Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

The Electron Gap: Why Energy Is the Real Bottleneck in the AI Era

The critical, overlooked constraint that may determine which civilizations thrive in the age of artificial intelligence.

The Paradox of Digital Abundance

We speak of AI as if it runs on pure thought: ethereal, weightless, infinitely scalable. The reality is dirtier. Every token generated (each word or word-piece an AI produces), every inference computed (each time AI processes a question and produces an answer), and every agent orchestrated requires electrons, by the billion.

The “Electron Gap” describes the disparity between the strategic necessity of maintaining AI leadership and the physical capacity of national power grids to support it. Training a frontier AI model already requires more electricity than some small nations consume annually. Running those models at scale for millions of users, in real-time, compounds the demand. This is what happens when a technology’s marginal cost collapses toward zero everywhere except one layer: the value doesn’t disappear, it piles up on whatever’s left that’s actually scarce, and today that’s the electron.

This is not a metaphor but a question of physics.

Why Davos 2026 Made This a Priority

At the World Economic Forum’s January 2026 meeting, AI energy demand dominated discussions. Nearly two-thirds of leaders across technology, energy, and policy sectors agreed that a net-positive AI energy future is achievable by 2035, but only with immediate action. Elon Musk put the bottleneck bluntly: AI chip production is growing exponentially while new electricity comes online at 3 or 4% a year, so that “very soon, maybe even later this year, we’ll be producing more chips than we can turn on.” That exponential compute curve, the same one that powered a decade of 100x technological leaps, is now colliding with a linear energy supply. Policymakers now accept that AI strategy is inseparable from energy strategy.

Key tensions:

Growth vs. Grid: Economic competitiveness increasingly depends on AI deployment, but power grids weren’t designed for data center loads. This affects you because countries that can’t power AI will fall behind economically, potentially costing jobs and prosperity.
Green Transition vs. AI Transition: Both require massive energy investment. Can we do both simultaneously? If not, we face a choice between economic relevance and climate stability.
Sovereignty vs. Interdependence: Nations racing for AI supremacy may need to import electricity, or sacrifice other sectors to prioritize compute. A country could find itself choosing between powering hospitals and powering data centers.

The Electron Gap positions electricity as a strategic asset comparable to oil in the 20th century, not merely as a household utility. Some thinkers go further, arguing that in a world where labor is automated, energy itself could become the unit that backs money.

The Numbers

Training Costs

Model	Estimated Training Energy	Equivalent
GPT-3 (2020)	~1,287 MWh	120 US homes for a year
GPT-4 (2023)	~50,000 MWh (estimated)	4,600 US homes for a year
Frontier models (2026)	100,000+ MWh per run	Small cities
Projected 2030 runs	4-16 GW sustained draw	Millions of homes

Inference Costs

Training happens once; inference happens billions of times. The cost of a single answer has fallen faster than its reputation. The rule of thumb that a chatbot query uses ten times the electricity of a web search dates from 2023. By the companies’ own accounts a typical text query now takes about a third of a watt-hour, 0.34 Wh for ChatGPT and 0.24 Wh for Gemini, which is roughly what Google said a search cost in 2009. The growth is in what gets asked. The IEA’s 2026 analysis puts GPU electricity for a text answer from a large model at about 0.3 Wh, an answer with reasoning at 7.6 Wh, a standard agent task of four to six model calls at about 1.1 Wh, and the same task with reasoning switched on at roughly 50 Wh, some 160 times the plain answer. Collectively, large language models consume thousands of MWh per day in inference alone. As AI agents become autonomous and run continuous workflows, inference energy demand will dwarf training. Nvidia’s Jensen Huang gave the multiplier on his August 26, 2026 earnings call: an agent needs “probably 15 to 100 times” the compute of a person using the same model, and “most AI,” he said, had crossed from human-prompted to agentic in the previous month. As AI gets more useful and more widely adopted, our electricity bills, both personal and national, will rise accordingly.

Data Center Projections

In 2024, data centers consumed roughly 415 TWh, about 1.5% of global electricity. The IEA’s April 2026 update, Key Questions on Energy and AI, put 2025 at 485 TWh, a 17% jump in a single year, with consumption at AI-focused data centers up 50%. It projects roughly 950 TWh by 2030, around 3% of global electricity, with the AI-focused share tripling over the period. In the US, data centers used 4.7% of national electricity in 2024, and Lawrence Berkeley National Laboratory’s June 2026 update projects 11.8% by 2030, within a range of 9.5% to 15.3%. And these projections don’t fully account for the “agentic economy” where AI operates around the clock, not just when humans prompt it.

If those aggregates feel abstract, one company publishes its own. Google’s environmental report, released July 2, 2026, disclosed that its electricity consumption rose 37% in 2025, the largest single-year increase in the company’s history. Its data centers alone drew 42 million MWh, up from 30.6 million the year before. That is one firm consuming roughly what New Zealand or Denmark uses in a year, and Google’s cumulative power growth since 2019 now exceeds 250%: it has added more electricity demand in six years than in the two decades before them. The company’s own framing matters more than the total. Google warned the climb won’t be linear, because the AI buildout is accelerating faster than the grid is decarbonizing. Amazon reported record emissions the same week, same cause.

