterprise demand for an internal chatbot or office support assistant alone is probably not large enough. AI needs to reliably perform full business processes end-to-end — and that's an implementation problem, not a model capability problem.
I feel like the body of your argument goes in the opposite direction of the thesis. Frontier models will be commoditized, making OpenAI-vs-Anthropic into an unprofitable fight. Okay that’s a reasonable theory. Therefore… Europe should spend lots of money to join the fight?
The logical conclusion would be, *don’t* try to join the fight. Instead encourage companies that are complementary. Ie, companies that extensively use the OpenAI or Anthropic APIs, and thus are benefited by competition between the two.
Well, we don't want to join the fight! We are advocating staying several tiers behind, as one of the strategies (together with the ones in the first piece on "Smart Second Mover" along the lines you suggest) to succeed in a world where uncertainty is high and the imperative is keeping some options and some choice.
Well, you already have Mistral. The market of "companies who are forced by regulation to buy a European product" seems critical to them. If anything, it seems more likely that some government funded competitor would hurt Mistral and then fall apart, than that it would have an effect on the broader AI model market beyond Europe.
ElevenLabs and Helsing both seem promising, too - not competing at all in the fundamental model space, but at something slightly different. If I were designed European policy, I would try to be encouraging more companies along those lines, rather than trying to get government entities to write software that takes on OpenAI's market head on.
We are not saying that it has to be a government entity, it could and should be run efficiently. Even Mistral could be the starting point of that entity. What is needed is that Mistral or any other entity is prepared (and capable) to play the strategic hedge against uncertainty that we are describing and to be become potentially a public good.
You seem to be proposing https://apertvs.ai/ which has existed for quite some time? Training so-so language models is fairly well understood now. But you don't mention this effort right on your doorstep, which is my point: non-frontier models just don't get used much and certainly don't get written about. So I am not sure how building more models that are less capable than the leading models is helping.
Funding is another variable to throw into the mix. Thus far the frontier seem to have had uncapped access to the private capital markets and are working towards access to the public capital markets to continue their spree.
However, despite improving margins on inference the direct RoI on each marginal training run seems pretty questionable from the outside. It only makes sense if seen as a prerequiste to the next n runs, that will eventually lead to a defensible level of performance.
Will such a thing ever materialize? The public markets could easily get nervous about that on the basis of more granular disclosures.
On the other hand, recent compute shortages to serve demand will probably push labs to be more aggressive on capacity leases (debt-like?), increasing their fragility in any bumpy scenario.
Open weights aren't the same as open development—the actual training still happens inside a handful of companies. Project Tapestry (just launched by the AI Alliance, with Yann LeCun as Chief Science Advisor) is trying to change that: federated training where nations and institutions co-build a shared model but keep their data local. Curious what people here think. https://thealliance.ai/projects/tapestry
As one of the co-authors of the CERN for AI piece you cite, I've definitely moved away from a publicly-funded effort to directly compete with labs at the frontier, but I'm also sceptical of strategies that rely on open models.
At the very least, Chinese open models appear to be falling further and further behind the frontier: https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro
It looks like Chinese models are misreporting their evaluation results and have inflated performance as a result -> when independently evaluated, the gap is increasing between the frontier and open models.
Unclear how much of this is because of intense compute shortages. Regardless, seems like Europe should be building more compute.
I agree with your take on why the model layer won't capture value—and it maps directly onto something I've been working on about why different countries are optimizing for different AI finish lines.
Your three theses explain why the pattern observed empirically makes economic sense. The US is racing toward engagement and platform value (model layer), but as you point out, that layer is sandwiched between monopolist hardware suppliers and customers with zero switching costs. Meanwhile, China has been optimizing for exactly what you're prescribing for Europe: deployment in the implementation layer where org design, not IQ, is the binding constraint.
In my piece, I explored why America builds AI for attention (girlfriends, viral moments) while China wires it into factories, hospitals, and power grids. Your thesis 3—"intelligence is not the bottleneck"—is the theoretical foundation for why China's deployment track will capture more value than America's model-layer race. The marginal gain of moving from GPT-4o to GPT-5.2 is small compared to redesigning a hospital workflow or integrating AI into injection-molding quality control.
https://rajeshachanta.substack.com/p/spectacle-vs-scaffolding
So, China is way ahead in executing the "smart second mover" strategy you recommend for Europe. Their progress confirms you don't need frontier models to deploy effectively—you need good-enough models plus org capacity and a willingness to experiment.
