Issue 025 - AI infrastructure - Capital-to-capacity scaling
How much power could Amazon's $220 billion buildout need?
Reuters reported that Amazon's stock jumped after strong AWS growth helped justify a 10% increase in planned 2026 capital spending to about $220 billion. Much of the spending is tied to AI chips, data centers, and supporting infrastructure.
The problem
Estimate the continuous electrical demand that Amazon's $220 billion of 2026 capital spending could eventually add once the resulting data-center equipment is operating.
Convert your result into one-gigawatt power-plant equivalents, average U.S. household equivalents, and annual electricity use.
Does $220 billion imply an expansion comparable to a few large industrial facilities, or an electrical system operating at the scale of cities or states?
Because Fermi problems target an order of magnitude, I normally use no more than two significant digits and write most calculations in scientific notation; the Fermi reference explains both conventions.
Before checking sources
Matt's first pass
I assumed that somewhere between 30% and 50% of that investment will go toward construction of new data and computing centers. That is:
$220 billion x 0.30 ~= $66 billion
$220 billion x 0.50 ~= $110 billionNext, I assumed an average facility cost of about $100 million, which gives:
$66 billion / $100 million ~= 660 facilities
$110 billion / $100 million ~= 1,100 facilitiesThen I assumed each facility consumes energy at an average rate of about the same as 1,000 homes. I also assumed homes consume power at about 1 MW, or 1,000 to 1,200 kW, so about 1 GW per facility.
power ~= 660 to 1,100 facilities x 1 GW/facility
~= 660 to 1,100 GWThat means the new energy consumption resulting from the new facilities could be in the range of 660 to 1,100 GW, or about the same consumption as about 660,000 to 1.1 million homes.
Calibration Score
Matt's Calibration Score: 25 / 100
Higher is better: earn points for accurate pegs, sound models, correct math, and a result close to the sourced answer. The image shows percent full of it: 100 minus the Calibration Score.
Pegs: 0/30. The household-power and facility-cost pegs were off by orders of magnitude.
Model: 15/30. Capital-to-capacity was the right family of model, but the assumed facility unit size distorted the calculation.
Math: 0/10. The MW/GW and household-power mistakes substantially changed the answer.
Result: 10/30. The final intuition that this was city-scale survived, but the numeric path was far off.
Grounding facts
The useful new peg is that an average U.S. household is about 1 kW continuous, not 1 MW. That single correction fixes many grid-scale intuitions.
A 4-GW always-on load is smaller than the entire U.S. grid, but it is not "just a few buildings." It is closer to adding the household electricity use of several million homes, and it requires generation, transmission, substations, backup systems, water/cooling plans, and long utility lead times.
After checking sources
Check and recalibrate
The main correction is the household peg. The EIA says the average U.S. residential customer used about 10,791 kWh/year in 2022. Converted to continuous power:
average household power ~= 1.1 x 10^4 kWh/year / 8.8 x 10^3 hours/year
~= 1.2 kWSo a thousand homes average about:
1,000 homes x 1.2 kW/home ~= 1,200 kW
~= 1.2 MWThat is about a thousandfold below 1 GW. A 1-GW data center is not "1,000 homes"; it is closer to a million homes.
The second correction is facility cost. A $100 million data center is not plausibly a 1-GW facility. JLL's 2026 outlook gives about $11.3 million per MW for shell-and-core construction, before AI technology fit-out that can add up to $25 million per MW. For an AI-heavy build, use an all-in cost near:
all-in AI data-center cost ~= $30M to $40M per MWNow estimate how much of Amazon's $220 billion turns into new data-center capacity. If 30% to 70% is relevant, that is:
AI/data-center capital ~= $220B x 0.30 to 0.70
~= $66B to $154BConvert capital to power capacity:
power capacity ~= capital / cost per MW
low case ~= $66B / ($40M/MW)
~= 1.7 x 10^3 MW
~= 1.7 GW
central case ~= $130B / ($35M/MW)
~= 3.7 x 10^3 MW
~= 3.7 GW
high case ~= $154B / ($30M/MW)
~= 5.1 x 10^3 MW
~= 5.1 GWA reasonable answer is therefore roughly 2 to 6 GW of continuous electrical demand, with a central estimate around 4 GW. If nearly all $220 billion eventually becomes AI data-center capacity, the number could rise toward 6 to 8 GW, but hundreds of GW would require much cheaper capacity or much more capital than this spending plan implies.
Convert 4 GW into power-plant equivalents:
power plants ~= 4 GW / 1 GW per large plant
~= 4 large power plantsConvert it into household equivalents:
households ~= 4 x 10^9 W / 1.2 x 10^3 W/home
~= 3.3 x 10^6 homesConvert it into annual electricity use:
annual electricity ~= 4 GW x 8.8 x 10^3 h/year
~= 3.5 x 10^4 GWh/year
~= 35 TWh/yearUsing the 2-to-6-GW range, the annual electricity range is roughly 18 to 53 TWh/year. That is not national-grid scale, but it is absolutely city-to-state scale: several large power plants, millions of homes, and a meaningful chunk of new electricity demand.
Post-check reflection
Matt's reflection
I totally blew it on household power, and then on the energy consumption of a single facility, by several orders of magnitude on both. My mental pegs really failed me.
In the end, it is not very surprising to me or contrastive with my mental models that Amazon's intended investment might represent an equivalent energy consumption to a few million homes. That seems about right.
Recommended memory peg
Remember 1 GW running continuously uses about 8.8 TWh/year and powers roughly 800,000 to 900,000 average U.S. homes. For AI data centers, use $30M to $40M per MW as an all-in capital-to-power peg when chips and fit-out are included.
Reader results
Bars show how submitted estimates sort into the answer choices from the gut-check prompt.