It's already mid-October, and predictions for new IT capacity in US data centers in 2026 remain between 10 and 21 gigawatts (GW), with a median of 12.8 GW. With only about two months left until the end of the year, these predictions still differ by more than double.
The divergence widens as the forecast period progresses: the 2027 forecast ranges from 12.5 to 39 GW, with a median of 18.0 GW; 2028 from 16.6 to 54 GW, with a median of 22.6 GW; and the cumulative total from 2026 to 2028 is estimated at 40 to 114 GW, with a median of 53.7 GW – a difference of nearly three times.
Nine out of ten predictions have cumulative values between 40 and 61.5 GW, with the most optimistic estimate at 114 GW. Even excluding this, significant differences remain regarding the growth rate. Based on the illustrative assumption of $50 billion in capital expenditure for full-stack AI infrastructure per gigawatt of IT capacity, the projected gap in 2026 corresponds to approximately $550 billion in investment, and the cumulative gap of 74 gigawatts over three years corresponds to approximately $3.7 trillion.
If OpenAI and Anthropic each reach 5 gigawatts by the end of 2026 and 10 gigawatts by the end of 2027, with an annualized revenue of $20 billion per gigawatt, their combined annualized revenue will reach $200 billion and $400 billion respectively. If each is short by 1 gigawatt, assuming the monetization level remains unchanged, their combined year-end annualized revenue will be $40 billion less.
