Five firms own over 70% of the world's AI computing power, the UN rights chief says
Volker Türk named no companies, but Epoch AI's estimate points to Amazon, Google, Meta, Microsoft and Oracle at 71% — while the big AI labs rent most of the compute they use.

The News
Five technology companies own more than 70% of the world's AI computing power, United Nations High Commissioner for Human Rights Volker Türk said on Thursday, November 19, at the RightsX summit in Geneva. Türk said artificial intelligence is governed in a way he called almost medieval — a small number of powerful players setting the rules, with little oversight from the rest of the world. Neither Türk nor the AFP news agency report of his remarks named the companies, or cited a source for the figure.
The closest published match for the figure comes from Epoch AI, a research group that tracks AI computing. Epoch estimates that Amazon, Google, Meta, Microsoft and Oracle owned 71% of cumulative AI compute at the end of 2025 — measured in H100-equivalents, a way of counting computing power in units of Nvidia's H100 chip. Google led, with about 5 million H100-equivalents, roughly 25% of the world's total. To be clear: those five names are Epoch AI's identification, not Türk's. The UN rights chief named no companies.
The distinction matters because owning compute and using it are different things. The frontier AI labs — OpenAI and Anthropic — rent most of the computing power they run, mainly from cloud providers. Epoch's separate usage research estimated that OpenAI, Anthropic and xAI together made up about 20% to 30% of operational AI compute at the end of 2025. Türk wrote to states and AI developers in September about the concentration of power; the AFP report does not say which companies received his letter, or whether any replied.
Five firms hold the keys to the world's AI computing power — and even the labs building the models rent the keys.
Why It Matters
The remarks matter because they put a number on a concentration of power that is usually described only in vague terms. More than 70% of the world's AI computing capacity in five hands means that a handful of companies decide who gets to train the biggest AI models, at what price, and on what terms. For a UN human rights chief, the concern is that decisions with public consequences — what AI can do, who can build it, who profits — are being made by companies, not by governments or the public.
The Epoch AI figures give the claim its shape. Counting chips owned, rather than chips used, Google comes out on top with about a quarter of the world's total, and the five named companies together hold 71%. But the rental market complicates the picture: OpenAI, Anthropic and xAI do not own most of what they run, yet together they account for 20% to 30% of the computing actually in use. Power in AI is not just about who owns the hardware — it is about who controls access to it.
The limits of the claim matter too. Türk named no companies and cited no source, so the 70% figure cannot be checked against his own evidence — the Epoch estimate is the closest match, and it is a different measurement, from a different date, counting owned chips rather than all computing capacity. And a speech is not a policy: Türk's September letter to states and developers has produced no public reply that the reporting could confirm. The concentration is real; the precise number is an estimate, and it should be read as one.
Analyst's Take
Our call is Mixed. The concentration Türk describes is real and important — five companies holding over 70% of AI computing power is a fact worth debating, and the rental model means even the most famous AI labs depend on their landlords. But the headline number is unverified as Türk stated it, and the company names come from a research estimate, not from the UN. Treat the warning seriously and the number carefully.
What to watch next is whether Türk's letter gets a public answer, and whether any government turns the concentration argument into policy — for example, rules on access to computing power, or support for smaller labs. Reader poll: Should AI compute be less concentrated? Vote: Yes, break it up, No, it is fine, Regulate, don't break up.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is AI computing power?
AI computing power, or compute, is the processing capacity used to train and run AI models — mostly the work of specialised chips in datacenters. The more compute a company controls, the bigger the AI models it can build and the more customers it can serve.
What is an H100-equivalent?
An H100-equivalent is a unit for counting computing power: one unit equals the power of one Nvidia H100 chip, a widely used AI chip. If a company owns 5 million H100-equivalents, it owns computing power equal to 5 million of those chips, even if its actual hardware is a mix of different chips.
What does cumulative AI compute mean?
Cumulative AI compute is the total computing power a company has built up over time — all the AI chips it owns, added together — rather than the computing power it uses in a single day or month. Epoch AI's 71% figure counts chips owned, not chips in use.
What is a cloud provider?
A cloud provider is a company that owns huge datacenters and rents computing power to others over the internet. Amazon, Google and Microsoft are the biggest cloud providers, and AI labs like OpenAI and Anthropic rent most of their computing power from them instead of owning it.
👁
Source: TechTimes