The Conversion Gap: How China, Japan and South Korea Turn AI Inputs into AI Capability, 2017–2025
This paper argues that national AI sovereignty depends more on a state's policy capacity to efficiently convert resources into capability than on resource accumulation alone, revealing that Japan's lag behind China and South Korea stems from systemic inefficiencies in transforming knowledge and compute into AI models rather than a lack of inputs.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the race to build artificial intelligence, nations are often judged by what they can gather. They count the researchers in their labs, the billions of dollars poured into projects, and the massive banks of computers humming in data centers. It is a logic of accumulation: if you have more of the raw materials, you should build the best machine. For years, this has been the standard way to compare countries like China, Japan, and South Korea, all of which have poured immense resources into becoming leaders in this technology. But having the ingredients does not guarantee a finished meal. A country might possess a vast stock of talent and computing power yet fail to turn those resources into working, intelligent systems that change the world. The real question is not just how much a nation has, but how efficiently it can transform what it has into actual capability.
This is the puzzle that Alfredo Merlet set out to solve in a new study examining the years 2017 through 2025. He looked past the simple totals of money and people to measure something more subtle: the conversion rate. Imagine a factory where the goal is to turn raw ore into finished cars. One factory might buy twice as much ore as another, but if it produces fewer cars, it is clearly less efficient at the job. Merlet applied this same logic to artificial intelligence, breaking the process down into three distinct stages. First is invention, where researchers create new ideas and publish papers. Second is industrialization, where those ideas are turned into actual software models that can be used. Third is operationalization, where those models are adopted by regular people and businesses. By tracking how much output came from each unit of input at every stage, the study revealed a surprising story about why some nations are pulling ahead while others, despite having more resources, are falling behind.
The results paint a picture of three very different paths taken by East Asian neighbors. China, starting with a massive workforce and a government that sets clear, aggressive targets, managed to accelerate its progress dramatically. By 2024 and 2025, Chinese developers were releasing hundreds of new AI models, moving from a follower to a peer of the United States. They achieved this not just by spending more, but by coordinating their efforts so that research quickly turned into deployable products. South Korea, despite having fewer researchers and less computing power than Japan, achieved the highest efficiency of the group. The country managed to produce 87 AI models, including 23 that were considered notable or groundbreaking, a number that far exceeded its larger neighbor. This success was driven by a highly coordinated relationship between the government and its major corporations, allowing them to focus their resources intensely on specific technologies and get results quickly.
Japan, however, presents a stark and confusing contrast. On paper, it should have been the most successful. It has a larger research workforce, more computing power, and a longer history of industrial excellence than South Korea. Yet, between 2017 and 2025, Japanese organizations released only 52 AI models, and a mere six of those were considered notable. In the final two years of the study, no Japanese organization released a single model that the researchers classified as notable. The data shows that Japan is not failing because it lacks money or scientists. In fact, when looking at how well Japanese researchers turn their work into patents, they perform reasonably well. The breakdown happens later in the chain. The gap appears when Japanese researchers try to turn their published papers into working models, and again when they try to use their massive computer power to build those models. They have the ingredients, but the kitchen is not producing the dish.
The study suggests that the reason for this gap lies not in the technology itself, but in how the government manages the process. The researchers compared the countries' performance against a measure of "policy capacity," which is essentially the government's ability to set a clear direction, coordinate different groups, and move quickly to support new industries. South Korea scored extremely high on this measure, reflecting a government that is assertive and organized in its approach to AI. Japan, by contrast, scored significantly lower. For most of the period studied, Japan relied on gentle guidance and voluntary guidelines rather than binding targets or concentrated programs. It was only in 2025 that Japan established a dedicated headquarters and a new law to push AI forward, a move that the study interprets as a late recognition of the problem. The evidence indicates that having the best research and the most powerful computers is not enough if the state cannot effectively organize them to produce results.
This finding challenges the common assumption that AI dominance is simply a matter of who can spend the most money or hire the most experts. The study shows that sovereignty over artificial intelligence is not just about owning the tools; it is about the ability to govern how those tools are used. A country can invest heavily in research and infrastructure and still produce very little if the system for turning those investments into real-world applications is weak. For emerging economies watching this race, the lesson is clear: building a strong AI future may depend less on matching the massive budgets of global leaders and more on building the administrative capacity to coordinate resources effectively. The difference between a nation that funds AI and one that truly builds it may come down to the quality of its planning and the speed of its execution, rather than the size of its bank account.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.