Who Owns Artificial Intelligence? Between Concentration and Openness

Epoch AI, a research institute, estimated that training GPT-4 cost about $79 million and training Llama 3.1-405B about $170 million; the latest edition of Stanford's AI Index reports these figures. They are estimates, not published invoices, but they are enough to show that building a frontier model is not a project for a university lab or a start-up.

Why does it lean towards concentration?

A large language model rests on three scarce resources at once: extensive data, expensive computing power, and people who know how to train and tune it. The effect shows in cloud computing, where most of this work is built. In the second quarter of 2026, Amazon, Microsoft and Google together held 63% of the global market, according to Synergy Research Group: 28%, 20% and 15% respectively.

This does not make those companies guilty by size. It means that the question of who owns the tools of knowledge is no longer only technical. Whoever supplies the model that answers millions of questions a day influences what people see and what stays out of view. That deserves watchful attention, not panic.

The other side: open and smaller models

The same picture also contradicts the pessimists. The gap between the best closed model and the best open-weight model on the Chatbot Arena leaderboard narrowed from 8.0% to 1.7% by February 2025. On another track, the smallest model scoring above 60% on the MMLU test in 2022 was PaLM, with 540 billion parameters; in 2024 Microsoft's Phi-3-mini reached the same threshold with 3.8 billion, a 142-fold reduction in two years, according to the same report.

This has practical consequences. A small model can run on a local device without a distant data centre, and a researcher in any city can build on an open model without a large company's permission. Caution is in order, though: one leaderboard and one benchmark do not sum up model quality, and "open weights" does not necessarily mean the training data or method is available.

Why is openness not enough?

Downloading an open model is not the same as being able to train a new one, tune it, or run it at scale. Computing, data and skills remain barriers even when the weights are open. In our assessment, openness lowers the cost of entry but does not remove it.

The OECD's AI Principles, adopted in 2019 and updated in May 2024, point to this dimension when they call on governments to foster an inclusive, AI-enabling ecosystem and to build human capacity and prepare for labour-market transition. It is a matter of policy and education before it is a matter of hardware.

Decentralisation has limits

Decentralisation is not an absolute good. A model that anyone can modify can also be modified by someone who intends harm, so access to tools should come with agreed safety and transparency standards and clear accountability. The aim is a responsible distribution of capability, neither free-for-all nor monopoly.

What can be done?

• Shared computing: make computing power available to researchers and start-ups at fair prices. In the Kingdom, Humain announced in August 2025 that it had begun building two data centres, in Riyadh and Dammam, each with an initial capacity of 100 megawatts; its first facilities were expected to go live in the second quarter of 2026.
• National skills: servers without people who can train, tune and review remain mere consumption.
• Fair competition: access to major infrastructure should not be monopolised on discriminatory terms.
• Public awareness: the more people understand these models' limits and where they err, the less blindly they rely on them.

Taking precautions means asking early: do we want to be tenants of other people's minds, or partners in building them? The answer takes shape in the investment, education and regulatory decisions made today.