01 / NATIONAL SCALE
Start with the wave.
A draft projects $10.3 trillion of US AI infrastructure spending across eight years. Historical investment episodes help put that scale in context, with their different coverage kept visible.
TWO ERAS. TWO BETS ON THE FUTURE.PYRAMIDS
PYRAMIDS
vs AI.
An ancient building programme.
A national AI investment wave.
What does it cost to build the future?
EXPLORE THE COMPARISON A national AI investment wave.
What does it cost to build the future?
LABOUR · CAPITAL · POWER
Different resources.
Shared ambition.
Shared ambition.
Labour capacity and GDP
are different denominators.
are different denominators.
US AI WAVE · DRAFT SCENARIO
3.63%
Projected annual investment / GDP
2025–2032 · national build-out
2025–2032 · national build-out
01 / WHAT DOES THE FUTURE COST?
Shared ambition.
Different claims on society.
Egypt organised labour around royal monuments. The AI wave coordinates capital, power and expertise around useful computation. The revealing comparison is what each programme demands—and what it promises in return.
The AI scenario is a large national investment wave. Egypt’s pyramid programme was a sustained mobilisation of labour, concentrated in its largest building dynasty. We can estimate each allocation. Labour capacity and investment / GDP have different denominators, so these figures cannot establish which society sacrificed more.
02 / NATIONAL PROGRAMME
Count the monuments.
A volume-based scenario extends the Khufu labour model across a named inventory of royal pyramids. Switching between the Fourth Dynasty and the wider Old Kingdom reveals how much the averaging window matters.
03 / PROJECT SCALE
Open the model.
Then inspect one pyramid complex and one campus. Change workforce, population, seasons and construction time; follow the capital replacement and power requirements beyond the initial build.
02 / NATIONAL INVESTMENT · HISTORICAL CONTEXT
An AI wave on
a national scale.
One Brookings conference draft projects $10.3 trillion across 2025–2032—chips, campuses, power and networks. Its aggregate investment / GDP share is:
3.63%
PROJECTED PERIOD AVERAGE · GROSS INVESTMENT / US GDP
September 4, 2026 draft, presented September 24–25. A scenario subject to revision.
PERIOD AVERAGE · INVESTMENT / OUTPUTDRAFT TABLE 1
Period averages reproduced from Van Nieuwerburgh, Table 1 ↗. They indicate historical scale; periods, asset coverage and accounting differ.
THE COMPARISON THE TABLE SUPPORTS
3.63% projected versus 2.24% historical average. The eight-year AI scenario exceeds the selected 1870–1890 railroad construction average.
THE PEAKS AND COVERAGE MATTER
Railroad shares exceeded 4% in 1870 and 1881. That series omits some equipment; the AI forecast covers a broader asset mix. This is not evidence of a record over historical peaks.
$10.279tn ÷ 8 = $1.285tn per year; $10.279tn ÷ $283.448tn of cumulative GDP = 3.63%. The draft models roughly 183 GW completed by 2032 plus spending on later completions. Gross spending excludes subsequent hardware refreshes. Tables 1 & 5; Appendix C ↗
03 / EGYPT · FROM A PROJECT TO A PROGRAMME
A dynasty’s work.
A kingdom’s long average.
Add the monuments before annualising the work. Their volumes give a rough way to scale a Khufu-sized labour budget across the royal building programme.
FOURTH DYNASTY · c. 2575–2465 BCE
Seven principal pyramids. One century.
Sneferu’s three pyramids, Khufu, Djedefre, Khafre and Menkaure: about 8.73 million m³ of original monument volume.
450,000 × (8.73m ÷ 2.58m) × 1.25 ÷ 110
≈ 17,282 labour-years per year
≈ 17,282 labour-years per year
Khufu’s 450,000 labour-years are the lab’s base annual crew × 20 years. The 25% addition is an assumed allowance for unlisted projects and differences in associated works. The starting calibration already includes quarrying, transport and off-site support.
