Open-Source AI: A Question of Who Holds the Power

On Monday 27 January 2025 Nvidia's shares fell 17% and its market value dropped by about $589 billion, the largest one-day fall in value for a US company on record. More than a trillion dollars was wiped from the technology sector. The spark was an open Chinese model from DeepSeek. It was not merely a market disturbance; it was evidence that whoever has access to a model can change the balance of an entire industry.

What does "open" mean?

On 28 October 2024 the Open Source Initiative issued its first definition of open-source AI. It requires a system to grant four freedoms: to use it for any purpose without asking permission, to study it, to modify it, and to share it. It also requires the "preferred form for modification" to be available: sufficient information about the training data, the complete code, and the parameters.

By that yardstick, publishing weights for download is not enough to earn the label. Take Meta's Llama 3.1 licence: it requires anyone whose monthly users exceed 700 million to request a licence from the company, requires the phrase "Built with Llama" to be displayed, and requires "Llama" at the start of the name of a model improved with its outputs. These restrictions sit uneasily with the definition's freedom to use a system "for any purpose without asking permission". The popular term is wider than the standard.

What happened in January?

In its technical report on the V3 model, DeepSeek stated that full training took 2.788 million H800 GPU hours and that the cost came to $5.576 million on the assumption of a rental price of $2 per GPU hour, and that this figure "excludes the costs associated with prior research and ablation experiments on architectures, algorithms, or data". The number in the headlines therefore measures one run, not a whole project. Even with that caveat, an openly available model reached serious competitiveness at a modest declared training cost relative to comparable projects.

The law distinguishes open from risky

Article 53(2) of the EU AI Act exempts providers of general-purpose models released under a free and open-source licence, with public weights, architecture and usage information, from the technical documentation duty and from informing downstream developers. The exemption does not cover the copyright policy or the summary of training content. And it falls away entirely for models with systemic risk. The law is young, and how it applies to real cases will become clear through practice and regulatory interpretation.

One reason for caution is that releasing weights cuts both ways. In a 2024 paper Sayash Kapoor and colleagues argue that open models strengthen innovation, competition, the distribution of decision-making power and transparency, but that their customisability and the difficulty of monitoring their use contribute to risks as well as benefits. They propose measuring the "marginal risk" that open models add beyond what existing technologies already allow, and conclude that current research is not sufficient to estimate it, with much of the disagreement coming from differing assumptions.

Opportunity and capability

An open model gives the Arab researcher a ready starting point and the chance to feed it with their own language and heritage, a point treated in the article "Arabic Artificial Intelligence: Do We Make or Consume? Sovereignty and Arabic Language Models". But the file is no substitute for capability: the compute, trustworthy data and human skills needed to run and develop large models matter as much as the model itself. That is our reading, not a direct research finding.

How should we read announcements of "open"?

Ask four questions. Were only the weights released, or the code and data information too? What do the licence terms restrict in use and in commerce? What does it actually cost to run? And who bears responsibility if it is misused? Whoever can answer them has moved from slogan to judgement.