The Second Electrification
PriyankaFour years later, his empire was bankrupt. Insull fled the country on a steamer, was dragged back to face trial, and died broken in a Paris Metro station with 85 cents in his pocket. His collapse helped trigger the New Deal, the creation of the SEC, and the Rural Electrification Act -- the federal intervention that finally brought electricity to the 90 percent of American farms that the private market had refused to serve.
That arc -- from private empire to public backlash to eventual democratization -- took about fifty years to play out. We may be starting the second revolution now.
The 13x Gap
Here is the reality of AI's energy appetite, stripped of hype.
Global electricity generation in 2024 totaled roughly 30,000 terawatt-hours. Data centers consumed about 460 TWh of that -- approximately 1.5 percent of global production. That sounds modest. It is not where this is heading.
The International Energy Agency projects data center consumption will roughly double to 1,000 TWh by 2030. Gartner forecasts electricity demand for data centers growing 16 percent annually through the end of the decade. And these are considered conservative projections, because they were modeled before the latest generation of AI training runs began demanding clusters that consume hundreds of megawatts each.
But the real thought experiment is bigger. What happens if AI eventually consumes 20 to 30 percent of the world's electricity?
Twenty percent of global generation is 6,000 TWh. Thirty percent is 9,000 TWh. The current data center footprint would need to multiply by a factor of 13 to 20. That gap -- between where we are and where the most aggressive AI scaling trajectories point -- is the central fact of this story.
Six thousand terawatt-hours is roughly equal to the combined electricity output of the United States and the European Union. Nine thousand is roughly the entire output of China. We are talking about conjuring an additional civilization's worth of electricity, and feeding it all into machines that think.
A Pattern We've Seen Before
The first electrification of America began in 1882, when Thomas Edison's Pearl Street Station lit up a square mile of lower Manhattan. It took another fifty years before electricity reached most American homes.
That half-century was not a smooth diffusion curve. It was a story of concentrated private power, geographic inequality, political conflict, and eventually, forced redistribution.
The early decades followed a familiar logic: electricity went where it was profitable. Cities got power first. Industry got it second. Factories in the Northeast and Midwest electrified rapidly, and the productivity gains were enormous -- manufacturing output per worker roughly doubled between 1900 and 1930 as factories converted from steam to electric drive. The economic case was overwhelming.
But rural America was left in the dark. Stringing wire to remote farms was expensive and the customer base was sparse. Private utilities refused to extend service. By 1930, nearly 90 percent of urban homes had electricity. Only 10 percent of farms did. Farmers milked cows by hand, pumped water manually, and lit their homes with kerosene. They could see the electrical future transforming the cities. They simply couldn't access it.
The man who profited most from this arrangement was Samuel Insull. Starting as Thomas Edison's personal secretary, Insull moved to Chicago in 1892 and began assembling what would become the largest utility empire in American history. His insight was simple and powerful: electricity had massive economies of scale. Bigger plants serving bigger territories produced cheaper power. So Insull built bigger, acquired competitors, and layered holding company upon holding company until his empire served millions of customers across a third of the country.
Insull was genuinely brilliant. He pioneered off-peak pricing, demand management, and the regulatory compact that defined American utilities for a century. He also extracted enormous personal wealth from a system that left rural communities without power and gave urban consumers limited choice.
When the Depression hit, Insull's pyramid of leveraged holding companies collapsed. Investors lost everything. The political backlash was ferocious. Insull's fall became a symbol of unchecked private power -- literally and figuratively. Congress passed the Public Utility Holding Company Act of 1935 to break up the empires. Roosevelt signed the Rural Electrification Act in 1936, providing federal loans to string wire to farms. By 1942, half of American farms had electricity. By 1950, 87 percent did. By the mid-1950s, electrification was essentially universal.
The timeline is worth sitting with. Edison's Pearl Street Station: 1882. Universal rural electrification: roughly 1955. Seventy-three years from invention to broad access. And it required a depression, a political revolution, and direct federal intervention to close the gap.
Insull's Ghosts
The parallels to 2026 are uncomfortable in their precision.
Today's hyperscalers -- Microsoft, Google, Amazon, Meta, Oracle -- are assembling the energy infrastructure for AI at a speed and scale that would have made Insull envious. In 2026 alone, these five companies are projected to spend $600 to $700 billion on capital expenditures, with roughly 75 percent -- about $450 billion -- directed at AI infrastructure. To fund this, they raised $108 billion in debt in 2025 and are projected to issue $1.5 trillion more.