BASE · ANNUAL ADULT LABOUR CAPACITY1.73%
About 1.90 million total labour-years, spread over 110 years and one million adults per year.
SCENARIOS: 0.77% LOW / 5.87% STRESSThe same inventory method gives about 0.43% across the wider Old Kingdom. A long average smooths away the Fourth Dynasty concentration; it does not describe the busiest year.
| Assumption / result | Low | Base | Stress |
|---|---|---|---|
| Khufu calibration · total labour-years | 275,000 | 450,000 | 630,000 |
| Additional programme allowance | 0% | 25% | 50% |
| Annual adult capacity, held constant | 1,100,000 | 1,000,000 | 495,000 |
| Programme labour-years / year | 8,449 | 17,282 | 29,034 |
| Annual labour share | 0.77% | 1.73% | 5.87% |
A programme proxy, not a state budget. Volume is only a rough labour driver: height, materials, bedrock, reuse and temple craftsmanship change effort per cubic metre. The allowance covers missing work only hypothetically. Population and productivity are held constant; non-royal construction and ongoing cult operations are outside the estimate. The stress case deliberately stacks a large workforce, small population and large allowance.
Inspect the monument inventory and scaling assumptions
Original geometric volumes are drawn from Mark Lehner’s The Complete Pyramids (1997), “Major Pyramid Statistics,” p. 17, using the linked table transcription. The inventory includes royal pyramids with a reported volume in Dynasties 3–6; unquantified projects and non-pyramidal royal tombs are excluded from the volume sum. The allowance is our assumption, not a measured correction.
| Monument | Dynasty | Original volume · m³ |
|---|---|---|
| Sneferu · Meidum | 4 | 638,733 |
| Sneferu · Bent Pyramid | 4 | 1,237,040 |
| Sneferu · Red Pyramid | 4 | 1,694,000 |
| Khufu | 4 | 2,583,283 |
| Djedefre | 4 | 131,043 |
| Khafre | 4 | 2,211,096 |
| Menkaure | 4 | 235,183 |
8,730,378 m³ ÷ Khufu’s 2,583,283 m³ = 3.38 Khufu volumes
The Met’s chronology supplies c. 2575–2465 BCE for the Fourth Dynasty and c. 2649–2130 BCE for the Old Kingdom. These are averaging windows, not an annual construction schedule. Volume inventory ↗ · Chronology ↗
National programmes put scale in view. Now open one project’s assumptions to see how those estimates are assembled.
04 / PROJECT SCALE · EXPLICIT CLOCKS.
A complex.
A campus. Per year.
A complex and a campus both depend on wider infrastructure. Annualising their construction keeps each year’s commitment in view.
EGYPT / KHUFU’S COMPLEX
A monument and its network
- Pyramid, associated works and local services
- Quarrying, transport and harbour operations
- Off-site food, tools and supply labour
22,500/yr × 20 yrs= 450,000 TOTAL LABOUR-YEARS · BASE CASE
First estimate the annual work: 10,000 permanent + 5,000 seasonal + 7,500 off-site = 22,500 labour-years per year. Multiplying that annual estimate by a conventional 20-year build gives the total.
UNITED STATES / MODELLED 1 GW CAMPUS
Compute and its infrastructure
- Data-centre buildings and cooling
- Associated power infrastructure
- Servers, accelerators and networking
$41bn ÷ 4ILLUSTRATIVE CAMPUS COST ÷ ASSUMED BUILD YEARS
$10.25bn per year. In constant 2025-dollar terms, that is about 0.033% of the paper’s $30.762tn US GDP baseline. The four-year construction window is adjustable in the lab.
SMALL IN THE ECONOMY. LARGE ON THE GRID.
A 1 GW campus running at full electrical load would use 8.76 TWh/year: about 0.20% of 2025 US utility-scale electricity generation, alongside the benchmark’s 0.033% of GDP in annual construction investment. A small national capital share can still create a concentrated local power demand. EIA denominator ↗
8.76 ÷ 4,429 TWh × 100; national generation is a scale reference, not spare grid capacity. The lab adjusts average electrical load.05 / THE FERMI LAB · INTERACTIVE
Make the assumptions
do the work.
