The 25-Year Energy Wall
A Cal Bay AI℠ Essay
Introduction
Software improves in months. Power plants, transmission lines, and substations are built on a 25- to 30-year industrial cycle.
That mismatch is now the defining constraint of the artificial intelligence buildout. For years, the limiting factor was the chip — how much computing could be packed into a processor. Today the limit has moved downstream, to electrons and the equipment that delivers them: generation, transmission, transformers, and the skilled workers who install them.
The consequence is a collision. The largest technology companies are spending at a pace the public grid cannot absorb without pushing costs onto ordinary ratepayers — and as resistance grows, more and more of the buildout is moving behind the meter, onto private power that never touches the public grid.
Code ships in months. Copper, steel, and concrete ship in decades.
Part I: From Bits to Gigawatts
Density
A traditional data center rack draws roughly 8 to 12 kilowatts. Racks built for modern AI accelerators can draw 40 to more than 120 kilowatts — ten times as much power in the same footprint.
| Traditional Rack | AI Accelerator Rack | |
|---|---|---|
| Power Draw | About 8–12 kW | About 40–120+ kW |
| Cooling | Air | Liquid piped directly to the chips, or full immersion |
| Load Behavior | Steady | Training clusters can swing by many megawatts in moments, stressing electrical equipment |
Cooling and Water
All that power becomes heat. Evaporative cooling consumes large volumes of water, which has turned many proposed campuses into local water fights. The industry is shifting toward closed-loop liquid cooling and immersion systems that use far less water — at a higher up-front cost.
The Interconnection Deficit
Before a new data center or power plant can connect to the grid, it must wait in an interconnection queue run by the regional grid operator — PJM in the Mid-Atlantic, ERCOT in Texas, MISO in the Midwest, CAISO in California, and others.
| Reality | What It Means |
|---|---|
| Long Waits | Research from Lawrence Berkeley National Laboratory has found that projects built recently spent roughly five years from interconnection request to operation. |
| Five-to-Seven-Year Lag | A large power agreement signed today can mean an energization date well into the next decade. |
| Phantom Requests | Speculative developers file applications for projects they may never build, clogging studies and slowing planning for everyone. |
Part II: The Equipment and Labor Bottleneck
Transformers and High-Voltage Equipment
| Component | The Bottleneck |
|---|---|
| Large Power Transformers | Lead times that once ran about a year now average well over two years. Wood Mackenzie’s 2025 survey found roughly 128 weeks for large power transformers and 144 weeks for generator step-up units. |
| Grain-Oriented Electrical Steel | The specialized steel inside transformers has very limited domestic production — Cleveland-Cliffs is the only U.S. producer — leaving supply exposed to trade disruptions. |
| Switchgear and Breakers | High-voltage switchgear, circuit breakers, and bus systems for gigawatt-scale feeds face their own long backlogs. |
| Gas Turbines | As more projects seek on-site gas generation, orders for large turbines have stretched manufacturers’ delivery schedules years into the future. |
The Skilled Trades Choke Point
The grid is built by hand. The country faces a structural shortage of high-voltage substation technicians, utility line workers, industrial pipefitters, and instrumentation and controls technicians.
That shortage creates a mismatch at the center of the AI economy. As The Structural Bifurcation describes, AI is displacing white-collar work — but a displaced analyst cannot become a licensed high-voltage technician overnight. The skills take years of apprenticeship to build. That makes skilled trades one of the most durable forms of leverage in this transition.
Part III: The Public Grid Conflict
Who Pays
When utilities expand transmission to serve new data centers, the cost is often spread across all customers through rates approved by state public utility commissions.
| Flashpoint | What’s Happening |
|---|---|
| Northern Virginia | The world’s largest data center market has become a center of debate over transmission lines, land use, and rates. |
| Georgia | Rapid load growth forecasts have driven major new generation plans and public argument over who bears the cost. |
| The Pacific Northwest | Data center demand competes with homes and industry for limited hydropower and transmission capacity. |
| The PJM Grid | As The Manufactured Eschatology notes, projected data center demand helped drive capacity prices up roughly tenfold in a single auction. |
The result is a growing wave of local moratoriums, zoning fights, and lawsuits — including over water permits — that can halt new campuses before they break ground.
