Gutting AI Safeguards Won’t Help Europe Compete

This notion ignores basic economics. Foundation models form the fundamental building blocks for various AI applications, and their developers in Europe promise that gutting the AI Act will allow them to catch up to the big players on the world stage. The catch-up most likely won’t happen, but the watering down of the regulation will have real effects in reducing the ability of European policy makers to protect their citizens from harm. 

Dispassionate economic analysis tells a much fuller story than appeals to national pride. In reality, the most capable foundation models are already well-positioned to become natural monopolies, making the need for regulation even more dire. Many factors will determine whether Europe can begin producing foundation models of the same caliber as its U.S. rivals. Refraining from imposing any safety obligations on these developers is not one of them.

There are three main costs that go into building foundation models, and that developers and European policy makers need to pay attention to: the fixed costs associated with model training, the additional costs linked to customizing the model for individual clients, and the variable cost attached to actually operating the model. 

Training foundation models is expensive. The fixed costs largely stem from acquiring enough computational power for pretraining, sourcing training datasets, attracting talent, and other infrastructure expenses. Ultimately, the bill can add up to tens or hundreds of millions of dollars, with soaring computational-power prices and competition for key talent forming the largest chunk. For example, training GPT-4 cost an astonishing $100 million, according to OpenAI CEO Sam Altman. As a result of these fixed costs, foundation models exhibit strong economies of scale. 

Once foundation models have been trained, they need significantly fewer resources to operate. This means that the biggest players on the world market who can afford to scale up are then able to reap the benefits of much lower average costs. Consequently, we observe significant barriers to entry for new companies, which are plagued by the need to achieve sufficient scale in order to be cost competitive. As such, early entrants into the playing field tend to gain first-mover advantages that make the market hard to challenge. 

Acquiring sufficient data poses one of the greatest challenges of the training stage. Currently, the most data-intensive language model is Google’s FLAN, trained on 1.87 trillion words. Foundation models have a voracious need for data, but the public internet only contains around 100 trillion words total, which makes proprietary datasets extremely valuable for maintaining a competitive edge. At the moment, U.S. big tech companies hold this advantage, controlling most data generated online across platforms, searches, emails, images, videos, and other documents. Foundation model providers that want to get their hands on this invaluable proprietary data will face a strong incentive to vertically integrate with incumbents. 

Another crucial variable is computational power or “compute,” which, so far, appears to scale along with model performance. To make their models more capable, developers must add ever more compute, but this is costly to acquire, limited in supply, and has technical constraints. To illustrate,
Microsoft
invested $13 billion in OpenAI, most of which OpenAI used to pay for training compute. 

The last key ingredient for becoming a big player in the foundation model market is talent. Currently, demand far exceeds supply for researchers and engineers capable of developing the best foundation models and the necessary server infrastructure. Foundation model producers are even willing to hire engineers without AI experience and train them in-house. A 2021 McKinsey survey found that 32% of organizations reported it “very difficult” to hire AI data scientists. New entrants would likely need to poach talent from market leaders, but workers are unlikely to switch without even more lucrative compensation offers. 

These factors are what plague the European foundation model-market—not regulation. Continental European players have not invested as much in compute, data, and talent  as incumbents across the Atlantic. Regulation of advanced AI models would not create a competitive disadvantage for European companies since it would apply to all players who produce models above a certain size and capability. If anything, it would hit American AI companies harder than European ones since they are investing in larger models. Furthermore, it may actually be easier for European companies to navigate European regulations than for foreigners. Finally, ensuring that the foundation models produced by the world leaders are safe would empower Europe’s thousands of smaller and medium-sized enterprises, which are building innovative applications on top of these powerful foundation models. 

What should policymakers do so that AI development benefits consumers, aside from ensuring the safety of advanced AI models? First, they should aim to ensure contestability for the foundation-model market in Europe. This may mean focusing on merger reviews, any privileged access to models, acquisitions made by foundation-model companies, and the existence of predatory pricing. As a last resort, foundation model companies may need to be broken up. 

Data governance rules can also help prevent undesirable concentration in foundation model markets by limiting privacy violations and exclusive data accumulation, especially as models shift from using open web data to proprietary datasets. In addition, authorities should ensure talent can flow freely by discouraging or banning noncompete clauses and scrutinizing other restrictive contract terms. As their economic role grows, the most capable model providers may need to be regulated as public utilities with outsized societal risks. 

If Europe wants to benefit from recent advances in AI, it must first impose strong regulations that correspond with European values. This will allow it to harness the productivity benefits of existing foundation models. Ultimately, watering down the AI Act requirements on foundation models would do nothing to improve EU competitiveness and would lower global standards, to everyone’s detriment.

Guest commentaries like this one are written by authors outside the Barron’s and MarketWatch newsroom. They reflect the perspective and opinions of the authors. Submit commentary proposals and other feedback to [email protected].