AI Infrastructure Bottleneck

A present-tense shortage

Ben Horowitz frames the near-term constraint on AI as physical rather than digital: "We're pretty much out of electricity now in the United States. Like not 12 months from now, like right now."¹ The list of what the country does not have enough of is long: electricity, where data center demand is rising vertically while grid capacity build is not; rare earth minerals, whose supply chains are concentrated in China; manufacturing capacity generally; memory, where a new DRAM factory takes five years to build and available supply is already absorbed; and power transformers, whose underlying technology has not changed since the invention of electricity. The order of relief matters: chips will become sufficient before electricity does, but even with enough chips the system stays memory- and electricity-constrained for years.

Horowitz's shorthand for the strategic picture: "the China graph is like this and the US graph is like that." Demand for compute is vertical, and China's infrastructure capacity is growing aggressively to meet it while the US graph is not, which is the framing behind a major venture firm's unusually large capital raise treating infrastructure investment as a matter of national competitiveness rather than only a return-generation question.

Different from the dot-com overcapacity problem

The AI-era shortage inverts the dot-com era's failure mode. In the dot-com period, enormous fiber capacity was built and mostly sat dark because demand never materialized as projected; the bottleneck was on the application and end-user side. In the AI era, GPUs are all lit and running, demand is real and growing faster than capacity, and the bottleneck sits on the supply side: electricity, transformers, memory, rare earths. Instead of building ahead of demand and crashing, the risk is demand outrunning the capacity to build. As Horowitz puts it, "almost everything is a bottleneck," distributed across multiple points in the stack simultaneously rather than localized in one place the way bandwidth was in the 1990s.

The classic economic response, that the cure for high prices is high prices, still applies, but the latency is extraordinary: a DRAM factory takes about five years, a new power plant seven to fifteen years, a transformer company building new designs three-plus years. High prices today will eventually signal enough investment to relieve the constraint, but that relief will not arrive on a twelve-month horizon.

The building-materials layer

A dimension the compute-and-electricity framing misses: before a GPU cluster or transformer can be installed, the building has to exist, and data centers consume construction materials at scale. Brad Jacobs, discussing a roughly 17 billion dollar acquisition of a building-products distributor, put it directly: "Data centers need roofs too. Data centers need waterproofing very much so. Data centers often need lumber-related products. Data centers are big consumers of building products."² At the time of the acquisition, data center revenue was a single-digit percentage of the acquired company's total but was described as fast growing, spanning insulation for thermal management, waterproofing against what would otherwise be catastrophic water intrusion in a server hall, roofing, and structural lumber.

This means the AI buildout creates demand up and down the physical supply chain: compute, power, cooling, memory, and also the building envelope around all of it. The building-materials layer carries the same structural tailwind as electricity and memory but is tracked separately, under commercial construction rather than AI capital expenditure, which means investors focused on chips and power generation are not necessarily pricing in the roofing and waterproofing demand the same investment wave is generating.

Why it matters

For investors, the infrastructure gap is investment-grade rather than venture-scale: it is closer to a nation-state problem than a return-generation one, which is part of the reasoning behind unusually large infrastructure-focused capital raises. For founders, the bottleneck determines who can scale, since companies that secure electricity and compute access early gain a structural advantage that cannot be bought away later, the inverse of the usual assumption that money alone solves software problems. For AI development timelines generally, the pace of capability deployment is gated by the pace of infrastructure build: model improvement without sufficient inference infrastructure is theoretically unbounded, but real-world deployment capacity is bounded by watts and wafers.