Is Europe late to AI?

In 2024, there was skepticism about Europe's ability to develop advanced generative artificial intelligence models (large language models, LLMs) and be at the forefront of applications. The reason is the large investments required to develop AI in terms of computing power, data, and attracting talent. However, on January 20, the Chinese startup DeepSeek introduced the R1 model (not to be confused with the astromech droid or repair robot from Star Wars ), with performance comparable to the best models from Open AI , Meta , or Anthropic , but trained at a much lower cost and with less advanced chips.

DeepSeek challenges the notion that only Nvidia's most advanced chips and software can lead to technological advancement, and that only large platforms can develop cutting-edge models and offer the best applications. In fact, Olivier Blanchard, former chief economist at the IMF, once stated optimistically that the day of R1's launch marked the largest prospective impact on total factor productivity in history.

DeepSeek changes the perception that the AI ​​market is necessarily a winner-takes-all market. These have been the markets where large digital platforms have triumphed, capturing market share and profits, such as Google with its internet search engine, Microsoft with its laptop software, Meta in social media, and Amazon with e-commerce.

In these markets, there are increasing returns to scale and scope (where the same technology can be used for multiple products), network externalities, and self-reinforcing learning economies (platforms with more data further refine their algorithms, which are able to capture more users and data). These platforms constitute an ecosystem that is protected, among other factors, by consumer inertia and switching costs.

There was skepticism about Europe's ability to create advanced AI models, but with ambition, it can be done.

There are differences between the aforementioned markets and the AI ​​market. In traditional digital markets, the fixed costs of generating products are very high, but the variable costs are very low. In AI, fixed costs are high in the initial training phase and in the refinement and adaptation to specific tasks, while operational costs are lower but significant (such as the energy consumption required to answer LLM questions). Furthermore, network externalities in the AI ​​market are more moderate than in traditional digital markets.

As the cost of generative AI falls, it will be used more widely. Greater competition at lower costs will benefit users and increase demand for AI services. However, it will also reduce operators' margins, and therefore the impact on AI companies' profits is ambiguous at first, depending on market expansion.

Read also The dollar: whose problem is it? Xavier Vives

This tells us that the AI ​​market structure could have two levels: the first is oligopolistic, where a few companies have a large market share and provide the most advanced models ( this is what I argued in this section, April 27, 2023 ). Of the current seven or eight companies, three or four could consolidate; now there is fierce competition to see who remains in the market. The final number of competitors will depend on the ratio of fixed costs to market size: the larger it is, the fewer competitors there will be. The second level will consist of companies that provide AI applications in a competitive market.

Europe should aspire to have at least one company at the top level and many at the bottom. To make this possible, we need a set of complementary factors (such as, for example, in Silicon Valley): computing power, a wealth of talent (with unexpected help from the Trump administration), favorable data processing (data protection laws and AI regulation may need to be reformed), better startup financing through the capital markets, and more labor flexibility for startups. With ambition, it can be done.