They are not merely buying servers. They are becoming energy companies. Meta signed agreements for 2,609 megawatts of nuclear power in the first quarter of 2026 alone, including a 1,200-megawatt advanced nuclear energy park. Microsoft has invested in nuclear fusion research. Amazon has been acquiring nuclear-adjacent real estate. These are not side projects. They are the core of the business.
Like Insull, the hyperscalers have discovered that controlling the energy supply is controlling the means of production. Like Insull, they are building vertically integrated empires -- from reactor to chip to model to product. Like Insull, they are extracting favorable rates and regulatory treatment that smaller players and ordinary consumers cannot access.
And like the first electrification, the benefits are concentrating geographically and economically before they spread.
The Loophole That Congress Forgot
Here is the part of this story that almost nobody is talking about.
After Insull's collapse, Congress didn't just express concern. It passed one of the most aggressive structural regulations in American history. The Public Utility Holding Company Act of 1935 -- PUHCA -- was designed to do one thing: prevent any company from building the kind of vertically integrated energy empire that Insull had assembled. Under PUHCA, holding companies that owned utilities faced strict geographic restrictions, mandatory SEC registration, disclosure requirements, and corporate structure limitations. The law essentially banned the business model that had made Insull possible.
PUHCA worked. For seventy years, the American electricity system was fragmented by design. No single company could control generation, transmission, and retail distribution across multiple states the way Insull had. Utilities were boring, regulated, local. That was the point.
Then, in 2005, Congress repealed PUHCA as part of the Energy Policy Act. The rationale was market efficiency. Deregulation was fashionable. The old holding-company abuses seemed like ancient history. The SEC's oversight authority over energy holding companies was eliminated.
Nobody in 2005 was imagining that, twenty years later, technology companies would become the largest energy buyers on Earth -- and would begin vertically integrating from nuclear reactor to consumer product in ways that Insull never dreamed of.
Consider what Meta is actually building. It now has power purchase agreements with nuclear operators, is funding the construction of new reactors, and is the end consumer of all that electricity -- funneling it into AI models that generate revenue through advertising and consumer products. It is simultaneously the generator, the transmission customer, and the consumer. It is, structurally, a holding company that spans the energy stack. The same is true of Microsoft, Amazon, and Google, each of which is signing generation agreements or investing in energy production at scale.
The law designed to prevent the next Insull was repealed just in time for the next Insull to arrive. Except this time, it's five Insulls, they're worth more than most countries, and they showed up wearing hoodies instead of three-piece suits. The regulatory framework doesn't see them as utilities. It sees them as technology companies that happen to buy a lot of electricity. The distinction is structural. The consequence may not be.
The Compute Standard
But the regulatory gap is only the domestic story. The international story is stranger and possibly more consequential.
In July 1944, delegates from 44 nations gathered at the Mount Washington Hotel in Bretton Woods, New Hampshire, and designed the postwar monetary order. The system they built had a simple architecture: the U.S. dollar would be convertible to gold at $35 per ounce, and every other currency would peg to the dollar. America had the gold. Everyone else had to come to America to get it.
The Bretton Woods system lasted 27 years. It collapsed in 1971 when Nixon severed the dollar-gold link. But for nearly three decades, it defined who had power and who didn't. Nations with dollar reserves could trade, invest, and project influence. Nations without them were clients of those that did.
Something structurally similar is emerging now -- except the reserve asset is not gold or dollars. It is compute.
The ability to train and run frontier AI models requires three things: advanced chips (overwhelmingly manufactured by NVIDIA and fabricated by TSMC in Taiwan), massive energy infrastructure, and the engineering talent to wire it all together. In 2026, these three inputs are concentrated in a way that makes the early Bretton Woods gold distribution look democratic by comparison. The United States controls the chip design. Taiwan controls the fabrication. A handful of American corporations control the largest compute clusters on Earth. And the U.S. government controls who is allowed to buy the chips at all.
This is not a metaphor. In April 2026, the U.S. imposed a cap of 75,000 H200 chips per year on sales to China, plus a 25 percent tariff. Broader export controls, first imposed in October 2022 and tightened repeatedly since, restrict AI chip sales to dozens of countries based on a tiered classification system. Tier 1 nations -- close U.S. allies -- get relatively open access. Tier 2 nations face caps. Tier 3 nations are effectively embargoed.
This is, functionally, capital controls for the compute age. The United States is deciding which nations can accumulate the reserve asset and which cannot. The mechanism is different -- chip quotas rather than currency pegs -- but the structure is identical to Bretton Woods. The nation that controls the strategic commodity sets the terms for everyone else.
The Sovereign Compute Race
The rest of the world has noticed. And the response looks exactly like what happens when countries realize they're on the wrong side of a reserve-currency arrangement: they start trying to build their own reserves.
