Skip to content
LumenaryAcademy
Frontier 80 min

Powering the AI data centre

How much electricity AI needs, how it travels from the grid to a transistor at under a volt, how the heat gets out again, and how data centres are connecting to — and starting to help — the grid.

Code tested · 5AI-draftedAwaiting review by an electrical engineer

Your guides: Chloe · The Physicist, Margaret · The Bench Engineer, Gideon · The Design Reviewer

MotivationH

A gigawatt for a campus

A large data centre of 2015 drew perhaps 20–50 megawatts. AI campuses announced in 2024–26 are planned at hundreds of megawatts to several gigawatts each — the output of one or more large power stations, at a single site, running around the clock. In Hardware for AI you met a rack that draws 120 kW and a chip that draws over a kilowatt at less than one volt. This lesson follows that power from the transmission grid to the transistor, and the heat from the transistor back out to the air, and then looks outward: how much electricity data centres use, how they get connected to grids that were not planned for them, what they are signing up for, and how they might help the grid rather than strain it.

It draws on almost every lesson of the Power and energy level: three-phase distribution, transformers, power electronics, grid stability and storage.

Note

Everything in this lesson upstream of a server's circuit board — medium-voltage switchgear, transformers, UPS systems, busways and battery rooms — carries lethal voltages and enormous fault energies. It is designed, built and maintained only by trained, authorised electricians, with equipment isolated and locked off before anyone works on it. The calculations here are to understand the design, never a guide to touching it.

ExplanationH

Demand: what is measured and what is projected

The most cited global estimate is the International Energy Agency's Energy and AI report of April 2025 (IEA). It put data centres' electricity use at about 415 TWh in 2024, roughly 1.5% of world electricity, growing about 12% a year over the previous five years; the United States accounted for about 45%, China 25% and Europe 15%. Its base case reaches about 945 TWh by 2030, a little under 3% of world electricity, with AI-optimised servers the main driver, and its scenarios for 2035 span roughly 700 to 1,700 TWh. For the United States, a December 2024 report from Lawrence Berkeley National Laboratory (Shehabi et al.) estimated 176 TWh in 2023, 4.4% of the country's electricity, and a range of 325 to 580 TWh (6.7–12%) by 2028.

Three cautions about such numbers:

  • The range is wide for good reasons. It depends on how fast chips become more efficient, how many are manufactured (limited by HBM and packaging), whether demand for AI services keeps growing, and how much financing continues. Projections made in 2023 were revised sharply upwards in 2024; some may be revised down.
  • Requests are not demand. Developers often ask several utilities for connections for the same project, so utility queues overstate what will be built.
  • Globally modest, locally enormous. A few per cent of world electricity is less than, say, the growth expected from electric vehicles or air conditioning. But data centres cluster: Ireland's Central Statistics Office has reported that they used about a fifth of the country's metered electricity in 2023, and northern Virginia, Dublin, Amsterdam, Frankfurt and Singapore have all restricted or delayed new connections at times.
DefinitionsH

From the grid to the transistor

transmission grid110–400 kV ACsubstationto 10–35 kVtransformer andswitchboard, 400–480 VUPSwith batteriesbackup generatorsbusway and PDUto the racksrack power shelfAC to ~50 V DC48→12 V, thenregulators to ~0.8 Vchip~1,000 Aeach step down in voltage is a step up in current for the same power
The power chain of a large AI data centre (simplified, one path of many in parallel). Voltage falls in steps from hundreds of kilovolts to under one volt; currents rise in proportion. Diesel or gas generators and the UPS batteries carry the load through a grid outage.

