Against Comparative Advantage
Should middle powers build foundation models?
TL;DR
Middle powers like Germany and Japan have announced plans to build sovereign foundation models. These plans are usually dismissed. Middle powers lack the compute, talent and capital to compete at the frontier. They should accept the foundation model race is over.
This post makes the strong case for sovereign foundation models. Building sovereign foundation models is a logical conclusion when middle powers need access to frontier AI and cannot rely on US closed-source systems or open-source alternatives.
Sovereign foundation models should not compete at the frontier. Accessing frontier AI built in the US should remain the priority. But sovereign foundation models can provide an always-on fallback in domains where marginal differences in capability matter.
Middle powers must think in terms of dynamic comparative advantage as they bargain for access to frontier AI built overseas. This means investing in areas of potential growth as well as areas of existing strength. The stakes are too high for middle powers to simply double down.
POSCO Gwangyang Steelworks, South Korea. Source.
Introduction
In 1969, the World Bank concluded that South Korea had no comparative advantage in steel production. More specifically, the Bank argued that ‘in view of the obvious comparative advantage, it would be desirable to shift the emphasis from iron and steel to machinery’.
South Korea lacked iron ore reserves, and was located far from the main sources of supply. Steel production was capital intensive, and required a skilled workforce. Domestic demand was limited, and the largest nearby market was served by the world’s most efficient producer, Japan’s Nippon Steel.
The Bank made a compelling case. It caused the US Export-Import Bank to decline a loan to the Pohang Iron and Steel Company (POSCO), a new integrated steel mill in South Korea. POSCO was central to South Korean President Park’s vision for the country as a ‘steel power’. Now the project was at risk.
But POSCO defied comparative advantage. In 1981, the World Bank described POSCO as ‘the world’s most efficient producer of steel’. In 1985, POSCO had lower production costs than Nippon Steel, two thirds lower than US Steel. In 1986, POSCO signed a joint venture to modernize the US Steel plant in Pittsburg, California.
Several governments have announced plans to build sovereign foundation models. Japan has the National Foundation Model Initiative. Germany has the Sovereign Open Source Foundation Model Initiative. Saudi Arabia has HUMAIN. The list goes on.
These plans are often dismissed. Countries other than the US and China lack the compute, talent and capital to build frontier AI. Middle powers like Germany and Japan should accept the foundation model race is over, and invest in areas of existing comparative advantage.
I am sympathetic to these arguments. I first joined the UK Government as part of the Foundation Model Taskforce: we decided that BritGPT was not a good use of taxpayer money. But I worry they risk sounding like the World Bank before South Korea’s steel miracle.
So this short post is an exercise. I want to try and make the strong case for sovereign foundation models. What is the alternative for middle powers?
Access Matters
Middle powers need access to frontier AI. This argument has been made in depth elsewhere, so I will just summarize the main points.
Foundation models are infrastructure. They are non-rivalrous within compute constraints, meaning many people can benefit from access simultaneously. They enable a wide range of productive activity, acting as an input to private, public and social goods. Failing to secure access to infrastructural inputs means falling behind. As Dean Ball and Anton Leicht argue, there have been no high-income countries without abundant electricity.
But this is an argument for access to foundation models, not frontier AI. Middle powers need access to frontier AI in competitive domains where relative capabilities matter. Domains where small gaps create exploitable vulnerabilities or compounding advantages over time.
This includes national security, where middle powers cannot accept overmatch by adversaries with access to frontier AI. It also includes economic competition, where foreign firms with access to frontier AI could beat incumbents without meaningful moats. Access to frontier AI will not decide every market, and adoption also matters. But it could prove decisive in winner-take-most sectors where marginal capability leads compound fast.
President Park was inspired by the Meiji Restoration, Japan’s response to the arrival of American steam-powered warships its own fleet could not contest. Without access to frontier AI, middle powers risk setting sail against steam.
Access Options
Middle powers have two main options to access frontier AI. They can rely on closed-source systems developed in the US, or build on the best open-source models. Both options have risks.
Option 1: Rely on the US.
Access to frontier AI developed in the US is not guaranteed. As Anton Leicht argued recently, structural forces push against broad access to frontier AI built in the US.
Safety risks are encouraging American AI companies to implement managed access regimes. This includes Project Glasswing from Anthropic and Daybreak from OpenAI, but is not restricted to frontier AI systems with significant cyber capabilities: OpenAI also released its life sciences system, GPT-Rosalind, with managed access.
Distillation attacks help to align the incentives of American AI companies with safety, but also encourage more restricted access to frontier AI. One of the steps Anthropic has taken to prevent distillation attacks is to crack down on educational accounts, and require stricter verification for research organisations and startups.
Geopolitics also favours restricting access. So long as there is not perfect alignment between US and middle power objectives, there will be an incentive to limit access to frontier AI. This is not new in international relations. Defence programs like the F-35 use tiered access, depending on geopolitical alignment and investment.
Access also promises to become a source of leverage. There is already evidence that the US Government is using access as a tool of foreign policy: it has been reported that Treasury Secretary Scott Bessent used a trip to Japan to inform the Japanese Government that selected Japanese banks will get access to Anthropic’s Mythos.
Mythos is unusual in being so binary. Access is a spectrum: there are many forms of access that could be offered or withdrawn in future. Managed access regimes implemented for safety will allow more fine-grained restrictions, adding more rungs to the escalation ladder between ‘cut off’ and ‘no constraints’. Token budgets and latency already differ across users.
Middle powers cannot compel access to frontier AI built in the US. Some have argued that middle powers can secure access by controlling supply chain chokepoints, such as ASML. As I have written previously, middle powers have leverage only if they can endure vertical and horizontal escalation by the US if they limit access to a chokepoint. This is a difficult condition for any middle power to meet.
