Bigger than the railroads, on a spreadsheet

A scenario, not a print from the accounts

Van Nieuwerburgh is not reporting what the United States spent. He is pricing a pipeline. A 200 megawatt training campus, in his estimate, costs about $8.2 billion. About a third is the building and the power hookup. About two thirds is the IT system. A gigawatt campus is about $41 billion. A continuous 200 megawatt load uses about as much electricity in a year as 170,000 average American households.

He starts from a project list as of July 2026. About 57 gigawatts already operating. A planned pipeline of about 509. The pipeline is not a forecast. The three-way split is of the post-2024 pipeline, about 527 gigawatts, which includes about 19 gigawatts already operating in 2025 and 2026. 182.8 gigawatts become operational from 2025 through 2032. Another 117 finish later. 227 are never built. That last slice is almost half the planned 509, not half of 527. He then applies the unit cost, lets it rise 4 percent a year, and spreads each project across the completion year and the three years before it. Nominal GDP is grown at the same 4 percent, so the ratio is not being juiced by a stagnant denominator. Money that falls inside the window, including spending on plants that switch on after 2032, comes to $10.3 trillion. That is 3.63 percent of cumulative GDP. The average of the eight annual ratios is the same number. Neither fact means the spending arrives in a smooth line. The sum is also larger than the cash the firms already throw off. That is an old problem, which is why the railroads show up in a book that is not about railroads. The paper that pays for a boom.

The other numbers in the same story

The newspaper that ran the chart also cites an estimate that U.S. AI investment hits 1.9 percent of GDP in 2026, and says the late nineteenth century railroad boom was the last time one new industry took a larger share. The archive text of that article has a hole where the estimator’s name should be. A secondary write-up names Goldman Sachs. The Goldman note that is public puts U.S. AI capex at 1.8 percent of GDP in 2026 and about $581 billion, not 1.9. That note was not the one the archive dropped. The hole stays a hole.

In the draft, 2026 investment is $1.432 trillion against an assumed GDP of $31.992 trillion. That is 4.5 percent, not 1.9. Capital spending at Oracle, Microsoft, Amazon, Meta, and Alphabet is projected to exceed $800 billion in 2026, the first year it would top their combined operating cash flow. Set $800 billion against the draft’s own 2026 GDP and you get about 2.5 percent. That capex includes spending that is not AI, and it excludes data centers the firms lease rather than build. The gap between cash thrown off and cash required is the reason the boom has to be financed with paper. That paper.

A fourth object is what the Commerce Department can already see. Through July 2026, the not-seasonally-adjusted sum for private data-center construction was about $37.7 billion, about $10 billion more than the same seven months a year earlier. The Journal called that sum seasonally adjusted. The seasonally adjusted series is an annual rate. July’s data-center rate was about $79 billion. Private construction as a whole was about $45 billion below the year-earlier seven months. Take the data centers out and the rest of private was about $55 billion below. The Journal’s $46 billion is closer to the decline in private construction as a whole than to everything else. About $38 billion of shells is not the $1.15 trillion scenario cell for 2025. Construction is the third of the campus that is concrete. The chips do not show up in that series.

The chart and the construction print are not in conflict. They are not the same measurement. Collapsing them into “we are already spending 3.63 percent of GDP on AI” is the mistake the chart invites. The Journal says the projection is tricky and that total spending might end up substantially lower.

There is a second version of the author’s own number. The abstract on SSRN still says a 188 gigawatt build-out finished by 2032 would cost nearly $9 trillion, or 3.2 percent of GDP a year. The conference draft the chart uses says 183 gigawatts, $10.3 trillion in the window, 3.63 percent. The unit cost is the same order. The dollar gap is mostly the window. Treat 3.63 as the chart’s figure. Do not treat it as the only figure with the author’s name on it.

How the old booms were counted

The historical bars are not one official series copied six times. The draft says the sources differ in scope and accounting, and that the estimates are approximate measures of scale, not harmonized national accounts.

Canals use Cranmer’s construction estimates, 1836 to 1841. Average 0.66 percent. Peak 0.90 percent in 1840. Six years, and then the boom was over. Railroads use Gallman’s construction series, 1870 to 1890. Average 2.24 percent. The annual share exceeded 4 percent in 1870 and again in 1881. The series leaves out at least some spending on locomotives and rolling stock. If you want a fair fight with an AI figure that is two thirds chips, the railroad number is the one that was trimmed.

