Wake up: The scale of what’s being built defies easy comprehension. So does the silence about what comes next.
There is a question sitting at the center of the most consequential infrastructure buildout since the interstate highway system, and almost nobody in American public life is asking it clearly. The question is not whether artificial intelligence is transformative. That debate is over. The question is this: if we are building a new national infrastructure layer, one that converts electricity into cognition at industrial scale, what social contract will make it politically sustainable when the cognition it produces begins replacing the labor that pays the mortgages?
The difficulty in framing this question is that it lives at the intersection of three different oceans, each enormous on its own, and the people paid to explain them to you tend to specialize in only one. The first ocean is physical: the campuses, the buildings, the sheer acreage being cleared and poured. The second is electrical: the gigawatts, the grid interconnections, the water and cooling demands that are already straining utilities from Virginia to Texas. The third is social: the labor displacement, the income distribution, the political fallout that history tells us accompanies every major technological upheaval, and that this time may arrive faster than any government is prepared to absorb.
This article tries to grab all three at once. It will not succeed entirely. But the attempt matters, because the people who should be making this attempt (the press, the Congress, the think tanks, the billionaires funding the buildout) are largely failing to connect these threads for the public that will live with the consequences.
I. The Boom: Not the Largest Buildings, but Among the Largest Machines
A natural instinct when confronting the AI data-center explosion is to reach for architectural superlatives. Are these the biggest structures humans have ever built? The answer is no, and the distinction matters.
The Boeing Everett Factory in Washington State remains the world’s largest building by volume, at roughly 472 million cubic feet. The New Century Global Center in Chengdu, China, holds the record for total floor area at about 1.7 million square meters. These are single enclosed structures. The AI data-center campuses now under construction are something different: sprawling industrial sites that contain multiple buildings alongside substations, cooling infrastructure, gas plants, parking, and access roads.
Epoch AI has measured three representative examples using satellite imagery. The xAI Colossus campus in Memphis spans about 0.39 square kilometers, with the main building covering roughly 0.07 square kilometers. OpenAI’s Stargate campus in Abilene, Texas, encompasses approximately 3.5 square kilometers of land, comparable in area to Central Park, but its eight planned IT buildings total only about 0.36 square kilometers. Meta’s Hyperion campus in Richland Parish, Louisiana, dwarfs both: after a quiet additional land purchase of roughly 1,400 acres reported by Fortune, the combined site exceeds 11 square kilometers, making it nearly four times the size of Central Park.
These are not the largest buildings ever conceived. They are among the largest machines. The distinction is critical because what makes them unprecedented is not their rooflines but their appetites, for electricity, for water, for grid capacity, and for capital. Meta’s Hyperion project alone may ultimately cost $50 billion, according to remarks President Trump attributed to Mark Zuckerberg at a Cabinet meeting. Entergy plans to build three new natural gas plants just to feed it.
When Zuckerberg announced Meta Compute, a top-level initiative to secure computing power for the company’s drive toward what it calls superintelligence, he wrote that Meta is “planning to build tens of gigawatts this decade, and hundreds of gigawatts or more over time.” A single gigawatt is enough to power roughly 700,000 to one million homes. Tens of gigawatts is not a building project. It is a civilization-scale energy commitment.
II. The Electrical Reality: A Thousand Percent and an Asterisk
The most authoritative accounting of U.S. data-center electricity consumption comes from Lawrence Berkeley National Laboratory’s December 2024 report, commissioned by the Department of Energy under the Energy Act of 2020.
The numbers tell a story of a trend that held steady for a decade and then broke violently upward. U.S. data-center electricity use sat at roughly 60 terawatt-hours in 2014, stayed nearly flat through 2016, then began climbing as GPU-accelerated servers entered the installed base at scale. By 2018, consumption had reached about 76 TWh, or 1.9 percent of total U.S. electricity. By 2023, it had more than doubled to 176 TWh, 4.4 percent of total U.S. electricity consumption.
Berkeley Lab projects a 2028 range of 325 to 580 TWh, which would represent 6.7 to 12 percent of total national electricity demand. At roughly 50 percent capacity utilization, that translates to a power demand of 74 to 132 gigawatts. The compound annual growth rate accelerated from about 7 percent (2014 to 2018) to 18 percent (2018 to 2023) and could range from 13 to 27 percent through 2028.
