The scale of the artificial intelligence infrastructure buildout, in terms of spending and corporate capital expenditure, continues to reach extraordinary levels. The pace of investment is notably more compressed than any previous infrastructure cycle, driven primarily by the largest cloud and hyperscale operators — Amazon, Microsoft, Google, and Meta — whose combined annual capex has climbed into the hundreds of billions of dollars.

A central question is exactly where the capital is being deployed. Unlike historical infrastructure projects such as railroad construction, which required steel mills and large workforces, the AI data center buildout has a markedly different cost structure. Industry estimates indicate that approximately 80% of data center costs go directly to Nvidia, with a substantial portion of the remainder directed toward rack components and related hardware. The physical construction component — the part that flows into the broader economy — may account for 5% or less of total spending. This concentration means the economic multiplier effect of the AI buildout is narrower than past infrastructure booms, with most dollars flowing to a small set of semiconductor designers and server-rack suppliers rather than diffusing broadly across construction labor and materials.

These facilities are also tremendously energy-intensive. The construction of power generation capacity represents a broadly dispersed expenditure that flows through factories and raw materials. However, the numbers surrounding power generation requirements are striking, and there are serious questions about whether capacity can keep pace with data center demand.

According to ZeroHedge, the energy mathematics behind the data center rollout raise significant concerns. Texas alone faces 474 gigawatts of interconnection requests through ERCOT, with 90% attributable to data centers. Texas Governor Abbott subsequently froze the rollout of new data centers in the state.

For context, ERCOT's all-time peak demand record stands at 91.1GW, set on July 22, 2026, with normal demand ranging between 40 and 80GW.

In a more measured scenario, Morgan Stanley projects that the entire United States will face a shortfall of approximately 80-100GW through 2030, equivalent to the capacity of roughly 100 nuclear power plants. This raises the possibility that even if data centers are constructed on schedule, they may not be permitted to connect to the electrical grid for years. Several major tech companies have already begun pursuing nuclear power agreements and next-generation reactor technologies as one response to this bottleneck, though these initiatives face their own multi-year development timelines and regulatory hurdles.