Almost nobody on earth can build one of these.
You use these tools every day. The number of organisations that can actually make one is small enough to list, and it gets smaller the further down you look. Here is the whole stack, counted, with sources.
Every layer below depends on the one under it, and every layer is narrower. People argue about which company they trust. The more useful question is how few there are to choose between.
The stack, counted
Click any row. The bar is roughly to scale.
Who actually builds them
A frontier lab trains its own models at or near the top of the compute distribution. A company that ships an AI product buys somebody else's. The distinction is now written into law: Europe presumes systemic risk above ten to the twenty five operations of training compute, and California defines a frontier model above ten to the twenty six.
The American labs are OpenAI, Anthropic, Google DeepMind, SpaceXAI (which absorbed xAI in February 2026), Meta and Microsoft AI. The Chinese labs are DeepSeek, Alibaba, Moonshot, Z.ai, ByteDance, Tencent and Baidu. Mistral in France is the only European organisation training at this scale. That is the list.
Ownership is worth knowing because it is not what the branding suggests. Microsoft holds about 27 percent of OpenAI on an as converted basis, valued around 135 billion dollars. Amazon and Google have each committed tens of billions to Anthropic, and neither has ever disclosed what percentage it owns. Both Anthropic and OpenAI filed confidential draft stock offering documents in June 2026, which means no public financials exist yet.
Two of the largest companies on earth have put tens of billions of dollars into a single AI lab, and neither will say what share of it they own. That is not a scandal. It is just the level of visibility the public currently has.
What one of these costs
Training compute for the largest models has grown about five times a year since 2020. Cost has grown about three and a half times a year. Epoch AI's estimate for a single recent training run, xAI's Grok 4, is roughly 490 million dollars and 310 gigawatt hours of electricity, with significant uncertainty attached.
The figure people quote in the other direction, DeepSeek's 5.6 million dollars, is real but describes only the final training run. It excludes the failed experiments, the research staff and the cluster itself. Quoting it as the cost of building DeepSeek is like quoting the petrol as the cost of the car.
The number that puts it in scale
Alphabet, Amazon, Meta and Microsoft are together guiding to somewhere around 600 to 685 billion dollars of capital spending in 2026 alone. Epoch estimates that across the big five, capital spending overtakes operating cash flow around the third quarter of 2026, meaning the buildout stops paying for itself out of profits and starts requiring debt.
Sixty one percent of all venture capital raised anywhere in the world in 2025 went to AI companies: 258.7 billion dollars out of 427.1 billion. In 2022 the figure was thirty percent.
Where the bill actually lands
This is the part that reaches people who have never opened a chatbot. The world's AI data centres drew about 30 gigawatts at the end of 2025, comparable to the peak power draw of New York State. American data centres used 192 terawatt hours in 2024, about 4.7 percent of national electricity.
The clearest measured consequence so far is in the PJM grid, which serves about 65 million people across thirteen states and Washington DC. Its 2028 capacity auction cleared at the price cap for the third year running, cost 16.4 billion dollars, and still came up 6,831 megawatts short of its own reliability requirement. PJM's independent market monitor attributes 29.4 billion dollars of the 63.6 billion in capacity charges across the last four auctions to data centres.
Who can actually make them stop
Almost nobody, and more than you would guess from the headlines.
The European Commission gained enforcement powers over general purpose AI model providers on 2 August 2026, two days before this page was written. It can demand documentation, demand access to a model to evaluate it, order mitigation, and in serious cases order withdrawal from the European market, with fines up to three percent of worldwide turnover or fifteen million euros, whichever is higher. This is the only such power anywhere in the world.
California can compel a large frontier developer to publish a safety framework and to report critical safety incidents within fifteen days, or twenty four hours where there is imminent risk of death. Penalties up to one million dollars per violation.
Everyone else publishes guidance. The American CAISI and the UK AI Security Institute both work through voluntary agreements. The UK institute has priority access to top models because the labs grant it, not because anyone requires it.
As of today, no authority anywhere can stop a frontier training run before it happens, require permission to start one, or compel anyone to hand over a model's weights. Everything else is paperwork after the fact.
The mismatch
One independent estimate puts the number of people working full time on AI safety worldwide at about 1,100, roughly 600 technical and 500 not, across 115 organisations. The author says it undercounts work happening inside the labs. In the same year, four companies are spending on the order of 600 billion dollars building the systems.
Sources
Everything above is dated. The fast moving items are flagged in the notes so you can tell what will be stale first.