Choose a case, then test it. The hatched band marks the low, base and stress cases at the 20-year reference. The summary stays visible through this lab and the population controls below. Weighted cost discounts only the seasonal recruits’ alternative work.
SOURCE-ANCHORED INPUTS / EXPLICIT ASSUMPTIONS
How much labour did a pyramid take?
One labour-year means one person’s normal working year. A peak crew is present for only part of it. The base case begins at 22,500 labour-years annually.
NOT A STATISTICAL INTERVAL
Base presetPresets stay fixed; adjustments create a custom case.
THE PYRAMID CONSTRUCTION CLOCK
More years—or less work each year?
The crew model estimates annual labour first. A different question holds its derived total work fixed and spreads that budget over a shorter or longer build.
15 years30 years
22,500 / year × 20 years = 450,000 labour-years
Fixed annual crew: 22,500 labour-years each year. Changing duration changes total work; the annual share stays at 2.25%.
At this population, spreading the fixed 450,000-labour-year reference budget over 15–30 years gives 1.50–3.00% annually.
EGYPT · ANNUAL ADULT LABOUR CAPACITY2.25%22,500 labour-years ÷ 1,000,000 adult labour-years
0%5%10%
Hatched: 1.25–6.36% at the 20-year reference.
Base preset: within the 20-year reference envelope.ONE AI CAMPUS · ANNUAL GROSS INVESTMENT / GDP0.033%$10.25bn per year ÷ $30,762bn US GDP
At 2025 prices and GDP. Labour share tracks people’s time; investment / GDP tracks capital spending relative to annual output.
MAGNIFY ONE 0.5% GDP TILE
One campus, magnified.
0.033% of GDP fills about 1 of these 15 fragments.
Each fragment is 0.0333 percentage points. This tile uses US GDP. Egypt’s hatched band uses adult labour capacity; their percentages describe different resources.
EXPLORE THE PARTS / BUILD THE ARGUMENT
Assemble the estimate.
Select a layer or step through the calculation. The parts follow your current assumptions.
SELECT A SEGMENT · ANNUAL LABOUR-YEARS
050,000 / yr
LABOUR LAYERS · NOT CONSTRUCTION PHASES
01 / 04
PERMANENT CREW
10,000labour-years per yearA year-round core includes local services as well as the workforce on the monument.
10,000 people × one normal working year
Source anchor: Redding’s provisioning model ↗ANNUAL TOTAL22,500labour-years / year
OVER 20 YEARS450,000total labour-years
Base case: 10,000 permanent workers, 25,000 total peak workers for 4 months, plus a 50% off-site allowance.
LENS 01 / THE SECOND CAPITAL CLOCK
Building once. Replacing repeatedly.
IT replacement allowance: $4.65bn/yrConstruction averages spread the first bill. They do not pay for future equipment. Following the paper’s rounded split, 68% buys IT systems and 32% buys facilities and power. Its economic-life assumptions are six years for IT and twenty for the rest.
3 years8 years
$27.88bn IT · $13.12bn facilities and power
ANNUAL IT REPLACEMENT ALLOWANCE$4.65bn/yr
$41bn × 68% ÷ 6 years
A separate 20-year allowance for the other assets: $0.66bn/yr.
Twenty years of capital.
INITIAL BUILD$41bn
IT REFRESHES2
CUMULATIVE GROSS SPEND$96.76bn
ONE CAMPUS / BILLIONS OF 2025 US DOLLARS
PYRAMID / 20-YEAR BUILD · SEPARATE CLOCK AND UNIT
22,500 labour-years repeated across each of 20 construction years. A constant model year, not a reconstruction of annual work.
Four years to build, then 16 operating years. Refreshes in years 10 and 16 from the start of construction.
