Dirt, Power & Fiber Weekly
Sample Issue · June 2026Power is becoming the AI bottleneck.
A weekly read-through on the physical constraints behind AI infrastructure.
The Signal
For most of the AI cycle, the constraint conversation was about silicon: who could get the chips, train the models, and ship the capability. That phase is ending. The four largest hyperscalers have guided roughly $700 billion of combined capital spending for 2026 — nearly double the prior year — and the question has shifted from whether the capital exists to whether the physical system can absorb it. Berkeley Lab estimates data centers consumed about 4.4% of U.S. electricity in 2023 and could reach 6.7% to 12% by 2028.
The gap between announced demand and deliverable capacity is now measurable in utility filings. ERCOT's large-load interconnection queue reached 233 gigawatts — with under 2% of it operational. Exelon's utilities are studying roughly 43 gigawatts of large-load requests, with some Chicago-area delivery timelines stretching toward 2032. NV Energy logged about 22 gigawatts of data center inquiries against roughly 6 gigawatts of executed agreements. Duke Energy has contracted around 7.6 gigawatts of data center load in the Carolinas and projects data centers at up to a quarter of system demand by 2030.
The system is responding the way constrained systems do: through price and policy. PJM's capacity auction cleared at $329 per megawatt-day, more than eleven times the level two cycles earlier. Oregon implemented a first-in-the-nation data center rate class that raised large-load rates roughly 29% this month. Denver passed a one-year moratorium, Nashville advanced one, Illinois paused new data center tax incentives, and Utah raised development standards by executive order — all within the past year.
This is the Gridlocked frame: AI demand is everywhere, but buildable capacity is not. The binding constraint has migrated from the chip supply chain to the grid — to interconnection queues, transmission timelines, water strategies, and the local politics of energization.
233 GW
of large-load interconnection requests in ERCOT's queue — under 2% operational
~$700B
of combined 2026 capital spending guided by the four largest hyperscalers
6.7–12%
of total U.S. electricity that data centers could consume by 2028, per Berkeley Lab
11x
increase in PJM capacity prices over two auction cycles, to $329 per megawatt-day
Why It Matters
A data center is not built because cloud or AI demand exists. It is built when a specific site clears a stack of physical and institutional constraints — and each layer of that stack can fail independently of demand. Capital that cannot be energized is not capacity; it is a queue position.
That difference — between demand and deliverable capacity — is the Gridlocked thesis. Markets, sites, and companies sort on it. The tightest colocation vacancy on record sits next to interconnection queues measured in years, and the spread between powered, entitled sites and speculative land keeps widening in every market we track.
What a campus actually needs
- Power, with a credible energization date
- Utility interconnection and grid delivery
- Land and site control
- A cooling and water strategy
- Fiber depth
- Permits and local acceptance
- Transmission and generation planning behind it all
Market Read-Through
Northern Virginia remains the benchmark and the warning sign: the deepest demand and fiber ecosystem in the world, increasingly gated by substation queues, land scarcity, and permitting fatigue. The market now matters as much for its constraints as its scale.
Texas shows the gap at its widest. Dallas-Fort Worth and the Austin/San Antonio corridor hold some of the deepest hyperscale pipelines in the country, with vacancy near 2% and new supply almost fully preleased — yet ERCOT's 233-gigawatt queue is mostly paper, and new state interconnection rules for large loads are still being written. Delivered power, not announced demand, decides what converts.
Chicago is the spillover market where the constraint arrived after the land rush: record-low vacancy and a $20 billion exurban campus pipeline against ComEd delivery timelines reaching 2032, sharply repriced PJM capacity, and a state incentive pause that creates a grandfathered cost advantage for projects already signed.
The Mountain West is where water joined power as a first-order constraint. Salt Lake City and Reno both offer pre-entitled land, low-cost power, and named hyperscale anchors — and both spent 2026 absorbing water-disclosure laws, protest filings, and utility resource plans that quantify just how much of the aspiration can actually be served.
Markets to watch
Phoenix
Land and demand are compelling; water optics, cooling architecture, and power procurement decide which sites convert.
Charlotte
Roughly 7.6 GW of contracted data center load behind a single regulated gatekeeper, with rate-class and tax-exemption fights now live.
Salt Lake City
Named hyperscale anchors and cheap power met water-disclosure law, protest letters, and a statewide executive order in the same quarter.
Reno
Roughly 22 GW of utility inquiries against about 6 GW of executed agreements — the widest aspiration-to-commitment gap in the West.
Columbus
A credible land-and-power story where utility delivery timing remains the gating item.
Public Market Read-Through
Public markets express the power bottleneck in several distinct ways. These are read-throughs to different layers of the constraint stack — some names screen as beneficiaries, some as regulated exposures, some as constrained operators, and some as second-order exposures. Exposure to the constraint is not the same thing as upside from it.
Regulated utilities
Load growth can support investment programs, but outcomes run through rate cases, cost allocation, and regulatory approval — regulated exposure, not automatic upside.
Power producers
Long-dated power contracts and new generation tied to data center demand sit closest to the power constraint itself.
Electrical & grid equipment
Switchgear, turbines, grid hardware, and thermal systems are the picks-and-shovels layer of energization.
Data center operators
Constrained operators: demand tailwinds are real, but growth depends on power, land, and delivery timelines.
Infrastructure construction
Transmission, substations, and site work are where queue backlogs become physical backlog.
Water, cooling & site systems
Second-order exposure to the cooling and water strategies that increasingly gate western and high-growth markets.
Real estate & site control
Powered, entitled land and infrastructure-adjacent assets carry the scarcity story; generic acreage does not.
What to Watch Next
The constraint stack moves through filings, dockets, and council votes before it moves through earnings. Signals worth monitoring:
- Utility load forecasts and resource plan updates
- Interconnection queue size, attrition, and study timelines
- Transmission approvals and in-service dates
- Power procurement and clean firm power contracts
- Data center moratoriums and zoning changes
- Permitting friction and local opposition flashpoints
- Water and cooling policy headlines
- Grid equipment lead times and backlogs
- Hyperscaler capex guidance revisions
- Utility rate cases and affordability pressure
Sources consulted
Methodology note
Gridlocked is a research and screening tool. Market scores and exposure scores are directional framework inputs, not investment recommendations. Public-market examples are used to illustrate read-throughs to the infrastructure constraint stack, not as buy or sell recommendations. Read the methodology →
In upcoming issues
- Phoenix screens well until water enters the model
- Northern Virginia is the blueprint and the constraint ceiling
- The moratorium map: where local politics moved faster than the grid
Get the weekly brief
Track the physical constraints behind AI infrastructure before they show up in the market narrative.