Electrification uses Ulmer’s gross capital spending for electric light and power, 1905 to 1925. Average 0.50 percent. Peak 0.96 percent in 1924. It excludes the electrical machinery bought by the people who used the power. Calling this “the grid” in a newspaper deck is fair as a label and unfair as a total. Highways use Federal Highway Administration capital outlays, 1956 to 1973. Average 1.13 percent. Peak 1.32 percent in 1958. Telecom and fiber use BEA private fixed investment in communications equipment and structures, 1996 to 2003. Average 1.10 percent. Peak 1.39 percent in 2000, which is also the year the overbuild became obvious. The fiber is still in the ground. A large share of the companies that laid it are not.

Add all five historical averages and you get 5.63, which is more than 3.63. The Journal’s dek says data-center spending is greater than the canals, the railroads, and the grid combined. Those three averages sum to 3.40, and 3.63 clears that by a quarter of a point. The paper’s table is a different claim. 3.63 is larger than each of the five booms, including highways and telecom. Neither the dek nor the table says the scenario beats the sum of all five. The highway years on this chart, 1956 to 1973, still sit inside an economy that made things. The telecom years, 1996 to 2003, sit after the break. Where that break is dated.

Gross dollars are not the same bet

The draft’s own warning is the paragraph a bubble chart cannot show. Canals, railroads, and fiber often provided service for decades. GPUs may be economically obsolete in three to six years. Per dollar of gross investment, this build-out creates less net capital than many of the earlier booms. Buildings and power connections are long-lived. They are the minority of the cost. The asset the firms are actually trying to keep is not a building. It is the knowledge.

If the economic life of the IT share is three years rather than six, the draft’s own sensitivity says the annual revenue required to earn a 10 percent unlevered return, at a 50 percent cash-flow margin, rises from about $3.7 trillion to about $6.0 trillion. Same buildings. Worse chips.

What has to come back

The revenue exercise is not a forecast of AI sales. The draft says so. It asks what mature annual revenue the 2025 to 2032 completion cohorts would have to produce to recover the investment, replace depreciating capital, and pay a given unlevered return.

That slice is 182.7 gigawatts. Initial cost about $8.54 trillion. Grossed up for the construction period, the capital base at commissioning is about $9.61 trillion. At a 10 percent unlevered return, a six-year life on the IT share, and a twenty-year life on the rest, the cohorts need $1.86 trillion a year of operating cash flow once they are mature. At a 50 percent margin, that is $3.725 trillion of mature annual revenue. Against his assumed 2032 GDP, that is 9.2 percent of the economy, every year, to give this capital an ordinary unlevered return.

In market units the central case is about $5.50 per installed GPU-hour at full utilization, on his representative rack. He says current on-demand prices for high-end Nvidia capacity commonly run from roughly $6 to more than $10 per GPU-hour. At 80 percent utilization the required billed price is about $6.90. On today’s price list, the central case is not absurd. The open question is whether demand absorbs something like 180 gigawatts at high utilization, and whether providers keep the rate as they compete.

The draft puts current combined annual revenue at OpenAI and Anthropic around $100 billion. Reaching $3.725 trillion from that base is roughly 80 percent growth a year. The cash does not have to be booked by those two firms. Hyperscalers and other clouds can be the collectors. Moving the revenue from the lab to the landlord does not shrink it. Who collects is also who can lock the customer. The stack.

Who is actually on the hook

The interesting risk is not “AI works” against “AI does not.” It is that the debt is being written against a stack whose useful life is shorter than the bond, inside vehicles designed so the obligation does not show on the tenant’s balance sheet until the tenant has to pay it.

The five hyperscalers can no longer fund the guided pace out of cash generation. Once capex exceeds operating cash flow, the marginal dollar comes from leases, joint ventures, project debt, private credit, and special-purpose vehicles. Those structures expand the capital that can be raised. They also raise leverage on the asset even when the company’s reported leverage stays low.