The International Energy Agency reports that globally, data centers consumed approximately 415 TWh in 2024, about 1.5 percent of world electricity consumption, and projects that figure to more than double to 945 TWh by 2030, slightly more than Japan’s total annual electricity consumption today.
Reuters, drawing on state and local filings tracked by Cleanview, reports that more than 150 gigawatts of new data-center power capacity have been proposed across the United States, spread across 24 states with at least one gigawatt each. Current total U.S. data-center capacity sits just under 15 gigawatts. That is a proposed increase of roughly 1,000 percent.
But Reuters immediately appends the asterisk that makes this number honest: many of these proposals from hyperscalers “are more optimistic visions rather than firm commitments and might be retracted if development timelines or grid connections face delays.” The pipeline is aspirational. The grid is physical. And the gap between the two is where the real story lives.
Virginia, already the nation’s data-center hub, anticipates an eleven-fold increase in computing capacity, with nearly 35 GW in development. Texas follows with approximately 27 GW planned. Gas-rich Pennsylvania ranks third at about 14 GW. Bloom Energy’s 2026 power report estimates that total U.S. data-center power demand could roughly double from about 80 GW in 2025 to 150 GW by 2028. By 2030, about one in five data-center campuses are expected to exceed gigawatt scale, rising to one in three by 2035.
The speed mismatch is the bottleneck. Developers in Texas who responded to a Bloom Energy survey typically expected to be connected to the grid about a year before the utility expected to have that power available. This is not a planning problem. It is a physics problem dressed in legal and regulatory clothing. Transmission lines, substations, and generation capacity operate on timelines measured in years to decades. AI product cycles operate on timelines measured in months.
This mismatch is already showing up in household electricity bills. In the PJM electricity market stretching from Illinois to North Carolina, data centers accounted for an estimated $9.3 billion price increase in the 2025 to 26 capacity market. The average residential bill is expected to rise by $18 a month in western Maryland and $16 a month in Ohio. A Carnegie Mellon study estimates that data centers and cryptocurrency mining could lead to an 8 percent increase in the average U.S. electricity bill by 2030, potentially exceeding 25 percent in the highest-demand markets of central and northern Virginia.
III. The Money: $650 Billion This Year, $5.2 Trillion by 2030
The capital being deployed is difficult to contextualize because the numbers have outrun the analogies. Alphabet, Amazon, Meta, and Microsoft are collectively forecasting approximately $650 billion in capital expenditures for their 2026 fiscal years, roughly a 67 percent increase over their combined $381 billion in 2025. Amazon alone expects its 2026 capex to reach $200 billion. Google is aiming for $175 to $185 billion. Meta estimates $115 to $135 billion.
Gartner projects worldwide AI spending to total $2.5 trillion in 2026, a 44 percent increase over 2025, with AI infrastructure absorbing $1.37 trillion of that figure. By 2027, Gartner expects AI spending to exceed $3.3 trillion.
McKinsey estimates that total investment in AI-ready data-center infrastructure will need to reach $5.2 trillion by 2030 to meet projected demand. JLL’s 2026 Global Data Center Outlook puts the investment supercycle at up to $3 trillion over five years, including $1.2 trillion in real estate asset value creation and approximately $870 billion in new debt financing.
For historical perspective: adjusted for inflation, the entire U.S. interstate highway system cost roughly $500 to $600 billion. At peak, railroad investment consumed 7 to 10 percent of GDP. AI infrastructure spending in 2024 was estimated at about $500 billion, or roughly 1.6 percent of GDP, large, but not yet at railroad-mania levels relative to the economy.
Whether this spending is warranted or reckless depends on which constraint you believe is binding. If the constraint is physical, not enough compute to run the models that enterprises and governments want to deploy, then the buildout is rational even if some individual projects fail. JLL reports global data-center occupancy at 97 percent, with more than three-quarters of the construction pipeline already pre-committed. That does not look like a bubble in the traditional sense.
But a Bank of America survey of 162 fund managers found that 35 percent now say corporations are overinvesting in capital expenditures, a record proportion over the past 20 years of the survey. Nearly one-third of surveyed investors identified hyperscaler capex as the most probable cause of a systemic credit event. Morgan Stanley and Moody’s peg cumulative spending at $3 trillion or more; JPMorgan projects more than $5 trillion including related power supplies. The buildout can be physically necessary and still financially mispriced, especially if model demand, energy prices, or regulation break one of the assumptions baked into today’s capital plans.
