When you ask a chatbot a question, the answer comes from systems owned and shaped by a small number of organizations. Who they are, what they need and what pressures they face matter for everyone, because control over AI is becoming control over a growing share of how people find information, do their work and make decisions.

The ingredients of power

Building a leading model takes four things in large amounts: computing power, data, skilled people and money. Computing power, usually called compute, means huge numbers of specialized chips running for months in data centers. Estimates of the cost of training a single frontier model have risen from millions of dollars to hundreds of millions, and the largest technology companies now spend tens of billions of dollars a year on data centers. Few organizations can pay that, which concentrates the most capable systems in a handful of firms and in the countries that host them.

Chokepoints in chips

The supply chain for advanced chips is unusually narrow. The American firm Nvidia designs most of the chips used to train AI. Taiwan's TSMC manufactures most of the world's most advanced chips. And the Dutch company ASML is the only maker of the extreme ultraviolet lithography machines needed for the most advanced chips. Each is a point where one company, or one government's policy, can affect the whole field. With this in mind, the United States in October 2022 imposed export controls on advanced AI chips and chipmaking tools bound for China, and it has revised them repeatedly since. Supporters see a security necessity. Critics warn of a technology race that splits the world and pushes China to build its own supply chain.

Open or closed

Companies differ on whether to release their models. A closed model can be reached only through the company's own app or service. An open-weight model is one whose trained parameters are published for anyone to download and run. Meta's Llama models, the French firm Mistral's models and the Chinese company DeepSeek's R1, released in January 2025, are prominent examples. Supporters say openness spreads power beyond a few firms, lets outside researchers inspect the systems and allows people to run models privately on their own machines. Critics answer that once weights are released they cannot be recalled or kept from misuse, so anyone can strip out the safety limits. Both sides agree the choice shapes who holds power.

Hidden labor and real costs

AI can look like pure software, but people and resources sit behind it. Training relies on human workers who label data and rate outputs, many of them in lower-income countries. In 2023 TIME magazine reported that workers in Kenya were paid under 2 dollars an hour to label disturbing content, including descriptions of violence and abuse, so that a chatbot could learn to avoid producing it. Some described lasting distress.

Data centers also use large amounts of electricity and, in some places, water for cooling. The International Energy Agency estimated that in 2024 data centers used about 1.5 percent of the world's electricity, a share it expects to grow quickly, with AI a major reason. The burden falls unevenly, on the towns where data centers cluster and on their local power grids.

Then there is the question of whose work trained the models. Authors, artists, news organizations and others have sued AI companies for training on their work without permission or payment. The New York Times sued OpenAI and Microsoft in December 2023, and in 2025 Anthropic agreed to pay about 1.5 billion dollars to settle a suit by authors over books it had downloaded from pirate websites. Courts in several countries are still deciding when training on such material is lawful.

Living on someone else's platform

For ordinary users, control shows up in quieter ways. A company can retrain a model, change its rules, raise its price or shut it down, as the flattering update described in Level 3 showed. Relying on a single provider matters because it can change or withdraw the model, its rules or its price at any time. A school, clinic or city that builds its services on one company's model inherits that company's decisions. Some governments are responding by funding public research computing, so universities and smaller groups are not wholly dependent on a few firms.

Worked example: the city's assistant

A mid-sized city wants an AI assistant to answer residents' questions about permits and services. Vendor A offers a closed model with a polished app and a low first-year price. Vendor B offers to run an open-weight model on servers the city controls, at a higher setup cost. The council's technology officer lays out the trade. With A, the city depends on terms the vendor can change, and residents' questions go to the vendor's servers. With B, the city keeps its data and can switch models later, but must hire staff to maintain and test the system. The council chooses B for anything involving personal data, and writes into every contract a right to audit the system's answers and to take the city's data along if it ever leaves.