What Is This?
AI scaling is usually explained as a compute race:
better models + more chips + more capital -> more capability
That model is now too narrow. Frontier AI is becoming a physical-infrastructure race as well.
Chris Gillett's Works in Progress essay, “What's really slowing down the AI buildout,” makes the correction cleanly: America may have enough electricity in aggregate, but the hard problem is getting usable power to the specific places where giant data centres need it, on the timelines AI labs want.
The better model is:
AI scaling = models + chips + capital + power delivery + grid interconnection + local infrastructure
A GPU can be bought. A data-centre campus can be financed. But a gigawatt-scale load still has to connect to a grid built around slower, regulated, local, physical constraints.
Why Does It Matter?
The bottleneck changes the forecast.
If AI is constrained only by chips, then progress mostly depends on semiconductor supply chains, model efficiency, and capital expenditure.
If AI is also constrained by electricity delivery, then progress depends on:
- grid interconnection queues;
- transmission buildout;
- substation and transformer capacity;
- local distribution constraints;
- permitting;
- utility planning cycles;
- on-site generation and backup power;
- demand flexibility;
- where data centres can actually be sited.
That means AI timelines become less smooth and more regional.
not “how much electricity exists?”
but “where can 500 MW to 1+ GW be delivered, reliably, soon?”
That is a different question.
The Wrong Bottleneck Model
The simple AI-scaling model focuses on scarce compute:
chips -> clusters -> training runs -> better models
That is still true. But the largest clusters increasingly behave like industrial facilities. Gillett notes that OpenAI and SoftBank's Stargate campus in Abilene is expected to draw about 1.2 gigawatts at peak load. That is no longer “just a server room.” It is city-scale electrical demand.
The mistake is treating power as a commodity input that appears when the cheque clears.
Electricity is not only energy. It is delivered capacity.
energy = total amount consumed over time
capacity = how much can be delivered at once, in a place, through real equipment
AI buildout stresses capacity.
The Grid Stack
Think of the grid as a stack:
generation -> transmission -> interconnection -> substations -> local distribution -> data-centre power systems -> workloads
A failure or delay at any layer can block the whole project.
1. Generation
Someone has to produce the electricity: gas, nuclear, renewables, hydro, batteries, or on-site generation.
But generation alone is not enough. A region can have theoretical energy supply and still fail to deliver power to the load centre.
2. Transmission
High-voltage lines move power over distance. Transmission is slow to build, politically exposed, and often shared across many competing priorities.
3. Interconnection
New power plants and storage projects do not simply plug in. They enter interconnection queues, where grid operators study what upgrades are needed before connection.
Lawrence Berkeley National Laboratory's Queued Up project reported that, as of the end of 2025, more than 2,060 GW of generation and storage capacity was actively seeking grid connection. The same project notes that many queued projects are withdrawn, and those that do get built are taking longer to move through study and completion.
That is the core friction: the clean-energy and storage pipeline can be enormous while usable grid capacity still arrives slowly.
4. Local delivery
A data centre does not consume “national electricity.” It consumes power through local substations, transformers, feeders, switchgear, backup systems, and utility agreements.
This is where aggregate forecasts mislead. A country can have enough generation on paper while a specific county, utility territory, or substation cannot support the next campus without upgrades.
5. Workload flexibility
AI workloads are not all equal.
Training can sometimes be scheduled. Inference may need low-latency service. Some jobs could move across regions; others cannot. If data centres can flex demand during grid stress, they may be easier to integrate. If they require flat, always-on, high-reliability load, they become harder.
Why “Enough Electricity” Is Not The Same As “Usable Electricity”
The key distinction:
total electricity supply != deliverable power at a specific site by a specific date
The EIA has reported that U.S. electricity consumption is rising again after more than a decade of little change, with recent and forecast growth coming from commercial demand, including data centres, and industrial demand, including manufacturing.
EPRI's Powering Intelligence report makes the uncertainty explicit. It says U.S. data-centre load growth is highly uncertain because generative AI could raise computational demand, while efficiency gains could offset some of that demand. EPRI highlights three necessary strategies: data-centre efficiency and flexibility, close coordination between data-centre developers and electric companies, and better modelling tools for 5–10+ year grid investments.
That combination is the point.
AI electricity demand is not a single number. It is a planning problem under uncertainty.
forecast demand + uncertain efficiency + slow grid assets + local constraints = bottleneck risk
What This Changes About AI Timelines
1. Scaling becomes lumpier
Data-centre capacity will not appear evenly. It will cluster where power, land, fibre, permitting, tax incentives, water/cooling, and grid agreements line up.
That favours places with prepared infrastructure and utilities willing to move quickly.
2. Power procurement becomes strategic
The winners are not only the labs with the best models. They are also the firms that can secure power, sites, utility relationships, and buildout execution.
The new question is:
who can turn capital into energized capacity fastest?
3. Efficiency becomes more valuable
If power delivery is binding, then model efficiency is not just cost reduction. It is capacity expansion.
A system that gets the same capability with less energy can serve more demand inside the same physical envelope.
4. Regulation becomes part of AI strategy
Permitting, utility planning, interconnection reform, transmission policy, and reliability rules become upstream variables in AI progress.
This makes AI timelines more political than a pure chip-supply model suggests.
