When TSMC began ramping its first Arizona fab in late 2025, the company encountered a problem no amount of transistor engineering could solve: getting enough electricity to the site on schedule. The same challenge is playing out at semiconductor construction projects across the United States, from Intel’s sprawling Ohio campus to Samsung’s facility in Taylor, Texas. The bottleneck is not chip design. It is the physical infrastructure required to power, wire, and chemically supply the factories that make AI processors. And in 2026, that bottleneck is tightening.
A 2,000-gigawatt traffic jam
The clearest measure of the problem sits in federal energy data. The U.S. Federal Energy Regulatory Commission has documented over 10,000 active interconnection requests representing more than 2,000 gigawatts of proposed generation capacity. These are applications from power plants, solar farms, battery storage facilities, and other generators seeking permission to connect to the transmission grid. Each one represents energy that someone wants to deliver but cannot, because the queue for grid studies and upgrades has grown far faster than regulators and utilities can process it.
That 2,000-gigawatt figure is not a forecast. It reflects projects already filed and waiting. For perspective, total installed U.S. generation capacity sits near 1,300 gigawatts, according to the U.S. Energy Information Administration. The queue holds roughly 1.5 times the nation’s entire existing power supply, yet completion rates for queued projects have historically been dismal. Lawrence Berkeley National Laboratory tracks these patterns through its Queued Up dataset, which shows that a large share of proposed plants withdraw before ever reaching operation. The gap between filed capacity and delivered capacity is where AI chip production timelines collide with physical reality.
FERC has acknowledged the severity of the situation and finalized a rule to reform the interconnection process. In its own explanation of the new framework, the commission notes that the backlog has “grown dramatically,” driven in part by surging demand from data centers and technology companies that need reliable, large-scale power for AI training and inference workloads. New semiconductor fabs, which consume enormous amounts of electricity to run cleanrooms, extreme ultraviolet lithography tools, and cooling systems, are stuck in the same queue as every other industrial power consumer.
Grid delays hit chip factories and data centers alike
The practical consequences are blunt. A chip fabrication plant planned for a region without adequate grid capacity must either wait years for a new substation and transmission line or pay steep premiums for existing power contracts. Data centers that train and run AI models face the same squeeze. Both types of facilities demand not just large volumes of electricity but extremely reliable supply. Even brief outages can ruin semiconductor wafers mid-process or interrupt weeks of model training, losses that can run into the tens of millions of dollars.
FERC’s interconnection reform aims to speed approvals by clustering study requests and imposing stricter readiness requirements on applicants, thinning out speculative filings so viable projects move forward faster. But even optimistic timelines for implementation stretch across multiple years, and the sheer volume of pending requests means meaningful relief will arrive gradually.
For companies building AI accelerators or expanding fabrication capacity in the United States, the grid bottleneck translates directly into production delays. A factory that cannot secure power on schedule cannot produce chips on schedule. A data center that cannot connect to the grid cannot run the training jobs that generate demand for those chips. The constraint is circular: AI growth requires power, power requires grid access, and grid access requires clearing a queue that was never designed for this pace of demand.
That dynamic is already reshaping site selection. Regions with existing surplus capacity and robust transmission infrastructure are becoming more attractive, even if they lack traditional advantages like proximity to major tech hubs. Areas with strong engineering talent or favorable tax regimes risk losing out if their local grids are saturated. Electrons, not just engineers, are now determining the geography of AI hardware.
Some hyperscalers are trying to sidestep the queue entirely. Microsoft, Amazon, and Google have all explored or signed deals for on-site natural gas generation and, in some cases, small modular nuclear reactors. But those alternatives carry their own permitting timelines and capital costs, and none has yet delivered power at the scale a major fab or AI training cluster requires.
Copper and specialty gases add friction
Electricity is the most quantifiable constraint, but it is not the only one. Copper, essential for wiring in both power infrastructure and advanced semiconductor packaging, has been under sustained price pressure. The International Copper Study Group has flagged a tightening supply-demand balance, and benchmark copper prices have hovered near record levels through early 2026. Every new substation, every mile of transmission cable, and every advanced chip package competes for the same metal.
Specialty gases present a different kind of vulnerability. Neon, used in the excimer lasers that power deep ultraviolet lithography, became a headline concern in 2022 when Russia’s invasion of Ukraine disrupted supplies from major Ukrainian purifiers. Since then, chipmakers and gas suppliers have diversified sourcing, with new purification capacity coming online in South Korea, China, and the United States. The acute crisis has eased, but the industry’s dependence on a handful of purification facilities for several critical gases, including fluorine compounds used in etching and deposition, leaves it exposed to future disruptions. No major gas supplier has published detailed capacity-versus-demand figures for these niche products, making it difficult to quantify exactly how tight the market is at any given moment.
The absence of hard public data on material shortfalls is itself telling. Semiconductor supply chains are opaque by design; chipmakers and their suppliers guard procurement details closely. What is visible from commodity markets and analyst reports points to persistent tightness rather than acute crisis, a slow drag on expansion rather than a sudden halt.
The CHIPS Act factor
Washington has committed tens of billions of dollars through the CHIPS and Science Act to subsidize domestic semiconductor manufacturing. TSMC, Intel, Samsung, and Micron have all received preliminary or finalized awards. But subsidies fund construction; they do not automatically deliver the electricity, copper, or gases those factories need to operate. A fab can be built on time and still sit partially idle if its grid connection lags behind the construction schedule.
This mismatch between capital investment and infrastructure readiness is one of the less-discussed risks in the U.S. semiconductor strategy. Federal incentives have accelerated the decision to build, but the permitting and physical buildout of supporting infrastructure operates on a different, often slower, clock.
Where the evidence is strong and where it is not
The strongest evidence in this story comes from two federal sources. FERC’s interconnection data provides a direct, quantified measure of the electricity access problem, with specific numbers drawn from the commission’s regulatory filings. Lawrence Berkeley’s Queued Up research adds historical depth, showing not just how large the queue is but how poorly it has converted pending applications into operating power plants. Together, these sources establish that the grid constraint is real, measured, and growing.
Evidence on copper and gas shortages rests on a thinner foundation. Commodity market reporting and analyst commentary point to tightening conditions, but the absence of granular data from producers or industry associations means these claims carry more uncertainty. Grid-related constraints are well-documented; material shortages are plausible but less precisely quantified.
The distinction matters for anyone trying to assess when AI chip production will catch up to demand. Grid interconnection delays are structural and regulatory, with timelines measured in years and outcomes shaped by formal rulemaking. Material markets are more fluid: prices can spike or fall, new mines can open, and suppliers can reallocate output. If copper or neon shortages ease, they may remove some friction from chip manufacturing, but they will not solve the underlying problem that many new fabs and data centers simply cannot plug into the grid when they need to.
What to watch through the rest of 2026
The most reliable way to gauge the physical limits on AI growth is to track the evolution of the interconnection queue and the early results of FERC’s reforms, rather than fixate on short-term commodity price swings. Key milestones include FERC’s first batch of clustered study results under the new rules, expected to begin producing decisions later this year, and quarterly updates to the Lawrence Berkeley Queued Up dataset that will show whether withdrawal rates are changing.
On the materials side, watch for announcements from major gas suppliers about new purification capacity and for copper inventory reports from the London Metal Exchange and COMEX. If those indicators stay tight while the interconnection queue remains clogged, the conclusion is hard to avoid: the ceiling on AI chip production in 2026 is not set by what engineers can design. It is set by what the physical world can deliver.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