Can Europe imitate the China path? Yes but only if Europe can build the org and regulatory capacity to execute on implementation when they don't have China's state coordination or America's VC density.
Agree on everything. Execution is not independent of the institutions. Regulatory barriers etc. are crucial to success there.
As always, very thoughtful—thank you. A few points:
1. After the initial euphoria, most consumers—whose budget constraints are binding, with budget shares adding up to 100% and not fully adjustable—will choose to use the older, less expensive versions, much like they stick to old iPhones and cars.
2. This trend will intensify competition.
3. Additionally, market participants know that tacit collusion is very difficult when n > 2;
4. So the war of attrition will continue, but not forever.
5. In this war, the announcements of future spending signal a determination to outspend rivals. However, if the previous arguments hold, these signals may carry limited credibility.
Thesis 3 is the one that should reshape the investment conversation. If intelligence itself isnt the bottleneck, then the hundreds of billions being poured into producing more of it are being invested in the wrong constraint. The binding constraint is organisational, how companies restructure workflows to actually absorb what the models can already do. That makes the AI value chain look like a barbell: durable margin at the hardware chokepoints above and at the implementation layer below, with the model layer squeezed between customers who can switch and suppliers who can't be replaced.
The EMI parallel is the sharpest diagnosis in the piece. The labs are building in the competitive middle of their own value chain, compelled to reinvest every dollar into staying ahead of a frontier that the open-weight fringe compresses from below. Thats a structurally unprofitable position regardless of how impressive the technology is. The market is pricing the labs as if they occupy the chokepoint. The IO economics say they occupy the squeeze zone.
(Copying this from X!) I'm generally in favour of adoption-focused strategy, but a little worried about fully betting on this thesis for two reasons:
first, it seems like the divergence between labs and fast followers is widening due to structural forces that will get much stronger. fast-following was decently easy when no one had big AI datacenters online, flywheels hadn't started, and distillation and research was still a wild west. that means you could build a fairly efficient fast follower, and neither the capability gap nor the efficiency gap matters much.
But all of that is changing - the asymmetries in compute between leading US developers and the world are just starting to kick in, and they are baked into at least the next four years of global compute trajectories because of the lead times on buildouts and chip procurement. crackdowns on distillation, as well as the fact that leading developers get much more revenue to reinvest into talent and compute and much more data to further widen the gap with makes it much worse. I wouldn't be surprised if the gap to fast followers extends to 1.5 years in an otherwise accelerating environment, and I also wouldn't be surprised if that gap also meant that the frontier labs will be able to offer much more efficient sub-frontier models than any pure fast follower competitor.
And second, I'm not sure that 'how many tasks can you do with non-frontier models' is the relevant question, it's 'how prohibitively important are the tasks for which you do need frontier models because you're engaged in some directly competitive dynamic, e.g. between attackers and defenders or directly competing wrappers around frontier systems'.
To the extent that the latter become economically vital and the ability to run them remains scarce, you could just imagine a lot of value (and especially leverage) concentration with those that can provide comparatively leading models even in the world where most tasks are carried out by sub-frontier models, because value asymmetrically accrues around the tasks that do require access to comparatively-frontier instead of absolutely-good models. that's definitely a contentious prediction, but it seems devastating to this strategy if true.
Thanks Anton. On the competitive dynamics that produce a winner-take-all market: agreed. If you are at war, or even just playing deterrence, you need the frontier model and nothing less. The question is how much of the economy works that way. War is the limiting case where the marginal capability is worth almost anything. But medicine, education, manufacturing, law, tax administration: in those, getting it right ninety percent of the time is the whole game, and the gap between the frontier and the model one tier back buys little.
On the gaps expanding, I agree they may. But look at what is actually happening in the frontier market, because I believe it points the other way. OpenAI is some distance ahead of Anthropic on raw model performance --see e.g. math-- ( ChatGPT 5.5 Pro Extended vs. Claude Opus 4.7, leaving Mythos aside). Yet Anthropic has taken the lead in the market, because its harness is far superior. The better model is not winning. The better implementation is. TEvem where intelligence is most directly the product, it is the wrapping around the model, the workflow that decides who captures the value.