5. On-site power becomes more attractive
If grid connection is slow, data-centre developers have stronger incentives to explore dedicated generation, behind-the-meter arrangements, batteries, gas turbines, nuclear power-purchase agreements, and other local solutions.
Those do not remove constraints. They move the bottleneck to fuel, permitting, capital, cooling, reliability, and integration.
The Forecasting Trap
Power-demand forecasts vary because several uncertain variables multiply:
- how fast AI adoption grows;
- how much inference dominates training;
- whether model efficiency keeps improving;
- whether specialised chips reduce energy per task;
- how much demand shifts across regions and time;
- whether data centres become flexible grid resources or rigid loads;
- how quickly grid upgrades happen.
The unsafe forecast is:
AI demand number goes up -> therefore the grid breaks
The safer forecast is:
AI load growth creates local, time-sensitive power-delivery bottlenecks unless grid planning, flexibility, and infrastructure buildout improve
That is less dramatic, but more useful.
How To Use This
When reading an AI infrastructure announcement, ask:
How much power is required?
Is the project talking megawatts, hundreds of megawatts, or gigawatts?Is power secured or merely planned?
There is a difference between a site announcement, a power-purchase agreement, an interconnection agreement, and an energized facility.Where is the bottleneck?
Generation, transmission, interconnection, substations, local distribution, cooling, backup power, or permitting?How flexible is the load?
Can workloads shift by time or geography, or does the site require constant high reliability?What is the timeline mismatch?
AI capex can be announced in months. Grid assets often take years.What assumption makes the forecast wrong?
Efficiency gains, slower AI adoption, delayed interconnection, gas constraints, nuclear delays, water/cooling limits, or demand-response breakthroughs?
Practical Takeaways For Jamie
- Do not forecast AI from chips alone. Power delivery is now part of the capability curve.
- Watch energized capacity, not press releases. A billion-dollar campus matters only when it can draw reliable power.
- Treat efficiency as strategic. Better inference economics can relax infrastructure constraints, not just lower bills.
- Look for regional arbitrage. AI buildout will favour places where power, land, fibre, and permitting align.
- Apply this to opportunity work. The Micro-SaaS equivalent is: demand is not enough; delivery capacity and distribution bottlenecks decide who can actually serve the market.
Why Smart People Get This Wrong
They confuse electricity with deliverability
“America has enough power” is not the same as “this data-centre campus can get 1 GW by 2028.”
They model AI as weightless software
AI feels digital at the interface. At scale, it is steel, concrete, transformers, substations, cooling, gas turbines, fibre, and power contracts.
They overtrust single-number forecasts
Demand forecasts are sensitive to assumptions about adoption and efficiency. A single 2030 number hides the real planning uncertainty.
They ignore queues
A project in a queue is not a project connected to the grid. Queue volume can signal future supply, but it can also signal congestion and attrition.
They assume constraints stop progress
Constraints redirect progress. They change where data centres are built, which architectures matter, what firms prioritize, and where profit pools move.
What This Does Not Prove
This does not prove that AI scaling stops.
It does not prove:
- data centres will break the grid everywhere;
- electricity demand forecasts are precise;
- chips no longer matter;
- power is the only binding constraint;
- on-site generation solves everything;
- every AI lab faces the same bottleneck;
- policy reform can instantly clear interconnection delays.
The safer conclusion is:
AI scaling is increasingly governed by physical delivery constraints, and the binding constraint may be local grid capacity rather than aggregate national electricity supply
Key Terms
- Load: electricity demand from a customer or system.
- Peak load: maximum power drawn at a point in time.
- Megawatt / gigawatt: units of power. One gigawatt equals 1,000 megawatts.
- Interconnection queue: the process and waiting list for new generation or storage projects seeking grid connection.
- Transmission: high-voltage movement of electricity over distance.
- Substation: equipment that transforms voltage and routes power into local systems.
- Behind-the-meter generation: power produced on or near the customer side of the utility meter.
- Demand response: reducing or shifting load in response to grid conditions or prices.
Recall Questions
- Why is total electricity supply different from deliverable power?
- What are the layers of the grid stack that matter for AI data centres?
- Why do interconnection queues matter for AI scaling even when they concern generation projects?
- How can model efficiency act like extra infrastructure capacity?
- What should Jamie check before taking a new AI data-centre announcement seriously?
Best Resources to Learn More
- Chris Gillett's Works in Progress essay for the clearest narrative explanation of the bottleneck.
- Lawrence Berkeley National Laboratory's Queued Up project for interconnection-queue data.
- EPRI's Powering Intelligence report for uncertainty around AI and data-centre demand.
- EIA's electricity-consumption reporting for the broad U.S. demand context.
Sources
- Chris Gillett, “What's really slowing down the AI buildout,” Works in Progress, 2026. https://www.worksinprogress.news/p/ai-is-bottlenecked-by-the-grid
- Lawrence Berkeley National Laboratory, “Queued Up: Characteristics of Power Plants Seeking Transmission Interconnection,” queue data through 2025. https://emp.lbl.gov/queues
- Electric Power Research Institute, Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption, Product ID 3002028905, 2024. https://www.epri.com/research/products/000000003002028905
- U.S. Energy Information Administration, “After more than a decade of little change, U.S. electricity consumption is rising again,” 2025. https://www.eia.gov/todayinenergy/detail.php?id=65264