The historical analog that keeps surfacing is the railroad boom. One Chinese analysis frames it bluntly: “Technological breakthroughs ignite imagination, capital influx spurs prosperity, over-construction leads to oversupply, and finally, a reshuffle occurs in the burst of the bubble.” From 1865 to 1873, total U.S. railroad investment reached $2 billion against a total national income of about $9 billion, more than one-fifth of the national savings poured into rail construction. The crash that followed was devastating. But by 1900, the cost of rail travel had dropped to one-tenth of its 1870 level and freight efficiency had increased fivefold. The infrastructure survived the bubble. Many of the investors did not.
IV. The Jobs Question: Not Whether, but How Fast and for Whom
Now we arrive at the part of the conversation that the people building these campuses would prefer to have quietly, if at all.
The International Monetary Fund warned in January 2024 that AI is poised to affect nearly 40 percent of jobs worldwide. In advanced economies, the exposure rises to about 60 percent. In roughly half those cases, workers can expect productivity gains from AI integration. In the other half, AI will perform key tasks currently carried out by humans, potentially lowering demand for labor, suppressing wages, and eliminating positions outright. IMF Managing Director Kristalina Georgieva stated plainly: “In most scenarios, AI will likely worsen overall inequality.”
The World Economic Forum’s Future of Jobs Report 2025 projected that by 2030, AI and robotics will displace 92 million jobs globally while creating 170 million new ones, a net gain of 78 million, if you believe the creation numbers will materialize on schedule. Goldman Sachs estimated in 2023 that AI could replace the equivalent of 300 million full-time jobs, while potentially raising global GDP by 7 percent.
These numbers are large enough to be almost meaningless in the abstract. What makes them concrete is the emerging evidence on who gets hit first.
Anthropic CEO Dario Amodei warned in mid-2025 that AI could wipe out half of all entry-level white-collar jobs and spike unemployment to 10 to 20 percent within one to five years. Microsoft AI chief Mustafa Suleyman gave it 18 months for AI to be capable of performing all white-collar work. Venture capitalist Vinod Khosla predicted in February 2026 that AI will handle 80 percent of economically valuable work, and urged radical tax shifts to cushion the displacement.
These are not outside critics. These are the people building and funding the systems.
The measured evidence so far is less dramatic than the predictions but points in the same direction. PIMCO reported in February 2026 that entry-level hiring stalled in 2025 in the most AI-exposed sectors, and cumulative U.S. employment in those sectors has already declined by over 1 percent since 2022, versus a 4 percent increase in other sectors. Morgan Stanley’s AI adoption survey found that AI had led to the elimination of 11 percent of jobs among respondent firms, with an additional 12 percent left unfilled, partially offset by 18 percent new hires, a net 4 percent job loss globally. The study noted this was “an unexpected outcome as previous analysis indicated that AI would have a positive effect on employment growth.”
Challenger, Gray & Christmas tracked nearly 55,000 job cuts attributed to AI in 2025, with warnings that the figure may be underreported. Harvard Business Review published research in early 2026 finding that companies are laying off workers because of AI’s potential, not its performance, with 29 percent of surveyed firms hiring fewer people in anticipation of future AI capabilities and 37 percent committed to replacing workers by end of 2026.
The NBER study of customer support agents provides one of the clearest empirical windows into what happens when AI is actually deployed at scale in a specific workplace: productivity increased by about 14 percent on average, with a 34 percent improvement for the least experienced workers and minimal impact on the most experienced. The implication is double-edged. AI makes the junior worker almost as productive as the senior worker. That sounds like democratization until you realize it means the company needs fewer junior workers.
V. Why Nobody Is Asking the Question Every Day
Here is a peculiar fact. A Reuters/Ipsos poll from August 2025 found that 71 percent of Americans worry that AI will permanently put many people out of work. A Gallup survey from December 2025 found that 73 percent of Americans believe AI will reduce the total number of jobs in the United States over the next decade. A Pew Research Center survey found that 52 percent of workers say they are worried about the future impact of AI in the workplace.