The Three Constraints

1. Generation Capacity

We need more electricity, period. The current global grid wasn’t designed for this demand curve.

Short-term responses:

Extending coal and natural gas plant lifespans (climate costs)
Restarting mothballed nuclear plants (regulatory hurdles)
Rapid solar/wind deployment (intermittency challenges: the sun doesn’t shine at night and the wind doesn’t always blow, so we can’t rely on them alone for consistent power)

Long-term solution: Fusion energy. If fusion achieves commercial viability by 2035-2040, the Electron Gap closes. If it’s delayed, the gap becomes a choke point.

2. Grid Distribution

Even if we generate enough power, can we deliver it? Data centers concentrate demand geographically. Transmission infrastructure (power lines, substations, transformers) takes years to build.

This creates a perverse dynamic: AI companies locate near existing power infrastructure, but that infrastructure then becomes overloaded. The fight over who pays to upgrade it is now playing out in statehouses and capacity auctions.

The auctions have already delivered a verdict. PJM, the grid operator for 13 states from Illinois to New Jersey, held its capacity auction for 2028-29 on July 14, 2026: prices hit the federally approved cap of $325 per megawatt-day, the region again fell short of its reliability requirement, by 6,831 MW, and its independent market monitor attributed $6.3 billion of the $16.4 billion bill to data centers, on top of $29.4 billion across the previous four auctions. Everyone on that grid pays. The politics followed the bills. A Gallup poll in March found 71% of Americans oppose a data center in their area, more than the 53% who oppose a nuclear plant; more than 99% of the $31 million spent on campaign ads mentioning data centers this year attacks them; and on August 31 President Trump posted that communities fighting data centers must want to end up “backwards and poor,” the same afternoon his vice president conceded that 99% of the backlash comes from places where a data center means higher electricity prices. Six months earlier, seven companies had signed the White House’s nonbinding Ratepayer Protection Pledge, promising to build or buy their own generation and pay for the wires. The bills say the pledge has not caught up with the auctions.

September brought the regulators in. PJM, having missed its reliability target in two straight auctions, proposed an emergency “backstop” procurement to buy the missing capacity for data centers, with offers due from September 30. The day before, the Federal Energy Regulatory Commission sent parts of the plan back, and its chairman, Laura Swett, wrote that the commission “will not be forced into accepting a deeply flawed, eleventh-hour procurement mechanism with billion-dollar implications for consumers.” The procurement is postponed with no new date. Eight days earlier, Texas Governor Greg Abbott had ordered the state’s environmental agency to pause all permits for data center projects until the grid operator finishes reviewing them. Two of the largest power markets in America spent the end of September telling the industry to wait.

3. Cooling and Heat Dissipation

Computing generates heat, and massive computing generates massive heat. The energy cost of cooling often equals the energy cost of computation itself. This is why data centers are increasingly located in cold climates (Scandinavia, Canada) or near water sources.

Climate change compounds this problem: as ambient temperatures rise, cooling efficiency drops.

Geopolitical Implications

The New Resource Competition

Just as 20th-century geopolitics revolved around oil, 21st-century geopolitics may revolve around electricity. Nations with abundant, cheap, reliable power have an inherent advantage in AI deployment. If solar and fusion ultimately make energy cheap everywhere, the same dynamic could instead end resource wars altogether.

Winners: Countries with hydropower (Norway, Canada), geothermal (Iceland), nuclear expertise (France), or massive solar potential (Middle East, Australia).

Losers: Densely populated nations with aging grids and limited domestic energy resources.

Strategic Vulnerability

A nation dependent on imported electricity for its AI infrastructure is strategically vulnerable. Adversaries could target:

Power generation facilities
Transmission lines
Undersea cables connecting data centers to power sources

This creates pressure for “energy sovereignty,” meaning secure domestic supply chains for AI power.

Washington has started treating the hardware itself as the exposure. On August 26, 2026, the president declared a national emergency over foreign-made equipment in the bulk-power system, giving the Energy Secretary authority to block purchases of transformers, inverters, batteries, turbines, and their control software from covered foreign suppliers. The security logic is sound. The timing is awkward: turbines and transformers are already the scarce parts, and the order narrows who may supply them.

The “Compute Arms Race” Meets Physics

Governments pushing for AI supremacy face a physical ceiling. You can write policy papers faster than you can build power plants. The Electron Gap may force more realistic assessments of what AI leadership actually requires.

The Unscarcity Connection

The Electron Gap reveals an uncomfortable truth in the “abundance” narrative. Digital intelligence may be theoretically unlimited, yet the energy required to sustain it is finite, at least until fusion or other breakthrough technologies mature.

Foundation Infrastructure

The Foundation architecture assumes abundant clean energy. The Fusion Timeline 2024-2030 article explores when this becomes realistic. Until then, Foundation deployment must be staged geographically, prioritizing regions with existing clean energy capacity.