Hyperion is about 2 gigawatts and about $30 billion, excluding the IT gear Meta finances as tenant. Meta sold an 80 percent stake to Blue Owl for about $2.5 billion. The joint venture, Beignet, raised $27 billion of debt in October 2025, rated A+, one notch below Meta. The leases are five terms of four years, from 2029 to 2049. At each renewal Meta can walk. If it walks, it covers any shortfall against a contractual minimum value. Under the accounting the draft describes, neither the future lease nor the guarantee is recognized before it becomes effective, even though one of the two will have to be met.

The structure is not cheaper money. Beignet’s debt came at 6.58 percent, at least 100 basis points above what the draft thinks Meta would have paid on comparable unsecured debt. Over the life of the financing that spread is more than $5 billion of extra interest, paid back through the lease. Meta is paying up to keep an asset-light profile. That habit is older than this one deal. Control without the title.

A footnote in the draft cites a separate Journal tally. For four hyperscalers, about $2.4 trillion of off-balance-sheet liabilities against $604 billion recognized on the balance sheet. On that tally, about one fifth of $3 trillion in related obligations is currently recognized. Moody’s, as cited in the draft, puts hyperscaler lease commitments at roughly $970 billion, of which about $660 billion is future leases not yet on the balance sheet.

A further footnote describes a $500 billion credit facility stood up to finance Nvidia chips, with Nvidia providing a 25 percent residual-value guarantee, exposure of as much as $125 billion. The supplier of the short-lived asset is backstopping the lenders who finance the customers who buy the asset. Public prices have already stopped treating data-center landlords as dull infrastructure. The draft’s rolling betas on listed data-center REITs rise from about 0.5 against the S&P 500 to about 1.0.

Van Nieuwerburgh’s own line is narrower than the commentary around him. It would be premature, he writes, to conclude that this already poses systemic risk comparable to earlier credit booms. Off-balance-sheet structures matter because they can make correlated exposures hard to see before a downturn. A coverage ratio of 1.12 times does not need a depression. It needs a delay, a lower renewal, or a residual value that misses the floor.

The paper that pays for a boom

In 1910 Rudolf Hilferding named a stage of capitalism after the paper that large industry had to sell. The oddity underneath it is older. Shareholders own stock, which is a claim on assets and earnings. The corporation owns itself. Railroads, stockyards, and steel mills then cost more than the firms’ existing revenue. The money came from outside, as bonds that promised interest and shares that promised a price and a dividend. Capital markets pooled that paper and put it at the service of the plant. Ben Tarnoff retells this in “How Tech Rules,” a review in the October 22, 2026 issue of the New York Review of Books. The books under review are Cecilia Rikap’s The Rulers and Nick Srnicek’s Silicon Empires. They were not read here. The review was, from an archive snapshot. He is not measuring the chart on this page. He is describing the kind of claim the chart’s dollars become once they cannot be paid from cash on hand.

The five firms in the capex comparison cannot fund the guided pace out of operating cash flow. That fact is already in who is on the hook. Hilferding’s point is what happens next. The obstacle of fragmented savings is removed. The firm draws on the combined capital of everyone willing to hold a story about the future. The story, this time, is the scenario: 182.8 gigawatts, $10.3 trillion inside the window, 3.63 percent of cumulative GDP. A bond holder is not buying that average. He is buying a claim that the average will throw off cash before the chip is obsolete. The mismatch is in asset life. The cash required is in what has to come back.

Tarnoff adds two prices that are not spending. Five companies in California and Washington, he writes, have a combined market capitalization of about $20 trillion. Market capitalization is not the $10.3 trillion. One is a price on shares. The other is an outlay spread across a window. Putting them in one bar would be the mistake this page is trying not to make. They sit next to each other because both are paper about the same plant. He also notes that Elon Musk became the first person marked as a trillionaire by taking SpaceX public, and that a large part of that valuation rests on a promise to put AI data centers in space. That is another story about future compute. It is not a receipt, and it is not in Van Nieuwerburgh’s pipeline.

The highway years on the chart, 1956 to 1973, still sit in the economy that made the things it sold. Tarnoff notes that manufacturing’s share of American employment peaked in 1953, and that the manufacturing workforce kept growing in absolute terms until 1979. The break he dates to that decade is the outsourcing of the physical good and the keeping of the intangible. The telecom years, 1996 to 2003, sit after it. The AI bar is a further step. The intangible is the product. The campus is the cost of holding it. The old planners, J.P. Morgan and John D. Rockefeller in Hilferding’s account, administered whole sectors by holding controlling stakes. Ownership was the instrument. The instrument has changed.