Saudi Arabia has committed $100 billion to AI infrastructure, including data centers, chip design partnerships, and a sovereign AI fund. This is not a technology investment in the conventional sense. It is a strategic reserve play. The Kingdom is converting oil wealth -- the reserve asset of the 20th century -- into compute capacity, the reserve asset it believes will define the 21st. The UAE is running a parallel strategy, with Abu Dhabi's sovereign wealth funds backing AI infrastructure across the Gulf.
The European Union has launched a sovereign AI infrastructure stack -- a policy framework designed to ensure that European AI models can be trained on European compute, using European energy, under European data sovereignty rules. France, with its nuclear baseload, is positioning itself as the anchor of this effort.
India and Japan announced an AI Cooperation Initiative -- pooling resources to build shared compute infrastructure that reduces dependence on American hyperscalers. India's government has separately committed to building national AI compute capacity through a public portal aimed at democratizing access domestically.
The pattern is unmistakable. Nation after nation is treating AI compute the way Cold War states treated uranium enrichment: as a capability so strategically important that dependence on a foreign supplier is an unacceptable vulnerability.
The Dependency Trap
The nations that cannot build sovereign compute face a trap that historians of monetary systems will recognize immediately.
Under Bretton Woods, developing nations that lacked dollar reserves had two options: earn dollars through exports (subordinating their economies to American consumer demand) or borrow dollars (subordinating their fiscal sovereignty to American creditors). Either way, the dependency was structural. It shaped trade policy, foreign policy, and domestic politics for decades.
The compute dependency is structurally identical. A nation that cannot train or run AI models domestically must rent that capability from someone who can -- almost certainly an American hyperscaler. That means its AI-dependent industries (and eventually, most industries will qualify) operate on infrastructure controlled by a foreign corporation, subject to foreign law, and priced in a foreign currency. Sensitive national data -- health records, financial systems, military logistics -- flows through foreign servers.
African and South Asian researchers have already begun calling this "the new digital colonialism." The term is provocative but the structure is precise. When a nation's economic infrastructure depends on a resource it cannot produce domestically and cannot substitute, that nation is, in the ways that matter, a client state of whoever controls the resource. Oil created this dynamic in the 20th century. Compute may create it in the 21st.
The difference is that oil is a commodity. Markets are liquid, suppliers are numerous, and alternatives (solar, wind, nuclear) exist. Frontier AI compute is not a commodity. It requires specific chips from specific manufacturers, fabricated in specific foundries, assembled in specific facilities, powered by specific energy infrastructure, and operated by specific engineering teams. The supply chain has single points of failure at nearly every layer. This makes compute dependency more brittle and more coercive than oil dependency ever was.
The Geopolitical Redraw
If AI energy demand continues to scale, the global power map redraws along energy lines. The winners are countries with cheap, abundant, reliable electricity and political willingness to build.
The United States has massive land, diverse energy sources, and dominant tech companies. But its grid is fragmented and permitting is slow. PJM Interconnection, the grid operator serving 67 million people across 13 states, is already facing what analysts call a "data center demand crisis." Data centers are consuming 6.7 to 12 percent of total grid capacity in some regions -- and the number is growing fast.
France, with roughly 70 percent of its electricity generated by nuclear power, has a head start that no other European country can match. It could become the continent's AI energy hub almost by accident.
The Gulf States -- particularly the UAE and Saudi Arabia -- combine cheap solar and gas resources with sovereign wealth funds large enough to build entire AI cities from scratch. They are already doing so.
Canada and the Nordics offer cold climates that reduce cooling costs, hydroelectric baseload, and political stability. Canada is seeing a data center construction boom.
The losers are energy-poor, import-dependent nations. Japan, South Korea, and most of Southeast Asia face a stark choice: build massive new generation capacity or become permanently dependent on foreign compute. Several Southeast Asian nations -- Vietnam, the Philippines, Thailand -- are now revisiting nuclear power plans specifically to attract AI data centers. The AI energy race is pulling countries toward nuclear that had previously rejected it.
Sub-Saharan Africa, South Asia, and parts of Latin America risk near-total exclusion from hosting AI infrastructure, deepening an already severe digital divide. The geography of AI power may calcify into a new map of global haves and have-nots -- this time defined not by who controls oil, but by who can generate and deliver cheap electrons.
The Crowding-Out Problem
The most immediate tension is domestic, not geopolitical. Every megawatt consumed by a data center is a megawatt that doesn't flow to a hospital, a factory, or a household.
This is already visible. In the PJM service territory, data centers are securing long-term power purchase agreements at favorable bulk rates while residential electricity bills climb. In northern Virginia -- the densest data center market on Earth -- local utilities have warned that they cannot add new customers fast enough to match demand. Some new industrial and residential developments have been delayed because the grid simply cannot support them alongside the data center load.