Responsible Compute

The Electron Gap creates an ethical dimension to AI usage. Wasteful compute (training models that add no value, running agents that duplicate work, generating content no one reads) consumes finite energy resources.

This aligns with Constitutional Core principles:

Law 3 (Power Must Decay): Even AI capabilities should be constrained by physical reality.
Law 5 (Difference Sustains Life): Efficiency in compute usage enables broader access.

Systems That Bend

What happens if energy constraints prevent full AI deployment? The Systems That Bend article explores how Foundation systems can operate at reduced capacity, prioritizing survival essentials over optimization luxuries.

Solutions Under Exploration

1. Algorithmic Efficiency

Better algorithms can do more with less power. Advances in model architecture (sparse models, mixture-of-experts) reduce inference costs. The industry is motivated: electricity is a major operating expense.

2. Specialized Hardware

GPUs aren’t the only option. TPUs, custom ASICs, and neuromorphic chips offer better performance-per-watt for specific tasks. Hardware-software co-design can dramatically reduce energy consumption.

3. Distributed Compute

Instead of concentrating compute in massive data centers, distribute it across smaller, geographically dispersed nodes. This reduces transmission losses and enables use of local renewable sources.

4. Time-Shifting Workloads

Non-urgent computation (training, batch processing) can be scheduled for times when renewable energy is abundant: midday solar peaks, windy nights. “Follow the sun” strategies route workloads to wherever clean energy is available.

5. Nuclear Renaissance

New reactor designs (small modular reactors, thorium, molten salt) offer safer, faster-deploying options than traditional nuclear. AI companies are now investing directly: Amazon has committed to 12 SMRs in Washington state (960 MW total, operational in the 2030s) with a broader goal of 5 GW of nuclear by 2039. Meanwhile, SpaceX’s Colossus complex in Memphis ended June with 1.4 GW of compute capacity, more than the average draw of a million homes, and expects to pass 2 GW by year-end, much of it fed by on-site gas turbines that bypass the grid entirely. Colossus is just one front in the GPU arms race building these compute clusters. The turbines are also where the bypass strategy is fraying. The Southaven, Mississippi site runs on 69 temporary gas turbines that neighbors have sued over; SpaceXAI began switching them off on August 31 under a state order that removes all of them by July 2027 as a permanent 1.2 GW plant comes online, and The Information reported that sites assembled from mobile turbines, battery packs, and more than a hundred portable chillers ran for months without backup power or cooling, with uptime below the company’s own 99.9% target. Musk’s answer, confirmed on August 30, is to go one layer down the supply chain: SpaceX is building a foundry in Bastrop, Texas, to cast gas-turbine blades and vanes itself, a component only a handful of foundries in the world make, with order books running to 2029. Chips can be bought and megawatts rented. The shortage moved to the one part you cannot.

6. Fusion Acceleration

Fusion remains the ultimate solution. The Fusion Industry Association counts $14.24 billion raised by fusion companies since 2021, a record $4.48 billion of it in the twelve months to July 2026. The progress is tangible: in June 2026, Helion became the first fusion company to hold the state licenses needed to operate a power plant, securing a Radioactive Materials License and a Radioactive Air Emissions License from Washington State for its Orion plant, and its Polaris prototype has hit 150 million degrees Celsius. The dates are softer than the hardware. Orion was meant to deliver 50 MW to Microsoft in 2028; Helion’s website now says only that it will “begin initial operations” that year, and a company executive told a conference, “We will not, in no universe, have 50 megawatts hit full operation in 2028.” Commonwealth Fusion Systems is building SPARC near Boston, about 80% complete in late August and aiming for its net-energy attempt in 2027, with a 200 MW purchase agreement from Google for the follow-on ARC plant in the early 2030s. On July 30, 2026, CFS raised another $1 billion, reaching $4 billion in total funding, and the investor list mattered more than the number: pension funds, sovereign wealth funds, and industrial partners rather than venture capital. That is the money that finances toll roads and transmission lines, which means someone’s actuaries have decided fusion belongs on the infrastructure side of the ledger. Whether fusion arrives in time to prevent a lasting Electron Gap depends on whether this momentum holds. The stakes of that timing are explored in a sensitivity analysis of what happens if fusion takes 50 years instead of 10.

What This Means for You

For AI Practitioners

Factor energy costs into system design from the start.
Prefer efficient inference architectures over brute-force scaling.
Consider where your compute runs, not just how fast it runs but how clean.

For Policymakers

AI strategy must integrate with energy strategy. They’re inseparable.
Grid modernization is AI infrastructure.
“Energy independence” increasingly means “AI independence.”

For Citizens

The Electron Gap is why you should care about energy policy, even if you never touch an AI.
Support nuclear and fusion research: they are climate solutions and AI prerequisites at once.
Question narratives that treat AI as infinitely scalable. It isn’t, at least not yet.

GPT-4 Training: 50,000 MWh and the AI Energy Crisis