Control without the title

Rikap calls an intellectual monopoly an organization that systematically captures rents from a sustained concentration of intangible assets. Patents and trademarks are the part accountants already name. She is more interested in the rest: data about suppliers, franchises, customers, and workers, and the story a brand can enforce. Coca-Cola fits her definition. So do Disney, Pfizer, and Nike. The useful case here is Nike, because it separates the asset from the plant.

Nike tells contractors what to make and what it will pay. The Swoosh, the designs, and the intellectual property are the value. The shoes are made in other countries, by other companies. Rockefeller planned by owning pieces of many firms. Nike plans without taking on the cost of owning the mills. Rikap’s phrase is control beyond ownership. She reaches for it by way of dependency theory, the tradition Raúl Prebisch helped found. Poor countries stayed poor, in that account, because rich countries could keep extracting value after formal empire ended. Her move is to put Silicon Valley in the core and everyone who produces the knowledge, without holding the asset, in a periphery. Computers made that practical once industry was no longer a single factory. Tarnoff cites Miriam Posner for the further point that a modern supply chain is an artifact of software. Software itself is the clean intangible. A pattern that costs nearly nothing to copy is, in his telling, why Bill Gates was the richest man in the world by 1995. A railroad track does not throw off rent that way. A model can. That is why a firm will finance a $10.3 trillion plant to defend an asset that is cheap to reproduce.

Platforms took the outsourcing one step further. Google and Facebook do not only ship code their own staff wrote. The data is supplied by people searching, liking, and posting, who are not paid for it. Large language models, in Rikap’s suggestion as Tarnoff reports it, are the far end of that move. They are co-produced by everyone connected to the internet, through the training data, while the model remains the firm’s asset.

The split shows up in a pair of counts she gives for Microsoft, from 2012 to 2021. The company co-authored 88 percent of its more than 15,000 scientific publications with other organizations. It shared ownership of 1 percent of its 35,233 patents. The knowledge was produced in common. The patent was not.

Start-ups are the other periphery. Microsoft, Google, and Amazon invest in them and take preferential access to what the smaller firm learns, without the duties of an acquisition. Tarnoff, reading Rikap, says this pattern is what produced the current boom. In 2019 Microsoft made its first investment in OpenAI. OpenAI got the cloud it needed to train the models. Microsoft, in her account, steered research priorities, projects, and direction. ChatGPT arrived in 2022. The next year Microsoft’s stock rose nearly 60 percent, helped by its place as the exclusive reseller of that technology through the cloud.

Set that next to Hyperion. Meta sold 80 percent of the project vehicle and kept the use of the campus, paying a higher rate to stay asset-light. Blue Owl holds a title. Meta holds the option to walk, and the guarantee if it does. The building is in one set of hands. The intangible the building exists to serve is in another. The chart prices the building and the chips. It does not price the reseller right, the patent, or the training set. What the halls are for.

What the halls are for

The cloud, in this review, is not a metaphor for the internet. It is the data centers that began to accumulate in the 2000s so the platforms could run. In 2006 Amazon started selling storage and processing to other companies. Tarnoff reports that Amazon Web Services posted a profit of $45.6 billion last year. The review is in the October 22, 2026 issue, so that year is his, not a line in the Brookings draft, and it was not re-counted here.

Amazon, Google, and Microsoft own the largest of these clouds. Most intellectual monopolies, in Rikap’s claim as he reports it, stay inside one sector. These three do not. They sit on digital production, exchange, and consumption in a way that lets them subordinate organizations across industries, including other intellectual monopolies. Frank Pasquale’s phrase, which Tarnoff borrows, is functional sovereigns.

A lawnmower makes the scale ordinary. A search, a product page, a stock check, a reorder, a payroll run, and the ads that follow the buyer can all pass through the same landlords. Home Depot, he notes in passing, uses Google Cloud. Multiply one Sunday errand by ordinary commerce and the landlords are not only collecting rent. Rikap’s image is a digital twin of the economy, a live map from which the host can read trends the tenant cannot see as a whole. The $8.2 billion campus is one more island in that archipelago. The six-year chip inside it is the part that will have to be bought again. The twin is the part they are trying not to share. Making the archipelago larger now requires the government. The state has to help.

