Every year, US tech giants extract an estimated €19 billion from European economies through AI infrastructure - using European energy, land and resources while avoiding fair compensation. This invisible digital heist overshadows earlier concerns about AI's use of training data and intellectual property.

The mechanism is subtle but powerful: While providing valuable AI services to European businesses, US companies bundle these services with US-owned cloud infrastructure, creating a closed loop that captures value while bypassing European providers and tax systems.

Estimating the extracted value is hard, due to the lack of transparency and a strong practice of secrecy. However, taking a systems perspective can reveal the underlying extraction mechanisms and create a model to estimate the economic impact.

Overview of the systems-level view of money flow in AI services to infrastructure and resources. Source: Leitmotiv

1. 41.17% of European businesses with over 250 employees use AI services. Most services are delivered by US-based frontier model providers, which become deeply integrated into the European economy. Using survey data from a16z on enterprise AI service spending, we estimate European enterprises spend 39 billion a year on AI services.
2. AI service companies spend about half of this revenue on procuring infrastructure to deliver their services to European enterprises. This leads to 19 billion spent on cloud infrastructure. US-based AI service providers bundle their models with US cloud infrastructure, like Anthropic with Amazon Bedrock and GPT with Microsoft Azure.
3. These US cloud providers use European energy infrastructure, land, and natural resources to provide the infrastructure. However, the profits from providing the AI infrastructure and enabling the service delivery, which we estimate at 6.2 billion per year, are not taxed in Europe. European resources are utilized, but do not lead to direct value creation in the economy, creating an outflow of money and resources, an invisible digital heist.
4. The current shift, exemplified by DeepSeek, from computing-intensive training to rather use more computing time to answer questions (inference) is further increasing the opportunity to act. The shift is leading toward even more foreign-owned cloud infrastructure in Europe, because inference needs to happen close to the customer. This is highlighted by the recent news around DeepSeek, which used 10x less computing for training, but is using more computation for inference.

The missed opportunity for Europe is twofold. First, European infrastructure providers are left out of the 19 billion market for AI infrastructure. Second, the profits from the provisioning of infrastructure are not taxed which represents an estimated 1.34 billion (using the average EU corporate tax rate of 21.5%) in missed tax revenue. Given that energy infrastructure, land and natural resources are used, a return for society through taxation should be the minimum.

In this brief, we will explore the mechanisms of resource extraction, direct value creation as well as the market barriers which prevent European infrastructure providers from competing. Furthermore, we will give policy recommendations on how to separate the practice of bundling services and infrastructure and enforce transparency on the flow of infrastructure resources (digital resources) in order for Europe to regain control over its digital economy & infrastructure.

<aside> ⚡

This entire paper and the background materials used for the calculation are available open-source. We welcome improvements and suggestions.

</aside>

How do enterprises import AI services that bundle the required computing infrastructure?

To understand the flow of money, we take a hypothetical example of a large European enterprise with an annual IT budget of 100 million euros. To illustrate our example, we assume 1% of the budget is spent on purchasing AI services. For our example, we assume the enterprise is buying services from a US frontier model provider (such as Anthropic, OpenAI, Google, xAI, and others). All of the service providers operate using a legal entity in the US. Consequently, the 1 million in AI service spent is flowing to the US and becomes an import for Europe.

To deliver the AI service, the provider must install its model on computers with the capability to run them and enable the model to be queried (inference). These computers contain Graphical Processing Units (GPUs) which provide the necessary horsepower to offer fast inference for the customer. All Frontier service providers rent those computers from US cloud providers, bundling their models with the computing resources of the infrastructure provider. The reverse also happens, where a cloud provider does the bundling. This is the case with Anthropic’s Claude model, which is available within Amazon’s cloud platform as Amazon Bedrock.

<aside> ⚡

This is akin to an electric car manufacturer partnering with a utility and offering a car that is bundled with an electricity contract of the utility, locking customers into a single utility providing the electricity.

</aside>

Why are AI service providers not looking for cheaper suppliers and pursuing higher margins from bundling?

An obvious question that comes to mind is: If the cost of computing is bundled with the model to provide the service, wouldn’t the model provider have an incentive to partner with infrastructure providers who offer the lowest price for the compute? This would increase the profit margin of the model provider and thus would make economic sense.

There are two straightforward reasons why this doesn’t happen:

• As the bundling goes both ways, the model provider can tap into the vast pool of existing customers of the cloud infrastructure provider. In fact, as the case of Anthropic, 60-75% of its revenue comes from customers using the model through Amazon AWS. This means the infrastructure provider is both a supplier and a sales channel at the same time. If Anthropic were to choose another cloud supplier, it wouldn’t be unreasonable to assume that Amazon would in turn threaten them to stop offering their model through their cloud platform.
• Large global cloud providers are using their cash-generating abilities - both from their primary businesses (e.g. advertising, e-commerce and licensing) and their cloud business - to place large orders (bets) with chip manufacturers such as Nvidia, Broadcom or TSMC, choking off the supply chain for European cloud infrastructure providers, who can’t match the spending or borrowing capacity of the US counterparts. According to Omdia, Microsoft alone ordered 485.000 Nvidia Hopper chips in 2024. Nvidia’s production capacity is forecasted to produce 3 million Hopper chips in 2025.