$41bn initial + 2 × $27.88bn IT = $96.76bn
At the $41bn benchmark, 20 operating years with five-year IT replacement totals $124.64bn; four-year replacement totals $152.52bn. The initial build is included in those costs; construction time sits before that operating clock. Purchases exactly at the end are excluded. Future equipment prices, performance and resale value can change; this is a constant-cost replacement illustration, excluding operation and financing.
At the $41bn benchmark, a four-year IT cycle implies $6.97bn/year; six years implies $4.65bn. Annual allowances spread constant-cost replacement over time. The twenty-year view instead counts whole refreshes. Construction spending and subsequent replacement belong to their respective years. The national build-out estimate excludes later hardware refreshes. Appendices A, C and D ↗
LENS 02 / THE PHYSICAL LENS
Put the energy on a common clock.
A common physical unit helps, but it answers a different question: human effort during pyramid construction versus electricity during campus operation. These are partial energy accounts, not whole-project footprints.
HUMAN METABOLISM · INPUT DURING WORK
1.41 MWCALENDAR-AVERAGE EQUIVALENT
12.38 GWh per year at 275 W per worker during work.
1,500 hours3,000 hours
At a 75 W mechanical-output benchmark: 3.38 GWh/year, equivalent to 0.39 MW averaged across the calendar.
CAMPUS ELECTRICITY · OPERATING INPUT
800 MWAVERAGE ELECTRICAL LOAD
7,008 GWh per year at 80% of the 1,000 MW capacity.
50% load100% load
The 1 GW capacity includes cooling and other facility overhead. At full electrical load for a year it would use 8.76 TWh. Electrical load is not GPU utilisation; 80% is an illustrative setting, not a measured operating rate.
22,500 labour-years × 2,000 h × 275 W ÷ 1 billion = 12.38 GWh/year
12,375 MWh ÷ 8,760 hours = 1.41 MW calendar average
FAO gives 250–300 W metabolic consumption during sustained work and up to 75 W of sustained mechanical output. We use 275 W and 75 W as uniform conversion assumptions, not measured pyramid-worker rates. The default 2,000 h = 250 × 8 h is also an assumption. Off-duty metabolism, animals, fuel, materials and upstream energy are excluded; the mechanical figure is output, not extra energy to add to the metabolic input. Power source ↗
LENS 03 / SURPLUS EXTRACTION · SENSITIVITY ONLYA small share of the whole.
A small share of the whole.
A large claim on what is available.
Feeding households and maintaining farms already occupied much of an agrarian year. Test a narrower question: if only a fraction of adult labour could be redirected after essential work, how much of that pool would the pyramid claim?
5% sensitivity40% sensitivity
Choosing a fraction activates the discretionary view; 15% is the starting assumption.
SHARE OF ALL ADULT LABOUR CAPACITY2.25%
22,500 ÷ 1,000,000 × 100 = 2.25%
With seasonal opportunity weighting: 1.88% of this denominator.
This is a labour-capacity thought experiment, not a grain balance or a measured state tax budget. Household surplus was not all available to the crown, and off-site provisioning was not all additional production. For AI, investor budgets and local power capacity identify who bears the constraint more precisely than national GDP. This lens leaves the main labour and GDP readouts unchanged.
Inspect the low, base and stress assumptions
These combinations are ours. Redding’s excavation-based provisioning model supplies an anchor of 8,000–10,000 workers outside the flood season and 25,000 during four months. Evidence from the later Giza workers’ town is used here as a proxy for Khufu’s project, not a direct Khufu payroll. The off-site uplift avoids counting local bakers and administrators twice; its size is uncertain.
