AI capital expenditure is about to eclipse the cost of building the railways in the U.S. and the U.K.—with the internet added on top
The modern AI build-out is turning into one of the largest infrastructure programs in economic history. If current trajectories hold, cumulative AI-related capital spending over the next decade looks set to surpass, in real terms, what it took to build the national railway systems of the United States and the United Kingdom—and then keep going, effectively stacking a fresh layer of “digital railways” on top of the broadband internet we already built.
Why this comparison matters
– Railways were the 19th-century skeleton of commerce; the internet (and the telecom networks beneath it) became the 20th/21st-century nervous system. AI now demands a third wave of foundational assets—compute, chips, power, and upgraded networks—large enough to reshape capital markets, industrial policy, and the grid.
– Framing AI’s capex against railways and the internet helps calibrate what “build once in a century” really means.
How to compare costs across centuries
Any cross-era comparison requires judgment. A sensible approach uses three complementary lenses:
1) Inflation-adjusted dollars: Convert historical spend to today’s prices (CPI-based).
2) Share-of-GDP: Compare how large the build was relative to the economy then, and translate that share to today’s economy.
3) Functional scope: Align what’s actually being built (track and rolling stock vs. data centers and chips vs. power and grid).
What the AI build-out actually includes
– Compute: Hyperscale data centers, accelerators/GPUs, storage, networking, cooling.
– Chips and fabs: Leading-edge foundries, packaging, lithography, and the upstream equipment ecosystem.
– Power: New generation (gas, nuclear, wind/solar), grid reinforcements, substations, and transmission—plus on-site generation and backup.
– Connectivity and real estate: Fiber backbones, metro links, subsea capacity, land acquisition and specialized buildings.
The run-rate today—and the likely cumulative bill
– Hyperscaler capex has entered a new regime. Across the largest providers in the U.S. alone, public guidance already points to well over $200 billion per year of total capex, with AI infrastructure the dominant driver and further growth signaled into 2025 and beyond.
– Semiconductors add another layer. Leading-edge fabs and wafer-fab equipment together represent on the order of $100+ billion per year globally, a sizable and rising share of which is explicitly tied to AI demand.
– Power and grid are the sleeper costs. Data center load forecasts in major U.S. markets and the U.K. imply tens of gigawatts of incremental generation and multi-decade grid upgrades—easily translating into hundreds of billions of additional investment.
Even on conservative assumptions—flat to modestly rising annual outlays—cumulative global AI infrastructure investment through the early-to-mid 2030s plausibly lands in the multi-trillion-dollar range. For the U.S. specifically, scenarios that add up data centers, chips, and power system expansions can credibly reach into the low trillions over 10–12 years; the U.K. tally, while smaller in absolute terms, is material relative to its economy and grid.
What the railways actually cost
Estimates vary, but credible historical tallies converge on these orders of magnitude:
– United States: By the 1910s, formal valuations of U.S. steam railroads ran to tens of billions of dollars in then-current money. Adjusted for consumer prices, that is broadly on the order of several hundred billion dollars in today’s terms. Measured as a share of the economy at the time—roughly a quarter to a third of annual GDP at peak—an equivalent share of today’s U.S. GDP would translate to multiple trillions of dollars.
– United Kingdom: By the early 20th century, the domestic railway capital stock stood at roughly a large fraction of one year’s GDP. Mapped to today’s U.K. GDP, the implied replacement value also sits in the high hundreds of billions to low trillions of pounds, depending on methodology.
The internet “on top”
One way to gauge the cost of “building the internet” is to look at telecom and broadband investment since commercialization:
– In the U.S., industry groups estimate broadband and telecom providers have invested on the order of $2 trillion since the mid-1990s to deploy and upgrade internet-capable networks (fiber, cable, mobile, core).
– In the U.K., cumulative broadband and mobile investment is smaller in absolute terms but significant relative to GDP, with full-fiber rollouts and 4G/5G upgrades alone amounting to many tens of billions of pounds over the past decade, and total internet-era telecom capex in the hundreds of billions over the modern period.
The bottom line
– Against inflation-adjusted railway costs: The AI build, if you sum compute, chips, power, and network reinforcements through the 2030s, is on track to exceed the CPI-adjusted historical cost of constructing the railways in the U.S. and in the U.K.
– Against share-of-GDP equivalents: If AI infrastructure spending continues compounding from today’s run-rate, its cumulative scale in the U.S. could rival the share-of-economy footprint that the railways commanded at their zenith—something the internet era did not match in a single, concentrated push.
– With the internet added on top: In the U.S., a plausible AI capex path through the early 2030s can exceed the combined price tag of building the railways (in today’s dollars) plus the roughly $2 trillion poured into the internet era’s broadband and telecom networks. In the U.K., the same stacking logic holds at national scale, given the country’s concentrated grid upgrades, data-center pipeline, and investment in semiconductor and power supply chains.
Why AI’s numbers are so large
– Density of spend: Each cutting-edge data center can cost several billions, much of it for specialized chips and power/cooling. A single state-of-the-art semiconductor fab can cost $20–30+ billion. Transmission lines, substations, and firm generation add billions more per metro area.
– Stacked dependencies: Unlike the internet era—when telco/cable carried most of the capex burden—AI simultaneously pulls on multiple capex-intensive systems: silicon manufacturing, hyperscale campuses, and the electric grid.
– Combinatorial growth: Model sizes, parameter counts, and training intensity still climb; inference is shifting from sporadic to continuous, pushing utilization and power needs higher.
U.S. and U.K. specifics to watch
United States
– Power markets: Data center load growth in PJM, ERCOT, and the Southeast is advancing multi-gigawatt generation and transmission plans, with timelines stretching a decade or more.
– Industrial policy: CHIPS Act incentives have already catalyzed well over $100 billion in fab announcements; utility integrated resource plans are being revised upward for load.
– Real estate and water: Land, cooling, and water rights are becoming binding constraints in key metros, influencing site costs and timelines.
United Kingdom
– Grid constraints: London and the Southeast face tight connection queues; National Grid’s reinforcement plans and offshore wind build-out are central.
– Data center pipeline: Multi-gigawatt campus plans around Greater London and new regional clusters imply tens of billions of pounds in capex this decade.
– Policy levers: Planning reform, grid connection modernization, and power market design will determine pace and cost.
Risks and wildcards
– Productivity payoff: The more AI demonstrably raises output, the easier it is to finance continued build; a weak payoff would slow capex sooner.
– Power and permitting: Delays in generation, transmission, or fab construction could cap growth and shift where dollars land geographically.
– Technology curves: Breakthroughs in model efficiency, chip architecture, or cooling can bend the capex curve down; equally, new workloads (agents, on-device AI) can bend it up.
What this means for decision-makers
– Policymakers: Treat AI infrastructure like a national-scale utility build—align permitting, grid planning, and industrial policy accordingly.
– Utilities: Update load forecasts, accelerate interconnection, and secure firm supply; explore novel procurement with hyperscalers.
– Investors and operators: Expect continued scarcity in power, land, and leading-edge silicon; build optionality into supply chains and site selection.
– Society: As with railways and the internet, broad access and interoperability choices made early will shape who benefits for decades.
The 19th century laid steel rails, the late 20th spun glass fiber. The 2020s are about to bolt silicon and electrons together at a scale that, by the time the current wave crests, will have cost more than either of those past projects—indeed, more than both in the U.S. and the U.K., even with the internet added on top.