The Limits of Intermittent Power
AI campuses are designed to run around the clock with extremely high reliability. Solar and wind are essential parts of the grid, but on their own they are not available every hour.
| Source | Typical U.S. Capacity Factor | What It Means |
|---|---|---|
| Solar | Roughly 20–25% | Produces only when the sun shines |
| Wind | Roughly 30–45% | Produces only when the wind blows |
| 24/7 Data Center | Needs power nearly every hour of the year | Requires firm power or long-duration storage |
This exposes a gap in corporate climate claims. Many companies offset their use with virtual power purchase agreements and renewable energy certificates — contracts that pay for clean power somewhere on the grid over the course of a year. But they do not guarantee that clean power is actually flowing at the hour the servers draw it.
Part IV: The Behind-the-Meter Turn
Facing years-long queues, equipment shortages, and political resistance, the largest buyers are increasingly going around the public grid — placing generation and computing together on private sites.
Nuclear Deals
| Deal | How It Works |
|---|---|
| Constellation and Microsoft — Three Mile Island | A long-term agreement to restart Three Mile Island Unit 1 in Pennsylvania and buy its output. The power flows through the grid rather than behind the meter, but the deal shows how far buyers will go to secure round-the-clock supply. |
| Talen and Amazon — Susquehanna | Originally structured to place a data center campus directly beside the Susquehanna nuclear plant. After federal regulators rejected an expanded behind-the-meter arrangement, the deal was restructured as a large grid-connected power purchase agreement. |
These cases sit at the center of a live regulatory fight: whether a data center placed beside a power plant must still pay its share of the transmission network — or can operate entirely behind the plant’s own connection point. Federal regulators and grid operators are still working out the rules.
On-Site Gas Generation
Other builders are installing their own natural gas generation — fast-starting aeroderivative turbines and large reciprocating engines — fed directly from nearby pipelines. This can bypass utility queues entirely, at the cost of new emissions, fuel-price exposure, and the turbine backlogs noted above.
The Industrial Microgrid
| Component | Role |
|---|---|
| On-Site Generation | Firm, around-the-clock power |
| Battery Storage | Smooths sudden load swings and covers short outages |
| Solar | Lowers fuel use during the day |
| Black-Start Capability | Lets the site restart itself without the grid after a failure |
| Waste-Heat Reuse | Channels server heat into district heating networks or nearby industry instead of venting it — a practice already common in parts of Northern Europe |
Part V: The 25-Year Horizon — An Illustrative Scenario
No one can forecast AI power demand precisely. But a simple scenario shows the shape of the problem when demand grows faster than the grid can expand.
| Year | Compute Power Demand | Grid-Delivered Capacity | Behind-the-Meter Generation |
|---|---|---|---|
| 2025 | 45 GW | 35 GW | 5 GW |
| 2030 | 120 GW | 60 GW | 35 GW |
| 2035 | 210 GW | 95 GW | 80 GW |
| 2040 | 300 GW | 140 GW | 130 GW |
| 2045 | 380 GW | 190 GW | 175 GW |
| 2050 | 450 GW | 240 GW | 210 GW |
Illustrative scenario, not a forecast. The figures show a possible trajectory to make the structural gap visible.
Two patterns stand out:
- The early gap is the widest. In this scenario, the grid covers only about half of compute demand by 2030. That is where projects stall, queues lengthen, and rate fights intensify.
- Private power fills the gap. By 2050 in this scenario, nearly half of AI’s power comes from behind-the-meter generation — a fundamental shift away from the shared public grid.
Demand Outruns the Grid → Costs Hit Ratepayers → Public Resistance → Private Power → A Split Energy System
Part VI: What It Means
| Lesson | Why It Matters |
|---|---|
| Power Is the New Bottleneck | The pace of AI is now set less by chip design than by transformers, substations, and permits. |
| Land with Power Is the Prime Asset | Acreage near firm power and substation capacity commands a premium; land without it is worth far less to builders. |
| The Grid Is Splitting | A growing share of industrial power is moving off the shared grid, raising hard questions about who pays for the system everyone else still depends on. |
| The Trades Hold Leverage | Electricians, line workers, pipefitters, and controls technicians are among the scarcest — and least automatable — workers in the economy. |
| Communities Have a Voice | Rate cases, zoning hearings, and water permits are where the public decides the terms. Showing up matters. |
Conclusion
The artificial intelligence economy is often described as weightless. It is not. Every model runs on electrons delivered through steel, copper, and concrete — on a timeline measured in decades.
That wall will shape who wins and who pays. It will push the largest companies toward private power, push costs toward ratepayers who are not paying attention, and push value toward the people and places that control physical capacity.
Software moves at the speed of thought. Infrastructure moves at the speed of steel. The future belongs to those who understand both.
As The Physical Hierarchy of Compute argues, the answer for individuals and communities is not to compete with the giants at their own scale — it is to understand the physical system, claim a voice in how its costs are shared, and own the capacity that is within reach.