| Input or result | Low | Base | Stress |
|---|---|---|---|
| Year-round core, including local services | 8,000 | 10,000 | 12,000 |
| Total peak workforce | 20,000 | 25,000 | 30,000 |
| Months at peak | 3 | 4 | 4 |
| Extra off-site labour / on-site labour | 25% | 50% | 75% |
| Population | 2,000,000 | 2,000,000 | 1,100,000 |
| Working-age adult fraction (assumption) | 55% | 50% | 45% |
| Annual adult labour capacity | 1,100,000 | 1,000,000 | 495,000 |
| Annual pyramid labour-years | 13,750 | 22,500 | 31,500 |
| 20-year total labour-years | 275,000 | 450,000 | 630,000 |
| Annual labour share | 1.25% | 2.25% | 6.36% |
| Fixed-work share · 30-year to 15-year build | 0.83–1.67% | 1.50–3.00% | 4.24–8.48% |
Annual on-site labour = core + (peak − core) × months / 12
Annual total = on-site labour × (1 + off-site allowance)
Annual share = total ÷ (population × adult fraction)
Reference total = annual total × 20
Fixed-work sensitivity = reference total ÷ chosen build years
Population anchors: roughly 2 million in Warden’s Old Kingdom account; 1.1 million in Butzer’s estimate cited by Redding. The 45–55% adult fractions, support uplifts and scenario combinations are sensitivity assumptions. The denominator includes adults of all genders and domestic work, without claiming they were all available for conscription.
Limits: the annual-crew mode repeats a representative year; the fixed-work mode reallocates a derived reference total. Actual work varied by phase, and 15–30 years is a sensitivity window rather than a dated construction interval. Labour-years omit differences in skill, animal power, capital and land. Off-site provisioning is a gross allocation, not all extra food production. A smaller adult denominator raises the estimate; fewer working months and lower seasonal opportunity cost reduce it.
Base case: 22,500 labour-years annually, or 2.25% of adult capacity.
06 / TWO EFFECTS THAT PULL APART
Fewer adults.
Quieter seasons.
A headcount fraction answers who turned up. Labour-years answer how much time was used. Opportunity cost asks what else those people could have done.
1%20,000 ÷ 2 million people
2%20,000 ÷ 1 million adults
That example is headcount only. The shared model above also allows for seasonality and supply work.
↓ Lower alternative value lowers effective cost
Flood-season time was not free.
Reduced farming opportunities could make seasonal recruits less costly to divert. They still needed food, and alternatives such as maintenance, transport and household work remained. Test the value of forgone alternatives; do not subtract this from actual labour used.
0% · no alternative100% · full value
18,750 alternative-work-equivalent years annually, versus 22,500 actual labour-years.
Only the extra seasonal recruits receive this weight. Permanent and off-site labour stay at full weight. The 25% default is a sensitivity choice, not a measured lost harvest or GDP estimate.
↑ Fewer adults raises the labour share
ADULT LABOUR CAPACITY1 TILE = 0.5%
Hatched fill: 2.25% actual labour. Light underline: 1.88% opportunity-weighted. Partial tiles preserve both shares.
2,000,000 × 50% = 1,000,000 adult labour-years of capacity per year.
The summary above updates as you change these controls. All adults are assigned one normal working year; age structure and participation remain uncertain. Children’s work and unequal working time are not modelled.
07 / WHAT DID SOCIETY GET?
Durability is not
the same as return.
A monument can last without producing saleable services. A machine can wear out while producing something valuable. The question is who benefits, and at what cost.
EGYPT / RELIGION, LEGITIMACY & CAPACITY
A state made visible.
The intended return was a royal afterlife and the authority of kingship. Excavation reveals the administration and supply networks required to serve that religious and political purpose.
POSSIBLE SPILLOVERS · NOT A MEASURED ROI
- Logistics, craft knowledge and coordination
- Provisioning and redistribution, with unequal benefits
- Labour and surplus diverted from other uses
The stone monument did not need a recurring hardware refresh, but rituals, provisioning and its surrounding institutions still used resources. The benefits of that organisation and the burdens it imposed were unevenly shared.
AI / SERVICES, PRODUCTIVITY & CASH FLOW
Useful work must follow.
AI campuses aim to sell computation and support productive services. A customer-support study found a 15% productivity gain, but one firm’s result does not establish the return on a national infrastructure boom.
ASSETS WITH DIFFERENT LIVES
- Buildings and power infrastructure can serve for decades
- Accelerator economics can change within a few years
- Revenue must cover operation, replacement and capital; the lab makes that replacement allowance explicit
Social returns also depend on access, distribution and environmental costs. A large investment is a bet on demand, not evidence that the bet has paid off.
