Should countries outside the US build their own frontier or sovereign models, or concede that layer and compete elsewhere?

Daniel Khachab · Oct 28, 2024

Europe's lack of elite AI institutions matters for the foundation-model layer but not for the application layer, where everyone started from zero in 2023 and the skills are new.

Application-layer skills — integrating foundation model APIs, prompt engineering, designing such products — are new to everyone, so teams can upskill into the race.

Scope: concedes the point for the foundational layer; application layer only

25:59 20VC: Why SaaS is Dead | Why AI First Companies Will Win | We are in the Middle of a Cold War for AI Talent | Why Europe is F******* and We Need to Stop Whining with Daniel Khachab, Co-Founder @ Choco

Daniel Khachab · Oct 28, 2024

Foundation models are Europe's hardest gap because of a structural talent shortage and the absence of a founder willing to commit a lifetime to building one.

Building a merely functional model isn't the bar — it must compete with OpenAI, Anthropic, Grok and Llama, which requires a level of talent and founder commitment Europe lacks.

30:34 20VC: Why SaaS is Dead | Why AI First Companies Will Win | We are in the Middle of a Cold War for AI Talent | Why Europe is F******* and We Need to Stop Whining with Daniel Khachab, Co-Founder @ Choco

Mike Krieger · Mar 3, 2025 · hedged

Europe is unlikely to compete at the frontier lab level, but application-layer innovation on top of models is plausible there because nimble startups can outpace labs serving hundreds of millions of users — though it would require a different regulatory and startup ecosystem than exists today

Compute access constrains labs, while building applications on top of models is getting much easier and big labs have to move slowly at their scale

Scope: harder question to answer from a labs perspective; conditional on regulatory and ecosystem change

59:20 20VC: Anthropic CPO Mike Krieger: Where Will Value Be Created in a World of AI | Have Foundation Models Commoditized | When Do Model Providers Become Application Providers | What Anthropic Learned from Deepseek

Brendan Foody · Jun 1, 2026

Europe should accept it has lost the foundation model race and focus on post-training, distillation and applications rather than competing head-to-head with labs like Anthropic

There is real value in the work that happens after foundation models are built and in the application layer, but not in a direct frontier contest

60:34 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility, The Model is the Product | Token Spend Will Exceed Headcount Spend in 5 Years | The True Cost of Hiring AI Researchers in the Valley Today with Brendan Foody

Arvind Jain · Jul 11, 2026 · hedged

Demand for sovereign models is weaker now than a year ago, as many nations concluded they cannot build their own and are content to let their enterprises use OpenAI, Anthropic and other models

Nations that thought they could build models in AI's early stages figured out that isn't the way

Scope: says he is not an expert and unsure of the trend direction

42:43 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Arvind Jain · Jul 11, 2026

China is the only country outside the US to have produced meaningful models, with perhaps a small exception in France with Mistral

Sovereign model ambitions simply haven't materialised anywhere else

43:51 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Daniel Khachab · Oct 28, 2024

Producing energy and attracting a chip manufacturer are more within Europe's control than building a competitive foundation model, which depends on the right person appearing at the right time and being willing to take the risk.

Foundation models require a specific rare founder, an inherently structural and uncontrollable input, whereas infrastructure is a policy choice.

Scope: partially concedes Harry's point

31:33 20VC: Why SaaS is Dead | Why AI First Companies Will Win | We are in the Middle of a Cold War for AI Talent | Why Europe is F******* and We Need to Stop Whining with Daniel Khachab, Co-Founder @ Choco

Matan Grinberg · Jun 13, 2026 · hedged

Europe is probably too far behind to catch up on the frontier model lab side, but still has a large opportunity in infrastructure buildout and energy

Democratic countries move slowly — you need broad support, convincing, and legislation — which is a disadvantage in a race where the required actions (data centers, energy) are clear and have long lead times

Scope: 'probably' on the frontier lab side; the slowness has an upside: correction and balance against going too far in one direction

66:04 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founder with Matan Grinberg, Founder @ Factory

Harry Stebbings · Jun 29, 2026 · hedged

Energy sovereignty matters more than model sovereignty in the European debate about AI independence