Following the figure:

  1. Transmission and the substation. A large campus connects to the transmission grid (110–400 kV, depending on the country) through its own substation, whose transformers step down to a medium voltage of 10–35 kV for distribution around the site — the subject of The power grid and Transformers and machines.
  2. Low-voltage distribution. Transformers near each data hall step down to three-phase low voltage — 400 or 415 V in most of the world, 480 V in North America — feeding switchboards.
  3. The UPS. An uninterruptible power supply bridges the seconds to minutes between a grid fault and the backup generators starting. A double-conversion UPS rectifies to DC, charges a battery and inverts back to AC continuously, costing a few per cent of efficiency; many operators now run "eco" modes that bypass the conversion until a fault occurs. Lithium-ion has largely replaced lead–acid batteries in new builds.
  4. To the rack. Busways or power distribution units carry the power to each rack. In AI racks, power shelves rectify AC into DC at about 50 V (the "48 V" family; NVIDIA's NVL72 uses 54 V) and a vertical busbar carries it to every tray.
  5. On the board. An intermediate bus converter steps 48 V down to about 12 V (or lower), and a multiphase voltage regulator module (VRM) — dozens of interleaved buck converters, from Power electronics — produces the 0.7–0.9 V the chip needs, at over a thousand amperes.

Why so many steps, and why does the industry keep raising the rack voltage? The next section computes it.

ComputationC

The last centimetre: why I²R rules

Delivering power PP at voltage VV means a current I=P/VI = P/V, and a conductor of resistance RR wastes I2R=P2R/V2I^2 R = P^2 R / V^2. For a fixed conductor, the loss falls with the square of the voltage. Halving the voltage quadruples the loss, and the chip insists on less than one volt.

python
# 1. The last centimetre: deliver 1,000 W to a chip through a path of fixed resistance.
P = 1000.0                      # W
R = 0.5e-3                      # ohms: 0.5 milliohm of copper, connectors and package
print("supply voltage   current     I^2 R loss    loss as % of delivered power")
for V in (0.8, 1.8, 12.0, 48.0, 800.0):
    I = P / V
    loss = I**2 * R
    print(f"  {V:6.1f} V     {I:8.1f} A   {loss:10.3f} W     {100 * loss / P:8.3f}%")

# 2. The same physics at rack scale: 120 kW over a busbar of 0.1 milliohm.
P_rack, R_bus = 120e3, 0.1e-3
for V in (54.0, 800.0):
    I = P_rack / V
    print(f"rack at {V:5.0f} V: {I:7.0f} A, busbar loss {I**2 * R_bus:6.1f} W")

# 3. A load step: the chip's current jumps by 500 A in 1 microsecond.
L_path = 50e-12                 # henries: 50 pH of loop inductance between regulator and die
dI, dt = 500.0, 1e-6
print(f"\nL di/dt = {L_path * dI / dt * 1e3:.0f} mV of droop on a 0.8 V supply "
      f"({100 * L_path * dI / dt / 0.8:.1f}% of it) unless capacitors supply the charge meanwhile")

# 4. The chain of conversions from medium voltage to the die (illustrative efficiencies).
stages = [("MV/LV transformer", 0.99), ("UPS (double conversion)", 0.96),
          ("distribution and PDU", 0.995), ("rack power shelf AC to 50 V DC", 0.975),
          ("48 V to 12 V bus converter", 0.98), ("12 V to 0.8 V regulators", 0.90)]
eff = 1.0
for name, e in stages:
    eff *= e
    print(f"  after {name:32s} {eff:6.1%}")
print(f"of every 100 W drawn from the grid for IT, {100 * eff:.0f} W reaches the transistors")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Delivering a kilowatt at 0.8 V through just half a milliohm would waste most of the power: the copper itself becomes a heater. At 48 V the same path loses a fifth of a watt. So the rule of power delivery is carry power at the highest practical voltage and convert down as close to the load as possible. On today's boards the final regulators sit a few centimetres from the package, and the newest designs put them directly underneath it ("vertical power delivery"), shortening the high-current path to millimetres — the same reasoning that moved chip power wiring to the back of the wafer in Transistor scaling.