Scale this up. At 20 percent of global electricity, data centers would consume roughly as much power as the entire United States does today. Even if total generation grows, the competition for transmission capacity, water resources, and grid interconnection points would be fierce.
Water is a binding constraint that receives too little attention. A single large data center can consume one to five million gallons of water per day for cooling. Industry projections show a 170 percent increase in data center water consumption by 2030 -- and that is based on current growth trajectories, not the aggressive scenarios we are considering. In water-stressed regions -- the American Southwest, India, the Middle East, parts of China -- AI data centers would be competing directly with agriculture and municipal water supplies.
The political implications are explosive. Imagine explaining to voters in Phoenix that their water bill is rising because a tech company in San Francisco needs to cool servers that generate AI-written marketing copy. The optics are, to put it mildly, challenging.
The Social Contract Question
This is the deepest layer of the parallel. The first electrification generated enormous aggregate wealth -- but it took decades and a political revolution before that wealth was broadly shared. The same question now hangs over AI.
If 20 to 30 percent of global electricity goes to AI, the total capital reallocation is staggering. Generating capacity alone would require something on the order of $7.6 to $11.4 trillion in nuclear construction, or equivalent investments in solar, wind, and storage. Add transmission infrastructure, grid upgrades, water systems, and the data centers themselves, and the total investment required lands somewhere between $15 and $25 trillion over 15 to 20 years.
That is not abstract money. It is steel, concrete, copper, and human labor redirected from other uses. It is capital that does not go to housing, healthcare, transportation, or education. It is a civilizational bet that AI will generate enough value to justify the reallocation.
The bet pays off if AI delivers transformative productivity gains: automating routine knowledge work, accelerating drug discovery and materials science, making government services dramatically more efficient, reducing costs across the economy. In that scenario, the energy investment looks like the original electrification -- expensive upfront, with returns that compound for a century.
The bet fails if AI remains primarily a tool for generating marketing content, chatbot conversations, and modest productivity improvements for white-collar workers. In that scenario, we have redirected a meaningful fraction of the world's energy infrastructure to serve the profit margins of five companies. The political backlash would make the post-Insull reforms look gentle.
History suggests the answer will be somewhere in between, and that the distribution of benefits will be the decisive variable. The first electrification was transformative in aggregate -- but it took the Rural Electrification Act to ensure that transformation reached beyond the cities and the wealthy. Some equivalent intervention will likely be necessary for AI.
When Bretton Woods Breaks
Not tomorrow. The IEA's most aggressive projections put data centers at three to four percent of global electricity by 2030. Getting to 20 or 30 percent would require sustained exponential growth in AI compute demand beyond current projections, massive breakthroughs in energy infrastructure deployment, and -- crucially -- economic returns from AI large enough to justify the investment.
A rough timeline, if it happens at all: 2040 to 2050 at the earliest. More likely, we settle somewhere in the five to ten percent range by the mid-2030s -- which is already transformative enough to trigger most of the dynamics described above, just at lower intensity.
But the Bretton Woods parallel suggests a more specific prediction. The original system didn't erode gradually. It snapped. By the late 1960s, the gap between America's gold reserves and the dollars held overseas had become untenable. France started demanding gold for its dollars. Other nations followed. Nixon had to choose between deflating the American economy to defend the peg or breaking the system. He broke it.
The compute standard may face its own Nixon moment. Today, U.S. chip export controls and hyperscaler dominance give America effective control over the global compute supply. But sovereign compute programs -- Saudi Arabia's $100 billion bet, the EU's sovereign AI stack, India-Japan cooperation -- are the equivalent of France demanding gold. They are nations signaling that they will not accept permanent dependence on America's compute reserves.
If enough nations build enough sovereign capacity, the American-controlled compute standard fractures into competing blocs -- just as Bretton Woods gave way to floating exchange rates. The question is not whether this happens, but when. And whether the transition is managed -- like the Plaza Accord -- or chaotic, like the 1997 Asian financial crisis.
Is artificial intelligence a general-purpose technology on the order of electricity itself, or is it something less?
If it is the former, compute is the reserve currency of the 21st century, and the fights over who controls it -- fights over power in every sense of the word -- will define the next half-century of geopolitics.
If it is the latter, we will look back at this era the way we look at the fiber-optic bubble of the late 1990s: as a period of dramatic overinvestment that created useful infrastructure but bankrupted many of its builders.
Samuel Insull would understand both outcomes. He lived through the first version of this story. He built an empire on the conviction that electricity would change everything, and he was right. He just didn't survive long enough to see who got to enjoy it.