Stanford's annual artificial intelligence index recorded that AI data centres draw 29.6 gigawatts globally, roughly New York State's peak electrical demand. The United States operates more than 5,400 data centres, and nearly every leading chip used in them comes from a single manufacturer in Taiwan.
The two numbers describe different fragilities. Consumption in gigawatts explains why capacity contracts began being negotiated in power rather than equipment counts: the electrical grid became the real limit on expansion.
Manufacturing concentration describes the other. Depending on one supplier in a single territory is a risk governments began treating as strategic, which explains the billion-dollar fab plans announced across three continents in the following months.
It's worth noting what those plans don't solve in the short term. A fab takes years between announcement and volume production, so the fragility the index describes holds throughout the period when demand grows faster.
For anyone contracting capacity, the useful data point isn't the global total but the cause of the queue: whoever secures contracted power near available grid capacity determines how much of the market they can serve.