Scope: framed as 'it strikes me'

39:31 20VC: Leo Aschenbrenner's Largest Holding: Inside the $90BN Bloom Energy | Why Electricity, Not AI Models, Will Decide the Winners of the AI Race | Why We Are Not in an AI Capex Bubble | Energy Sovereignty and The Future of Power with KR Sridhar

Brendan Foody · Sep 15, 2025 · hedged

Sovereignty can let a model provider win a scoped part of the market, but the largest model companies will not be geography-specific ones

A provider like Mistral could be the best on European law's nuances, but the biggest companies will win on broad general-purpose capabilities people use daily to code and work

Scope: scoped part of the market only

45:08 20VC: Mercor: From $1M to $500M in 17 Months: The Fastest Growing Company in the World | How to Think About Margins and Revenue Sustainability in AI | Why Evaluation Benchmarks in AI are BS Today with Brendan Foody

Brendan Foody · Jun 1, 2026 · hedged

Sovereign models are only sometimes justified: jurisdictional localization is a real last mile, but frontier labs will simply hire thousands of local specialists and their transfer learning advantage makes them hard to stop

Labs already ask for models good at British or French law, and will hire 10,000 people in France to teach it; transfer learning from other domains is too powerful for local competitors to overcome

Scope: 'maybe in some cases'

61:04 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility, The Model is the Product | Token Spend Will Exceed Headcount Spend in 5 Years | The True Cost of Hiring AI Researchers in the Valley Today with Brendan Foody

Nick Frosst · Sep 1, 2025

Being a Canadian company is a commercial asset, because the US has shown willingness to cut off access to technology for political reasons and the line between American tech and the American government is blurring

Companies in Canada and around the world now want to work with non-American tech vendors; the government taking a 10% stake in Intel is an example of the blurring

Scope: 'I'm not an economist and I don't know if that's good for the country or not'

51:10 20VC: Cohere Founder on How Cohere Compete with OpenAI and Anthropic $BNs | Why Counties Should Fund Their Own Models & the Need for Model Sovereignty | How Sam Altman Has Done a Disservice to AI with Nick Frosst

Harry Stebbings · Jul 11, 2026

Demand for sovereign models is unequivocally on the rise, because the Trump administration banning Anthropic's latest models showed Europe it cannot rely on a US individual who could cut off access to intelligence

A US administration demonstrated it can revoke access, making dependence on US models a strategic risk

Scope: concedes no results yet since the ban was only a month ago

43:26 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Emad Mostaque · May 17, 2023

Every nation needs its own models and datasets rather than all models coming from Palo Alto, because local context is encoded in models and we are outsourcing our thinking to them

Stable Diffusion rendered 'salaryman' as a happy man while in Japan a salaryman is a sad man; models are like talented grads who occasionally go off their meds, and you want ones that understand your local context so they serve you better

18:23 20VC: Why the AI Bubble Will Be Bigger Than The Dot Com Bubble, Why AI Will Have a Bigger Impact Than COVID, Why No Models Used Today Will Be Used in a Year, Why All Models are Biased and How AI Kills Traditional Media with Emad Mostaque, Founder & CEO @

Nick Frosst · Sep 1, 2025

Language models are infrastructure, and countries are better served by models built in their own language, dialect and cultural context than by relying on Chinese or American models

A model with the cultural fluency of the country empowers its people and economy in a way an imported model cannot; it's like building national infrastructure

Scope: unsure what the end state actually looks like in practice

49:45 20VC: Cohere Founder on How Cohere Compete with OpenAI and Anthropic $BNs | Why Counties Should Fund Their Own Models & the Need for Model Sovereignty | How Sam Altman Has Done a Disservice to AI with Nick Frosst

Anjney Midha · Apr 14, 2026

Because the US Cloud Act gives the US government access to data on infrastructure managed by American companies, European military, defense and other mission-critical workloads cannot send that sensitive context to US-managed AI infrastructure, pushing those organizations to seek trusted local AI infrastructure partners

Sensitive supply chain, military and mission-critical context can't legally or safely cross the border into US-managed clouds, so those customers seek local providers