The state has to help

From the 1990s to the middle of the 2010s, Tarnoff writes, following Srnicek, there was a broad agreement he calls the Silicon Valley Consensus. The internet would extend an American-led world of borderless commerce and data. Regulation would stay light. National security would sit behind commercial access. China was a market to enter. In 2015 Mark Zuckerberg went to Beijing and asked Xi Jinping to give his unborn daughter an honorary Chinese name. Xi declined.

Two forces broke that agreement, in this account. Countries began treating exposure to world markets as a risk, after the pandemic, Russia’s invasion of Ukraine, and the US and Israeli attack on Iran, among other shocks. And generative AI turned out to be heavy. It needs data centers, and the centers need energy. The industry cannot raise all of that as a private capex plan. Tarnoff says the Trump administration has opened federal land to data-center construction, cut back environmental review, and offered financial support. Srnicek’s list of what the firms are also seeking abroad is data to train on, energy, water to cool the servers, land, and capital. Each of those needs a state, here or somewhere else.

That is the frame for the smelter that lost the power in Mississippi, in what the spending is already doing. A campus and a furnace wanted the same electricity. The campus got it. The contest is no longer only a local planning board. It is the resource bill of a technology that has recruited the government as a partner. The revenue side recruits the government too. The $3.725 trillion does not have to be booked by OpenAI and Anthropic, but it does have to be booked by someone, and the home market is not assumed to be enough.

The frictions had a timetable, in his telling. Obama announced a Pivot to Asia in 2011. Trump opened a trade war in 2018. Biden added export controls after 2021 and talked about democracies against autocracies. In early 2025, Tarnoff writes, the Biden administration’s AI Diffusion Rule sorted the world into three tiers. Close allies faced the fewest hurdles. Adversaries were embargoed. Everyone else had to apply for a license. More recently Trump reduced tariffs and eased some restrictions on the sale of AI chips, while most of the Biden-era export controls remained. This is his sequence, not a second reading of the rule text.

China is the exception the review will not treat as another contractor. The offshoring of the 1970s, which Rikap ties to the rise of intellectual monopoly, ran alongside Deng Xiaoping’s opening. China took the work, then climbed. Joint ventures, often required of foreign firms, transferred expertise the American side later resented. National champions became intellectual monopolies of their own. Export controls aimed at that climb do two jobs at once, in this account. They hit a rival state and a rival industry.

Hilferding had already watched monopolies outgrow the night-watchman state. High domestic prices, in his argument, pushed them toward foreign markets, and foreign markets came with a more active government. Tarnoff says the 2020s rhyme with that decade more than is comfortable, and then gives Srnicek’s brake. The United States and China are still too bound up to separate cleanly. Interdependence is why he does not treat the rhyme as a timetable for a war. It is not a reason to treat the stack as an ordinary export. Who gets to plan.

Who gets to plan

The chart invites a race. Canals, railroads, the grid, highways, telecom, then a scenario bar. Rikap’s question is who is allowed to decide which bar gets built. Hilferding thought the monopolists of his day had already done the hard work of unifying production. The political question was who would take the unified machine. Tarnoff hears that line in a proposal from Bernie Sanders. A sovereign wealth fund would take equity in the largest AI companies through a onetime tax, and pay citizens an annual dividend of $1,000.

The proposal is a plan for the paper, not a plan for the plant. What has to come back is $3.725 trillion of mature annual revenue in the central case, from a base this page already put near $100 billion at two labs. A large share of the asset dies in three to six years. A dividend that depends on that revenue ties household income to the same mismatch the bond already carries. Tarnoff writes that reasonable people can disagree about tying every household more tightly to Silicon Valley. He still thinks something of that ambition is required.

Rikap’s own demand is not a portfolio. She wants knowledge production governed as a commons. A wider share of the benefits, and a democratic say in what knowledge and which technologies get made. Her reminder, via Tarnoff, is that the choice is not between markets and planning. Someone is already planning. The draft itself throws out 227 gigawatts. That discard is a plan. The question she poses is whether the planners stay the firms chasing the next trillion, or whether anyone else gets a vote on the pipeline this chart is pricing. The things that would tell you which way it is going are in what to watch.

