How Concentrated Is AI? The Numbers Are Striking

1/ People say AI is concentrated. But how concentrated, exactly? The numbers paint a stark picture across every layer of the stack. Let's go through them. 🧵

2/ COMPUTE: Three cloud providers — AWS, Azure, Google Cloud — control roughly 67% of global cloud infrastructure. If you're training or deploying AI at scale, you almost certainly rent from one of them.

3/ GPU SUPPLY: Nvidia holds approximately 80-90% of the AI accelerator market. When Nvidia sneezes, the entire AI industry catches cold. TSMC fabricates nearly all of these chips. Two chokepoints for the whole industry.

4/ FRONTIER MODELS: As of early 2026, only about 5 organizations have trained models at the true frontier (GPT-4 class and above): OpenAI, Google DeepMind, Anthropic, Meta, and xAI. That's it. Five labs for the most powerful AI on Earth.

5/ FUNDING: In 2024, roughly 75% of all AI startup funding in the US went to just three companies: OpenAI, Anthropic, and xAI. The concentration of AI capital makes Big Oil look diversified.

6/ TALENT: A 2023 study found that over 70% of top AI researchers (by citation count) work at just a handful of companies: Google, Meta, Microsoft, OpenAI. The public universities that trained them see diminishing returns.

7/ TRAINING DATA: Common Crawl — the web scrape most LLMs train on — covers the internet, but who curates it matters. The companies that invest billions in data pipelines, annotation, and RLHF have a compounding advantage no startup can match.

8/ REVENUE: The AI-as-a-service market tells the story. Microsoft (via OpenAI), Google, and Amazon capture the vast majority of enterprise AI revenue. The "AI startup" story is largely a story of big tech's distribution advantage.

9/ For comparison: At its peak, Standard Oil controlled 91% of US refining. AT&T handled 80-90% of US phone calls. Today's AI concentration across compute, models, and talent rivals the great monopolies of the Gilded Age.

10/ Why does concentration matter? Because concentrated industries set the terms. Pricing, access, safety standards, whose problems get solved — these decisions sit with a vanishingly small number of actors.

11/ Elinor Ostrom showed that common resources can be governed without either privatization or state control. But that requires distributed power. When five labs hold the frontier and three clouds hold the compute, distributed governance is fantasy.

12/ Concentration isn't inherently bad. But this level of concentration, in a technology this consequential, without corresponding accountability? That's a redistribution problem we need to name clearly. /end

---

LinkedIn version:

How concentrated is the AI industry? The numbers are more striking than most people realize.

Three cloud providers (AWS, Azure, Google Cloud) control roughly 67% of global cloud infrastructure. Nvidia holds 80-90% of the AI accelerator market, with TSMC fabricating nearly all of them. Only about five organizations have trained models at the true frontier. In 2024, approximately 75% of US AI startup funding went to just three companies. Over 70% of top AI researchers by citation work at a handful of Big Tech firms.

For historical comparison: Standard Oil at its peak controlled 91% of US refining. AT&T handled 80-90% of phone calls. AI concentration across compute, models, talent, and capital rivals the great monopolies of the Gilded Age.

This matters because concentrated industries set the terms. Pricing, access, safety standards, whose problems get solved — these decisions rest with a vanishingly small number of actors. Elinor Ostrom demonstrated that common resources can be governed without privatization or state control, but distributed governance requires distributed power.

When five labs hold the frontier and three clouds hold the compute, the market structure itself determines who benefits from AI and who doesn't. Concentration isn't inherently bad. But this level of concentration, in a technology this consequential, without corresponding public accountability, is a redistribution problem we need to name clearly.

What aspect of AI concentration concerns you most?