Scope: applies to mission-critical/sovereign workloads in Europe

11:20 20VC: Anj Midha on Investing $300M into Anthropic | The Early Days of Anthropic & How 21 of 22 VCs Turned it Down | The Four Bottlenecks to Compute | What the China Has Smashed and Why We Should Be Worried

Anjney Midha · Apr 14, 2026

The investment thesis in Mistral is a fully independent European AI stack — sovereign land, power and shells, local compute, and openly released models that can be deployed and customized anywhere

Scope: Europe-focused, though open models deployable globally

13:49 20VC: Anj Midha on Investing $300M into Anthropic | The Early Days of Anthropic & How 21 of 22 VCs Turned it Down | The Four Bottlenecks to Compute | What the China Has Smashed and Why We Should Be Worried

Andrew Feldman · Oct 6, 2025

Sovereignty is a genuine strategic advantage for Mistral, and combined with the fastest inference hardware it makes their LeChat product compelling

There were too few AI labs doing interesting work in Europe, so Mistral played the 'Europe's leader' card well

44:05 Sovereignty combined with hardware speed is a genuine competitive advantage for mistral

Hemant Taneja · Sep 22, 2025

Compute productivity should be captured onshore within each geography rather than leaking to a single company or country, since that concentration creates imbalance.

An inclusive, abundant approach to AI creation requires that value from compute stays distributed across regions.

29:17 Compute productivity should be captured onshore not concentrated in one company or country

Harry Stebbings · Oct 28, 2024

Europe could build a competitive foundation model; the binding constraint is capital, not talent or founders.

Europe has DeepMind and world-class AI centres and resources in Paris, so a team can be stood up, and only one exceptional founder is needed rather than fifty; Mistral shows it's possible but is nowhere near well enough funded.

30:59 Europe has the talent for a frontier model capital is the binding constraint

Matt Clifford · Jul 1, 2024

Countries will have to decide whether they are AI makers or AI takers, and being a taker means shipping dollars to California to use Californian models on Californian compute

Follows from AI being the defining technology of the generation

34:06 Countries must choose between being ai makers or ai takers

Harry Stebbings · Dec 1, 2025

Model sovereignty is coming — national institutions like the German healthcare system will not run on American model providers

38:43 National institutions will not run on foreign model providers

Jonathan Siddharth · Dec 1, 2025 · hedged

Sovereign government AI programmes will require control over the nationality of the humans producing their training data, not just fine-tuning and evals

If you are the German government you presumably want German nationals contributing the SFT and reinforcement learning data

39:00 Sovereignty extends to the nationality of the data labor

Nick Frosst · Sep 1, 2025

Countries should have their own model infrastructure, in the same way it is good for a country to own its power plants

Language models are not that dissimilar from infrastructure like nuclear and hydro power plants, which countries rightly maintain domestically

52:06 Own model infrastructure like national power plants

Jonathan Ross · Sep 29, 2025

Model sovereignty is not enough to win — compute dominates model quality

A model ten times smarter than OpenAI's loses to OpenAI if OpenAI has ten times the compute, because you can't run the AI without compute

38:09 Compute quantity not model sophistication decides competitiveness

Jonathan Ross · Sep 29, 2025

European model sovereignty solves only for removing another party's control over you; it does not solve for having enough compute, so Europe must build enough compute for a player like Mistral to actually compete

Sovereignty removes external control but without sufficient compute you cannot run or scale the AI, so the model can't compete regardless of ownership

38:35 Sovereignty removes control but not compute shortfall mistral needs scale

Harry Stebbings · Jun 19, 2023 · hedged

For certain applications the importance of national datasets is very clear

Different nations have very different data

22:21 Different nations have very different data relevant for certain applications

Martin Mignot · Aug 11, 2025

Europe does need its own LLM provider, and there is a real enterprise market for it, though it will be smaller than OpenAI.

Tech sovereignty is a real constraint: government and quasi-government buyers may want or be required to use local providers, and there is meaningful localization and customization demand.

36:23 Europe needs its own smaller scale llm provider for sovereignty driven enterprise demand