Evidence: AERA’s workers’ town excavations; Brynjolfsson, Li & Raymond’s workplace study; Brookings on asset lives. The comparison of social returns is our interpretation.
08 / THE RETURN IS THE TEST
Building the future
is a bet on
what comes after.
The AI draft describes a national capital wave. The pyramid model describes labour mobilisation, with the Fourth Dynasty’s concentration visible inside the much lower Old Kingdom average. Scope and time change the story before any comparison of returns begins.
Egypt’s rulers sought a royal afterlife and an enduring sacred order. AI’s investors seek useful computation and higher productivity. Each required present resources, specialised skills and institutions capable of organising them.
The test is whether the promised return justifies what society gives up—and who gets to decide. A monument’s survival and a machine’s output answer different parts of that question.
PRIMARY EVIDENCE / ASSUMPTIONS / INTERPRETATION
Follow the evidence.
Sources accessed in October 2026; this edition is dated 24 October. The Brookings figures reproduce a selected conference draft and may change in later revisions. Archaeological observations, authors’ models and our sensitivity assumptions are identified separately.
Van Nieuwerburgh · Brookings conference draftDraft dated September 4, 2026; conference September 24–25. Tables 1, 2 and 5; Appendices A–D. Forecasts and assumptions may change. Historical series have different coverage; railroad peaks exceeded 4%.↗
Richard Redding · A Tale of Two Sites (2013)Author-uploaded chapter, pp. 315–319. Seasonal provisioning model and a separate herding workforce. These are model inputs, not a complete worker census.↗
AERA · Lost City of the Pyramid BuildersMark Lehner’s excavation project: settlement, workshops, supply systems and questions about state development. Attribution to Khufu is less direct than for later Giza kings.↗
Pierre Tallet · The logbook of Merer (2017)Primary papyri edition and English translation. Records Tura-to-Giza limestone shipments, provisions and dyke work during Khufu’s reign.↗
Leslie Anne Warden · Webs of Power (2015)Penn Museum. Roughly 2 million people in Old Kingdom Egypt; complex relationships between royal and local power.↗
Brynjolfsson, Li & Raymond · Generative AI at WorkQuarterly Journal of Economics, 2025. Study of 5,172 customer-support agents. Productivity gains vary; not a national ROI estimate.↗
IEA · Energy and AI (2025)Executive summary: grid connection queues and transformer/cable lead times. Used for the infrastructure bottleneck parallel, not an ancient constraint ranking.↗
US EIA · Electricity generation, capacity and sales2025 utility-scale generation: approximately 4,429 TWh, excluding small-scale solar. Used as a fixed annual reference for campus electricity, not available grid headroom.↗
FAO · Energy for AgricultureSection 2.4. Distinguishes human metabolic consumption from mechanical output. Work hours and the uniform conversion across tasks are this page’s assumptions.↗
Mark Lehner · The Complete Pyramids (1997), p. 17“Major Pyramid Statistics.” Link opens an attributed table transcription; source is Lehner’s archaeological synthesis. Volumes are geometric estimates. Scaling them into programme labour and adding 0–50% for other works are our assumptions.↗
Tara Prakash · Egypt in the Old KingdomThe Metropolitan Museum of Art, 2019. Fourth Dynasty monuments and approximate reign chronology. The 110-year Fourth Dynasty and 519-year Old Kingdom windows follow this chronology. Programme labour estimates are ours.↗
What changed in this edition?
The opening now leads with the national AI scenario and removes the labour-to-GDP scorecard. Railroad peaks and differences in coverage sit beside the averages. A named monument inventory replaces the constant-workforce dynasty extrapolation, with Fourth Dynasty and wider Old Kingdom views. The lab tests 15–30-year pyramid builds under both fixed-crew and fixed-work assumptions. “High” is now “Stress,” and the conclusion makes no cross-unit ranking. Brookings is explicitly identified as a draft subject to revision.