You're right to think that companies implementing Artificial Intelligence mostly import and export services, so at first glance, tariffs shouldn’t directly impact production or maintenance costs.
However, the situation is far more complex. Just because a GPU is designed in the United States (USA) doesn't mean it's also manufactured there. In fact, most GPUs and TPUs designed in the U.S. are manufactured in Taiwan and South Korea, as they lead in advanced lithographic processes. Both the packaging and shipping of these components also originate from Asia.
USA data centers using this hardware will be directly impacted by these measures. Ultimately, the consumer of services offered by companies like OpenAI, Meta, or Anthropic will be the one paying adjusted prices that reflect the new reality. The more expensive it becomes for providers to acquire these products, the more expensive their services will become in order to offset the impact of tariffs.
On another front, becoming dependent on USA AI providers could indeed affect us — if we're not prepared. One way Europe can respond more assertively to such policies is through its tech companies. In fact, the European Commission is preparing fines for tech firms that fail to comply with the Digital Markets Act (DMA). So, while this new norm of economic instability between countries might not directly affect us yet, we can expect retaliatory measures from U.S. tech companies at some point, and we must be ready for that.
Here are some strategies that governments and tech companies should follow to reduce dependency on major tech players:
Diversifying Suppliers
European alternatives to AI model providers can be explored — for example, Mistral AI, which offers open-source models like Mistral Large with 123 billion parameters and competes with models from OpenAI, Claude, and Google on several benchmarks. Another initiative is OpenEuroLLM, a European project aiming to democratize access to high-performance AI by enabling development and deployment of LLMs within the European framework.
Using Open-Source Models Hosted on Private Clouds
To eliminate reliance on external providers, open-source models can be deployed on private servers — for example, Meta's LLaMA models, DeepSeek’s R1, or the aforementioned Mistral Large by Mistral AI. In Spain, and as part of a European Union initiative, the government is training its own language models — such as the ALIA family (40 billion parameters) and Salamandra (with 2 billion and 7 billion variants) — to make them publicly available. This aims to promote open, transparent language models and boost the use of Spanish and regional languages in AI development and deployment.
Self-hosted AI
This marks a major shift in how organizations can integrate AI into their operations. Using models that don't require massive data centers and can run on local or household hardware offers several strategic benefits. First, a considerable cost reduction, as there’s no dependence on third-party services — many of which are managed by foreign providers like OpenAI or Anthropic — eliminating usage fees and avoiding policy restrictions. Second, substantial improvement in privacy and technological sovereignty: full control over the model reduces the risk of sensitive data leaks or exposure. And third, the maturity of the open-source ecosystem, which has produced highly competitive models when compared to proprietary alternatives like GPT-4o or Claude Sonnet 3.7.
Many of these models, such as Mistral AI's Mixtral, also introduce real technical innovations. In this case, they use Mixture of Experts (MoE) architectures — think of them as a team of specialists inside the model. When a query comes in, the system identifies the topic and activates only the relevant "experts," keeping the rest idle. This not only improves efficiency by reducing resource usage but also enables more targeted and relevant responses.
In summary, self-hosted AI not only represents autonomy and efficiency, but also a viable path toward more sustainable, secure, and adaptable solutions.
Implementing Small Language Models (SLMs)
To reduce the need for large computing capacities (and thus the costs associated with energy consumption), companies should begin exploring the use of SLMs (Small Language Models). These are smaller versions, often distilled from larger models, with fewer parameters or “neurons,” making them significantly lighter in terms of resource usage and more efficient to run. While they don’t reach the reasoning or generalization levels of large-scale models, SLMs are highly effective for specific tasks — particularly when deep contextual understanding isn't needed. They’re ideal for focused business applications, such as information extraction, text classification, or internal assistants in specialized domains.
Avoid Generic Solutions and Use AI for Specialized Applications
Large language models have become popular because of their ease of use and general knowledge across a wide range of topics. However, in the business sector, they are increasingly seen as tools for writing emails faster or digesting information quickly — and not much more.
The real potential of AI emerges when it's applied to specific processes within a company’s value creation. When AI is introduced at that level, it truly boosts competitiveness and is applied strategically within operations. Often, this doesn’t require massive external models; on the contrary, being specialized solutions, they can be optimized to reduce costs and external dependencies.
In the End…
Anticipation is key. Diversifying suppliers, using open-source models, and implementing AI solutions tailored to your value creation process — these are critical steps. Being ready to adapt to new trade regulations will enable European tech companies to continue operating with stability and efficiency. If your organization wants to prepare for greater autonomy, get in touch with us.