What the spending is already doing

The scenario can be wrong by a trillion dollars and the build-out can still shove other parts of the economy around. That part is reporting, and it should be labeled as reporting.

Data-center shells were up about $10 billion year over year through July 2026, on the not-seasonally-adjusted Census sum. The Journal said about $9 billion, and called the sum seasonally adjusted. Everything else in private construction was down about $55 billion on the current file, not the $46 billion the Journal printed. The $46 billion is closer to the decline in private construction as a whole. The Richmond Fed, the Journal says, has reported that data-center construction is straining labor supply in its district. One industrial story is specific and thinly sourced. Mississippi was in the running for an aluminum smelter and an estimated 1,000 permanent jobs. A data center near Vicksburg tied up the electricity. The smelter went to Oklahoma. The sourcing is a person familiar with the decision. It is still the right kind of example, because the constraint was power. Tarnoff’s review, not the Journal, says the industry is now asking the state to clear that kind of constraint. The state has to help.

LinkedIn, as cited by the Journal, credits AI with more than 750,000 U.S. jobs from 2023 through 2026 to date. The median salary on AI-related listings was around $180,000, against $80,000 for all listings. Employers added 117,000 jobs at data centers since the start of 2024, not counting construction. Household holdings of corporate equities and mutual funds were $63 trillion in the second quarter, the Journal says, citing the Federal Reserve, nearly double the level at the end of 2022. Import prices for computers, peripherals, and semiconductors were 20 percent higher in August than a year earlier. The Journal quotes Fed chairman Kevin Warsh naming hyperscaler borrowing as one reason long-term rates are higher. Electricity bills have risen in some places with a lot of data centers.

What to watch

• Energized gigawatts against announced gigawatts. The draft already discards 227 gigawatts. If cancellations run hotter, the sentence about the railroads becomes a description of a press release.
• The price per GPU-hour as the 2026 and 2027 cohorts come online, and the utilization under that price. A CEO saying demand is strong is not the print. The rate card is.
• The guarantees. Future leases that are not recognized until they start. Residual-value guarantees that appear when a tenant walks or a chip is sold below the floor.
• Whether hyperscaler capex stays above operating cash flow after 2026, and whether the gap is funded with bonds a holder can read or with another project vehicle.
• Which tier a country sits in, and whether it is being sold the whole stack or only a chip license. The sort is in the stack. A press release about an export license is not the same thing.
• Land, power, and credit the state is clearing that ordinary capex could not raise. Federal land and a waived review are a subsidy even when they are not a check.
• Whether the model, the patent, and the reseller right stay with the firm after the building has been sold into a vehicle. That split is the whole of control without the title.

The railroads were a bet on a network that was still there when the bonds came due. A large share of this one is a bet on a chip that will not be. That can still be a sensible use of capital if the cash from the chip, while it lasts, pays for the building and the next chip. It is a different bet than the chart implies. It is being financed as if the difference were an accounting detail. The review adds a second difference. The network being financed is not only the halls. It is a claim to direct other people’s work without holding the title, and a claim on foreign customers that now runs through the state. The halls can be a sensible use of capital and still be that claim.

Sources

1. Van Nieuwerburgh, Financing the AI Buildout. Brookings Papers on Economic Activity conference draft.
2. Putzier and Lahart, The Wall Street Journal. Text read from an archive snapshot.
3. SSRN abstract 6444363. The abstract still prints the other window, 3.2 percent. It was not updated when SSRN marked the paper revised on October 2.
4. ChainCatcher summary of the Journal article. It names Goldman. That name was not used to fill the hole.
5. Financial Express, 18 August 2026, on the Goldman note that prints 1.8 percent of U.S. GDP and about $581 billion.
6. Census, private construction put in place, not seasonally adjusted. The seven-month data-center sum is from this file. The seasonally adjusted file is an annual rate.
7. Tarnoff, How Tech Rules. New York Review of Books, October 22, 2026 issue. Archive snapshot of the review of Rikap, The Rulers, and Srnicek, Silicon Empires. The books were not read here. Figures in the sections on the paper, control, the cloud, and the stack are his account, not a second count of the Brookings draft.

End