Option 2: Rely on open-source.
If access to frontier AI developed in the US is not guaranteed, middle powers could rely on open-source. Middle powers already build on open-source: Llama 2 is the base for HUMAIN’s Arabic language models. But relying on open-source is not a viable long-term alternative.
There is no guarantee that frontier AI will continue to be open-sourced. Open-source models can be modified and shared without oversight. Governments and AI companies may decide that frontier AI should therefore not be open-source above a certain capability threshold. Meta recently confirmed that Muse Spark was not open-sourced in part because it triggered internal safety thresholds.
Economics also discourages open-source. Frontier models now cost billions to develop. Open-sourcing makes it difficult to recoup this investment, and comes at the cost of products that capture real revenue and profits. Eventually, economics will overcome any ideological commitment to open-source. The latest two models in Alibaba’s much-loved Qwen series have been released closed-source.
But let’s assume Chinese companies continue to open-source frontier AI, to undercut the leading US firms. Maybe state support allows company leaders to maintain an ideological commitment to open-source, and defy economics. After all, Premier Li cited building a ‘vibrant open-source ecosystem’ as an objective at the Two Sessions.
Middle powers cannot rely on open-source models from China. The domains where frontier AI matters most are the domains where middle powers can least afford to rely on China. Chinese models pose well documented security risks, including alignment with CCP values. Chinese models could also pose sleeper agent risks: despite progress in identifying backdoors, model poisoning remains an unsolved research problem.
So let’s relax another constraint, and assume American AI companies continue to open-source. Maybe the US Government supports open-source so long as there are open-source Chinese models, to win the diffusion race. After all, the US AI Action Plan recognized the ‘geostrategic’ value of open-source models.
Open-source will continue to lag the closed–source frontier. American AI companies will not release open-source models that compete with their closed-source products. Gemma 4 lags Gemini 3.1 Pro, and GPT-OSS lags GPT5.5. Open-source groups like Arcee and Reflection will meanwhile struggle to keep up. The current paradigm privileges access to deployment data, which can serve as reward signal for online RL, and high-quality RL environments, which the leading companies can afford to buy under exclusivity agreements.
Nvidia is perhaps the only company that could provide the compute and capital to close the gap. The Nemotron coalition announced at GTC 2026 is a step in this direction. But there is likely a ceiling on Nvidia’s support: it has an incentive to commodify the model layer, but must also defend its core business and protect customer relationships with AI companies.
Sovereign Foundation Models
It is harder to dismiss sovereign foundation models in context. Middle powers need access to frontier AI, and cannot rely on closed-source US systems or open-source alternatives. Building sovereign foundation models is a logical conclusion from a bad set of options.
So what should middle power projects to build sovereign foundation models look like? This question deserves a full post, so I will just sketch an outline.
Organisational Design
Middle power projects could be built around a sovereign AI lab. Each lab would be state-backed, but directly supported by domestic companies with an interest in staying competitive without relying on American AI companies. Japan’s National Foundation Model Initiative is supported by SoftBank, Sony, and Honda, for example.
The sovereign AI lab would aggregate talent and compute from across the ecosystem, capitalising on the consolidation that takes place as domestic AI companies struggle to compete with the leading US firms. This kind of consolidation was visible in Cohere’s acquisition of Aleph Alpha earlier this year.
Sovereign AI labs could be networked together, linked by informal agreements that create options for closer collaboration in future. Coordination costs and difficult questions around benefit sharing make it simpler to start national, not multinational, but the sovereign AI labs could ultimately be consolidated into a single middle power project.
Role and Function
The main function of a sovereign AI lab would be to build foundation models. The lab would not compete with US closed-source systems, but extend the open-source frontier. Accessing frontier AI developed in the US would remain the priority, but a sovereign AI lab would provide an always-on fallback for scenarios where access is severely limited.
Foundation models developed by sovereign AI labs would serve as engineering platforms, reducing R&D costs for the wider ecosystem. Domestic companies could provide private data and build specialized systems on the sovereign base model. The national security community could similarly build systems on the sovereign base model. In the extreme, sovereign AI labs could take on engineering challenges like building models on distributed or heterogeneous compute, similar to DeepSeek de-risking Ascend for the Chinese ecosystem.
Sovereign AI labs would have several benefits. They would increase the quality of the next-best alternative to frontier AI developed in the US, increasing middle power leverage and eroding the pricing power of American AI companies. If foundation models developed by sovereign AI labs narrowed the gap between the open and closed frontier, they could see global adoption, and encourage American companies to open-source more advanced systems. Specialized systems built on the sovereign base model could find practical utility as tools for the US closed-source systems, becoming core components of multi-agent setups.
Outlook
In 1991, the World Bank wrote a retrospective on its analysis of South Korea. The Bank acknowledged it did not expect ‘success in changing South Korea’s comparative advantage’. But ultimately ‘dynamic comparative advantage was realized’.
There are many reasons not to build sovereign foundation models. Profit incentives may encourage broad access to frontier AI developed in the US. The global race for diffusion could narrow the gap between the open and closed frontier. Securing the compute, talent and capital to extend the open-source frontier will be difficult. But building sovereign foundation models might be essential all the same.
Middle powers without guaranteed access to frontier AI must think in terms of ‘dynamic comparative advantage’. Dynamic comparative advantage acknowledges that comparative advantage is endogenous, created by choices over time rather than simply inherited. It focuses on areas of potential growth as well as areas of existing strength. In other words, middle powers need to make bets The stakes are too high to play it safe.
Many thanks to Herbie Bradley, Flynn Devine and Jack Wiseman for feedback on earlier versions of this post. All opinions and errors are my own.