The New Bottleneck Is Concrete, Not Compute
Data centers take 12 to 18 months to build. Connecting them to the grid now takes 5 to 7 years. That gap is reshaping who wins the next AI capex cycle.
The world's largest technology companies have made a collective bet: about 190 gigawatts of new power capacity, more than the entire U.S. nuclear fleet, dedicated to running AI. The chips are ready. The buildings can be poured in 12 to 18 months. The grid, for the first time in a generation, is the binding constraint.
This is the structural shift Bessemer Venture Partners' Atlas roadmap on the AI data center stack identified last month, and it extends past venture capital's usual scope. The framing has inverted: in the late 2010s, the bottleneck was silicon. In 2026, the bottleneck is concrete, wire, and the queue at the interconnection desk.
The numbers behind the inversion are blunt. According to Sightline Climate's Data Center Outlook research, as cited by BVP's roadmap, hyperscalers have announced about 190 gigawatts of data center capacity across 777 projects as of early 2026, with roughly 148 gigawatts still in planning, 21 gigawatts under construction, and only 12 gigawatts operational. That is not a forecast of build. It is a forecast of announcements. Most of those projects will collide with the same wall.
That wall is the gap between two clocks. A data center shell can be built in 12 to 18 months. Connecting it to the U.S. transmission grid now takes 5 to 7 years on average, according to BVP's roadmap, and the PJM Interconnection data is even worse: AI infrastructure projects that reached service in 2025 took more than seven years on average, with more than three years to reach an interconnection service agreement and another four years waiting to come online after approval, Data Center Knowledge reported on May 12. PJM spokesperson Jeff Shields told Data Center Knowledge that the timelines reflect a system that was not built for the speed at which demand is now arriving.
The consequence is not that data centers stop getting built. It is that they get built anyway, on a parallel track the grid does not control.
Hyperscalers have started voting with their capital. Alphabet announced an agreement to acquire Intersect, a developer of on-site power and energy systems, to lock in capacity that grid timelines cannot promise. This is not a green-energy gesture. It is a procurement decision driven by the math that waiting seven years for a grid tap is unacceptable when the model training window is measured in quarters.
That is why the federal government moved in April. On April 20, 2026, the White House issued three presidential determinations under Section 303 of the Defense Production Act, covering grid infrastructure and equipment, coal supply chains and baseload generation, and large-scale energy infrastructure. The determinations do not write checks. They declare the domestic supply chain for transformers, high-voltage circuit breakers, transmission conductors, protective relay systems, capacitor banks, and electrical core steel to be essential to national defense, which unlocks federal financing, preference rates, and procurement priority through the Department of Energy, according to a legal analysis by Foley & Lardner. The framing matters: grid hardware is now treated the way microelectronics and rare earths have been treated since the CHIPS Act.
Inside the queue, the binding constraint has migrated. PJM's January 2026 construction-metrics data shows that for projects under construction or development, permitting now drives 29 percent of milestone change requests, EPC procurement, equipment, and construction account for 28 percent, and supply chain for 23 percent. The interconnection desk is no longer where the schedule dies. The supply chain for transformers, the EPC labor pool, and the local permit are.
The scale of the mismatch is hard to overstate. PJM has processed more than 170,000 megawatts of generation requests since 2023, has roughly 30 gigawatts of transition-cycle projects still to clear, and received approximately 220 gigawatts of new applications for its next cycle, Data Center Knowledge reported. The grid operator projects summer peak demand rising from about 154 gigawatts in 2025 to roughly 210 gigawatts by 2036, with data centers and electrification doing most of the lifting. PJM is now evaluating 95 large-load adjustment requests totaling about 54 gigawatts, many of them AI campuses in the hundreds-of-megawatts to gigawatt range, the publication said.
This is where startup agency actually lives. The categories that have room to move are concrete, not abstract: software that orchestrates the interconnection queue and the upstream permitting process, transmission and power-conversion hardware that compresses siting timelines, modular and advanced cooling systems that relax water and climate constraints, behind-the-meter generation that turns a multi-year grid wait into a 24-month turnkey project, and skilled-trades labor and construction technology that breaks the EPC bottleneck. Each of these is a sub-market BVP's roadmap identifies as underbuilt relative to demand.
The capital is already moving. Data centers accounted for 78 percent of built-environment venture investment in 2025, according to Sightline Climate data cited by BVP: $4.5 billion of $5.7 billion deployed. That is a venture signal, not a grid signal, and the lag between the two is the opportunity Bessemer is betting on.
The friction the venture thesis tends to underweight is real. New generation, even on-site, has an emissions profile, and the AI campus siting fights of 2024 and 2025 in Northern Virginia, Phoenix, and Columbus have shown that local opposition can compress a project's political runway faster than the interconnection queue can. Cooling systems that draw on strained municipal water in arid regions are already drawing regulatory attention. Large flexible loads on a transmission system designed for steady industrial customers create a stability question that operators have only begun to study. And the entire buildout rests on an assumption about AI demand that the GPU cycle, the application cycle, and the capex cycle will all support through the end of the decade. If any of those assumptions slip, the 148 gigawatts of planning is the first to get cut.
McKinsey research, cited in BVP's roadmap, estimates that more than $5 billion is spent annually on complex infrastructure permitting in the U.S., a third of total federal permitting spend, and that more than $1.5 trillion of infrastructure capital is currently stuck in the permitting pipeline. Between March 2024 and March 2025, 16 data center developments were delayed or denied on permitting grounds. Permitting is now the largest single drag on the AI buildout, and the friction is bipartisan, state-level, and almost entirely outside the scope of the federal DPA action.
What to watch next is specific. The next PJM capacity auction will reveal whether generation is clearing at prices the AI campuses can actually pay. The first wave of Section 303-funded transformer orders will show whether U.S. electrical steel and core manufacturing can scale on a defense timeline. The Northern Virginia 2027 zoning round will set the political ceiling for behind-the-meter generation in the largest U.S. data center cluster. And the next hyperscaler capex call will answer the real question: whether the AI buildout is, at its core, a grid story or a generation story, and whether the bottleneck that defines 2026 is the bottleneck that defines 2028.
For now, the grid is winning. The chips are ready. The buildings are being poured. The queue is seven years long, and the line keeps getting longer.