Computing Voucher

Anyone who wants to develop a powerful AI program today needs an enormous amount of computing power. This power is delivered by specialized data centers with thousands of expensive chips. They can be rented by the hour, much like a car from a rental company. A computing voucher is a credit for exactly this kind of rental. The recipient therefore does not pay with money, but redeems a voucher allocated to them by a government, the EU, or a large technology company. The voucher is not denominated in euros, but usually in a certain amount of compute hours.

Computing power as the bottleneck of AI development

Access to large computing facilities now determines who can keep up in AI research. Training a single large language model costs, depending on its size, millions in compute time. A university or a young company can rarely afford this out of its own pocket. Buying the appropriate hardware outright is even more expensive and takes months.

Computing vouchers are meant to ease this imbalance. They shift access away from the question of who has the most capital, toward the question of who submits the most convincing idea. Politically, this is also about independence: Europe does not want to remain permanently dependent on American corporations providing the compute time and setting the terms.

For companies that issue such vouchers, there is also a business model behind it. Whoever trains their first model on a particular cloud platform tends to stay there. The voucher then acts like a free test drive that leads to a long-term rental agreement. Experts call this effect lock-in, meaning being stuck with one provider.

From application to billed compute hour

It usually begins with a call for proposals. A funding body describes what it wants to award compute time for, for example medical research or language models in European languages. Interested parties submit an application describing their project. A jury of experts evaluates the applications and distributes the allocations.

The voucher itself is not a piece of paper but an account. It holds a credit, often expressed in GPU hours. A GPU is a specialized chip that performs many computing steps simultaneously and is therefore particularly well suited for AI. If a training run takes ten hours on a hundred such chips, a thousand GPU hours are deducted from the account.

Conditions usually apply. The credit expires after a deadline, it is non-transferable, and the results sometimes have to be published. A common misconception is that a voucher is simply free money. It is tied to a specific facility and a specific purpose. Anyone without the right experts on their team often cannot make meaningful use of the credit at all.

European programs and offers from cloud providers

Several such programs are running within the EU. Through the joint supercomputer network EuroHPC, start-ups and research teams gain access to facilities such as the Jupiter computer in Jülich. National development banks and ministries also allocate quotas. In news reports, they often appear as part of broader AI strategies, alongside figures on funding amounts.

Alongside this, there are the providers' own offerings. Amazon, Microsoft, Google, and Nvidia hand out starter credits to young companies, often through founder programs. Sometimes part of an investment is even paid out directly in compute time instead of money. Such deals regularly spark discussion, because it remains unclear how much real capital actually changes hands.

For students or hobbyist developers, there are smaller versions. Platforms like Google Colab or Kaggle provide limited free compute time. It’s the same idea on a small scale: getting started should not fail because of the price of hardware.