Brussels, 20 March 2025
This Policy Brief analyses innovations in AI models over the past half year and examines the economic implications for AI companies and policymakers.
Author: Bertin Martens
Executive summary
By mid-2024, artificial intelligence large language models (LLMs) were running into diminishing returns to scale in training data and computational capacity. LLM training began to shift away from costly pre-training to cheaper fine-tuning and allowing LLMs to ‘reason’ for longer before replying to questions.
Fine-tuning uses chain-of-thought (CoT) training data that includes questions and the logical steps to reach correct answers. This increases the efficiency of learning for smaller AI models, such as DeepSeek. CoT data can be extracted from large ‘teacher’ LLMs to train small ‘student’ models.
These changes shift the cost structure of AI models from high pre-training costs to lower fine-tuning costs for model developers and more inference costs for users. While smaller models are cheaper to use, a positive AI demand effect is likely to exceed the negative price effect. Price competition between models will increase, resulting in tighter margins for AI firms. Specialised models can still fetch premium prices.
Cheaper LLMs are an opportunity for European Union companies to catch up in building smaller AI models and applications on top of LLMs. Increased demand for AI services will require more investment in computing infrastructure, including in the EU. Investing in large LLMs and the corresponding hyperscale infrastructure is riskier, especially as price competition between models increases.
Knowledge extraction between AI models puts pressure on model developers to protect their investments against free riding by others. It also creates a dilemma for policymakers: should they favour free riding to promote competition and innovation, or should they clamp down and reinforce monopoly rents to stimulate investment in AI models? Past policy will not be an appropriate response in a world that offers vastly expanded opportunities for knowledge pooling and innovation at lower cost.
1. Introduction: enter DeepSeek
DeepSeek is innovative, but in line with model evolution over the past half year – not an unexpected revolution. The start of 2025 was marked by several major announcements related to artificial intelligence. The release of the DeepSeek (2025) AI model on 22 January blew a trillion-dollar hole in the stock market1, on the basis that China’s DeepSeek would substantially undercut American AI giants. DeepSeek was soon followed by many copy-cat small and cheap AI models. Markets concluded somewhat prematurely that DeepSeek broke the AI model scaling law and would undermine the rationale for heavy investment in AI computing infrastructure.
But can small AI models really perform as well as big models, without access to huge quantities of the expensive Nvidia AI processor chips that dominate the sector? Big-tech AI firms were not impressed by this market turbulence and doubled down on their AI infrastructure spending2. Just a week before the DeepSeek release, OpenAI and Oracle announced a $100 billion to $500 billion AI infrastructure investment – dubbed Stargate – to catch up with big-tech firms3. Two weeks later, the European Union announced its own €200 billion AI investment initiative.
This Policy Brief aims to go beyond the DeepSeek hype. It analyses innovations in AI models over the past half year and examines the economic implications for AI companies and policymakers, in particular in the EU. It argues that DeepSeek is innovative, but in line with model evolution over the past half year – not an unexpected revolution. It still fits into the ‘transformer’ generative AI or large language model (LLM) paradigm of the last eight years (Vaswani et al, 2017).
Nevertheless, it has set in motion major changes in AI business models. The cost structure of AI models has shifted away from upstream pre-training costs towards more downstream fine-tuning costs for model developers and more computational ‘reasoning’ or inference costs to respond to the queries of end users. Moreover, AI models are increasingly free riding on, and extracting knowledge from, each other. Price competition between AI models has further increased because smaller models are cheaper to operate. End-user costs per token5 have dropped precipitously, making it more difficult for AI companies to extract revenue and make a profit from AI services. This creates a dilemma for AI developers. Should they protect their models against free riding by others, if at all possible, or should they resolutely go for more innovation to stay ahead of competitors?
All these changes may create opportunities for EU innovators to catch up in the global AI race. EU policymakers should support smaller AI models, built on top of large models, to reap innovation and productivity gains from AI, without risky investment in large foundational models.
Read the complete Policy Brief
Further links
Source – Bruegel Website