And yet this issue does not dominate the national political conversation. It does not lead newscasts. It does not produce sustained legislative action. Congress has introduced bills (the AI-Related Job Impacts Clarity Act, assorted proposals for workforce retraining and regulatory guardrails) but little has become law amid partisan gridlock and tight margins. A bipartisan 250-page House AI Task Force report released in December 2024 laid out recommendations. Almost none of it has been enacted.
Why?
The simplest explanation is experiential. Most American workers have not yet felt AI replace them personally. Pew found that 63 percent of workers say they do not use AI much or at all in their jobs. Gallup reported that only 15 percent of employees say it is very or somewhat likely that automation, robots, or AI will eliminate their job within five years, unchanged from 2023 despite surging AI adoption. A UC Merced study found that even when people are told AI-driven job displacement could arrive by 2026, their personal concern does not meaningfully increase.
There is a cognitive pattern here that Princeton economists Anne Case and Angus Deaton mapped in their work on deaths of despair. The slow-rolling catastrophe (deindustrialization, the decline of unions, the outsourcing of blue-collar work) did not produce a political earthquake in real time. It produced a decades-long erosion of community, wages, health, and meaning that showed up in suicide rates, opioid overdoses, and alcoholism long before it showed up in voting patterns. Case and Deaton described it not as a sudden shock but as “a slow-rolling catastrophe” in which “whole towns have closed; social life has been disrupted.”
The AI displacement may follow a similar path, but faster. The threat is legible to poll respondents as an abstraction. It has not yet become a lived daily experience for most people. And the institutions that should be translating the abstraction into urgency (the press, the political parties, the labor movement) are themselves under siege from the same forces. Newsrooms are cutting staff. AI-generated content is flooding information channels. The attention economy rewards conflict and spectacle over the slow, structural story of a labor market being re-engineered underneath the feet of the people who depend on it.
The AFL-CIO’s president, Liz Shuler, is trying to organize a response. The Atlantic reported in February 2026 that she is “trying, and mostly failing” to get CEOs and tech leaders to engage seriously with the workforce implications of the AI race they are running. The CEOs and tech leaders are, in the Atlantic’s framing, “so focused on winning the AI race” that the question of who loses is treated as a secondary concern.
VI. Universal Basic Income: The Answer That Is Not an Answer
Into this vacuum steps the idea of universal basic income, championed with varying degrees of sincerity by several of the same people building the systems that will produce the displacement.
Sam Altman, CEO of OpenAI, has called for “universal extreme wealth for all,” unlocked by AI. Elon Musk has imagined a future where “probably none of us will have a job” but we will all have “universal high income.” Anthropic’s Dario Amodei has warned that AI will act as a “general labor substitute for humans.” Andrew Yang, who made UBI the centerpiece of his 2020 presidential campaign, argues that “the argument for universal basic income strengthens continuously as AI accelerates.”
The empirical evidence on unconditional cash transfers is more encouraging than critics typically acknowledge, and less transformative than advocates promise.
Finland’s basic income experiment (2017 to 2018) gave 2,000 unemployed people 560 euros per month with no conditions. The employment effects were small. But recipients were more satisfied with their lives, experienced less mental strain, and had a more positive perception of their economic welfare.
Alaska’s Permanent Fund Dividend, which has provided annual cash payments to all residents since 1982, showed no effect on aggregate employment in a rigorous study using synthetic control methods. Part-time work increased by about 1.8 percentage points, but overall employment did not decline.
The Stockton Economic Empowerment Demonstration gave 125 residents $500 per month for two years. Full-time employment among recipients rose from 28 percent to 40 percent, an increase far exceeding the control group. Recipients also experienced reduced income volatility, better mental health, and greater financial stability.
OpenAI-backed OpenResearch conducted the largest U.S. randomized basic income study, distributing $1,000 per month to low-income individuals in Illinois and Texas over three years. The results defied easy categorization: stress reduction was significant in the first year but did not persist in the data, even as qualitative interviews told overwhelmingly positive stories. Employment shifted modestly (some recipients worked more, some worked less to spend time with family) and the effects largely averaged out in aggregate statistics.
None of this evidence supports the caricature that giving people money makes them stop working. But none of it demonstrates that cash alone solves the structural problems that AI displacement will create.