I understand the proposal for the ‘middle powers’ to build and maintain a ‘second tier’ model is about ‘insurance’, ie, against the case that model development becomes a monopoly. If this understanding is correct, it would imply that if the monopolisation scenario does not materialise, the enterprise of building and maintaining a second-tier model would be loss-making, would not it (just like you suffer.a net loss if you buy insurance against events that do not materialise)? In particular, as it has already been suggested, frontier model developers should always be able to offer second-tier models at a lower price than non-subsidised independent developers, should not they?
Excellent analysis. Working in applied AI for enterprises, I'd add that the binding constraint in the implementation layer is reliability of agentic deployment.
Large models still hallucinate and, on hard tasks, attempt shortcuts and "cheat" — these are core reasons for the high failure rates reported in enterprise adoption. LLMs working as agents need to reason about the next step or action in a business process, but the involved systems — ERPs, mainframes, legacy software — are rarely fully known. The LLM only has access to an API or UI; the inner workings and state transitions of these systems are not fully documented or discoverable. So the prediction of which action produces which state change is often incomplete, and at times wrong.
Any successful implementation layer needs to overcome this lack of full knowledge of the enterprise IT landscape. Agents need to become a combination of LLM and prediction system, capable of processing at very high reliability — say 99.9% — because otherwise errors accumulate too fast and business process outcomes degrade. The harness around the model must include a layer that explicitly represents the probabilistic nature of its knowledge about involved systems.
This connects directly to Garicano's Thesis 3: if intelligence isn't the bottleneck, reliability is. And if the reliability problem isn't solved quickly, enterprise demand for agentic AI may stall well before it justifies the current investment in frontier models and data centers. Enterprise demand for an internal chatbot or office support assistant alone is probably not large enough. AI needs to reliably perform full business processes end-to-end — and that's an implementation problem, not a model capability problem.
I feel like the body of your argument goes in the opposite direction of the thesis. Frontier models will be commoditized, making OpenAI-vs-Anthropic into an unprofitable fight. Okay that’s a reasonable theory. Therefore… Europe should spend lots of money to join the fight?
The logical conclusion would be, *don’t* try to join the fight. Instead encourage companies that are complementary. Ie, companies that extensively use the OpenAI or Anthropic APIs, and thus are benefited by competition between the two.
Well, we don't want to join the fight! We are advocating staying several tiers behind, as one of the strategies (together with the ones in the first piece on "Smart Second Mover" along the lines you suggest) to succeed in a world where uncertainty is high and the imperative is keeping some options and some choice.
Well, you already have Mistral. The market of "companies who are forced by regulation to buy a European product" seems critical to them. If anything, it seems more likely that some government funded competitor would hurt Mistral and then fall apart, than that it would have an effect on the broader AI model market beyond Europe.
ElevenLabs and Helsing both seem promising, too - not competing at all in the fundamental model space, but at something slightly different. If I were designed European policy, I would try to be encouraging more companies along those lines, rather than trying to get government entities to write software that takes on OpenAI's market head on.
We are not saying that it has to be a government entity, it could and should be run efficiently. Even Mistral could be the starting point of that entity. What is needed is that Mistral or any other entity is prepared (and capable) to play the strategic hedge against uncertainty that we are describing and to be become potentially a public good.
You seem to be proposing https://apertvs.ai/ which has existed for quite some time? Training so-so language models is fairly well understood now. But you don't mention this effort right on your doorstep, which is my point: non-frontier models just don't get used much and certainly don't get written about. So I am not sure how building more models that are less capable than the leading models is helping.
Funding is another variable to throw into the mix. Thus far the frontier seem to have had uncapped access to the private capital markets and are working towards access to the public capital markets to continue their spree.
However, despite improving margins on inference the direct RoI on each marginal training run seems pretty questionable from the outside. It only makes sense if seen as a prerequiste to the next n runs, that will eventually lead to a defensible level of performance.
Will such a thing ever materialize? The public markets could easily get nervous about that on the basis of more granular disclosures.
On the other hand, recent compute shortages to serve demand will probably push labs to be more aggressive on capacity leases (debt-like?), increasing their fragility in any bumpy scenario.
Fun times ahead.
Open weights aren't the same as open development—the actual training still happens inside a handful of companies. Project Tapestry (just launched by the AI Alliance, with Yann LeCun as Chief Science Advisor) is trying to change that: federated training where nations and institutions co-build a shared model but keep their data local. Curious what people here think. https://thealliance.ai/projects/tapestry