This is where the critique becomes important. A thoughtful argument against UBI, articulated by the SCORAI network, holds that cash transfers commodify the satisfiers of basic human needs. They say, in essence, that citizens should purchase what they need on the open market, perpetuating the premise that free markets are the best system for meeting human needs. After four decades of evidence that this premise fails the working class, building a new institution on the same foundation does not withstand scrutiny.
The alternative framework is universal basic services: guaranteed access to government-provided health care, education, housing, transportation, and social participation, funded publicly and delivered institutionally. In this model, public money remains in the public sphere, cash payments become less central to household survival, and the political economy of who captures the value of AI is confronted directly rather than papered over with monthly checks.
The uncomfortable truth is that UBI, as currently discussed in Silicon Valley, functions partly as a prophylactic against political unrest, a way for the people capturing the vast majority of AI-generated value to offer a stipend to the displaced without restructuring the systems that concentrate that value in the first place. Albert Wenger of Union Square Ventures, who funds one of the longest-running U.S. UBI trials, frames it as a safeguard against political unrest and argues it could “remove the necessity for minimum wage laws, rent controls, and subsidized housing.” Read that sentence again. The quiet part is loud: UBI as a replacement for the institutional protections that workers spent a century building.
The U.K. is already probing this territory. In early 2026, U.K. Minister for Investment Lord Jason Stockwood told the Financial Times that the government is considering UBI implementation to cushion workers in AI-threatened sectors. Whether this represents serious policy or rhetorical positioning remains to be seen.
VII. The Silence of the Republic
There is a pattern in American history. When a technological revolution displaces labor faster than institutions can adapt, the pain concentrates among people who lack the political power to force a reckoning. The Gilded Age crushed workers for three decades before the Progressive Era produced reforms. The mechanization of the 1920s, combined with speculation and inadequate safety nets, contributed to the conditions that made the Great Depression catastrophic. Deindustrialization hollowed out the Rust Belt over 40 years while economists assured everyone that trade produces net gains.
In each case, the people who saw it coming were dismissed as alarmists or Luddites. In each case, the delay in responding made the eventual adjustment more painful, more politically explosive, and more damaging to democratic legitimacy.
Complexity scientist Peter Turchin has demonstrated that throughout history, an oversupply of elites with bleak economic prospects leads to political instability, a pattern he traces from the late Roman Republic to the present. The AI displacement may produce a novel version of this dynamic: not an oversupply of elites, but a sudden demotion of the professional class that believed its credentials made it safe. When the lawyer, the radiologist, the financial analyst, and the junior software engineer discover that their ladder has been pulled up, the political consequences will not be gentle.
Case and Deaton’s work on deaths of despair offers the most sobering template. They found that “something going on in America is different, and uniquely toxic to the working class”, not just automation and globalization, which other developed nations also experienced without the same epidemic of despair, but a specific American combination of eroded unions, outsourced employment, collapsed communities, and a health care system that functions as an extraction mechanism.
AI displacement lands on top of this already-weakened foundation. The question is not whether it will cause pain. The question is whether the pain will be managed (through retraining, through income support, through institutional investment, through honest public conversation) or whether it will be left to compound in silence until it erupts in ways that make the populist upheavals of the last decade look like a warm-up act.
The press is not asking this question with the frequency or depth it demands. The wealthiest 1 percent, who own the platforms and fund the buildout, have deemed it a secondary concern relative to the competitive race for superintelligence. The public, surveyed and anxious but not yet mobilized, is waiting for someone to connect the dots.
This article cannot connect all of them. But it can name the shape of the thing. We are building a new infrastructure layer of civilization (measured in gigawatts, not square feet) at a pace that outstrips our ability to connect it to the grid, let alone to the social contract. The machines being assembled in Louisiana and Texas and Virginia are not the largest buildings humans have conceived. They are something more consequential: the physical substrate of an economy that may not need most of us in the way the current one does.
The question is what we intend to do about that. And the silence, so far, is deafening.
Related reading
- The $1.65 Trillion Nobody Was Supposed to Find: Read the Footnotes Before Someone Else Buys a Tech Stock in Your Name
- The Data Sovereignty Trap: Why America’s AI Infrastructure Bet May Miss the Point Entirely
- The War That Broke the AI Boom: Economic and Supply Chain Disruptions from the Iran Conflict
- American AI Domination? Not So Fast.