The same arithmetic at rack scale explains the 2025–26 move to higher voltages. A 120 kW rack at 54 V carries over 2,000 A, through busbars of tens of kilograms of copper, and loses about half a kilowatt in a busbar of 0.1 mΩ; at 800 V the current is 150 A and the loss two watts. NVIDIA has announced an 800 V DC architecture for racks of up to a megawatt from 2027, estimating that a 1 MW rack at 54 V would need up to 200 kg of copper busbar (NVIDIA), and Open Compute Project members have specified a ±400 V DC power rack. High-voltage DC brings its own engineering: DC arcs do not self-extinguish at a current zero as AC arcs do, so breakers, connectors and safety practice must change.

The block's third result is about speed rather than loss. When thousands of tensor cores start or stop together, the chip's current can change by hundreds of amperes in a microsecond, and even 50 pH of inductance between regulator and die gives a 25 mV droop — 3% of the supply, enough to cause timing failures unless banks of capacitors on the package and board supply the charge until the regulators respond.

ComputationC

PUE, cooling and water

Definition— Power usage effectiveness (PUE)

PUE=total energy used by the facilityenergy delivered to IT equipment≥1.\text{PUE} = \frac{\text{total energy used by the facility}}{\text{energy delivered to IT equipment}} \geq 1.

Introduced by The Green Grid in 2007 and standardised as ISO/IEC 30134-2, it measures the overhead of cooling, power conversion upstream of the IT equipment, lighting and offices. A PUE of 1.5 means 0.5 W of overhead for every watt of IT.

The Uptime Institute's annual surveys have put the industry-average PUE at roughly 1.5–1.6 for more than a decade; large modern facilities do much better — Google reports a fleet-wide 1.09 for 2025 (Google). The program below builds a PUE from its parts, then asks how much fluid it takes to carry a rack's heat away, and how much water evaporation would use.

python
# 1. PUE = total facility energy / IT energy, built up from the overheads.
def pue(it_kw, cooling_cop, electrical_eff, other_kw_frac=0.02):
    cooling_kw = it_kw / cooling_cop                 # electricity to move the IT heat out
    electrical_loss = it_kw * (1 / electrical_eff - 1)
    other = it_kw * other_kw_frac                    # lighting, offices, security
    return (it_kw + cooling_kw + electrical_loss + other) / it_kw

IT = 10_000                                          # kW of IT load
for label, cop, eff in [("older air-cooled hall, chillers all year", 3.0, 0.90),
                        ("modern air-cooled, economiser most hours", 8.0, 0.95),
                        ("warm-water liquid cooling, dry coolers", 25.0, 0.96)]:
    print(f"{label:42s} PUE = {pue(IT, cop, eff):.2f}")

# 2. Moving the heat: mass flow = power / (specific heat x temperature rise).
Q = 120e3                                            # W, one AI rack
water = Q / (4186 * 10)                              # kg/s of water warming by 10 K
air = Q / (1.2 * 1005 * 15)                          # m^3/s of air warming by 15 K
print(f"\n120 kW rack: water {water:.2f} kg/s ({water * 60:.0f} L/min), "
      f"or air {air:.1f} m^3/s ({air * 3600:,.0f} m^3/h)")

# 3. Water: rejecting heat by evaporation alone.
latent = 2.43e6                                      # J/kg, latent heat of evaporation near 30 C
litres_per_kwh = 3.6e6 / latent
print(f"evaporating water absorbs {latent / 1e6:.2f} MJ/kg, so 1 kWh of heat needs {litres_per_kwh:.2f} L")
campus_mw = 100
print(f"a {campus_mw} MW campus cooled only by evaporation: "
      f"{campus_mw * 1e3 * litres_per_kwh * 24 / 1e6:.1f} million litres a day")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Cooling is the largest overhead, and the table shows the levers: free cooling with outside air or water for most of the year instead of mechanical chillers, and liquid cooling with warm water that can be cooled by dry coolers without chillers at all.

Heat can be carried by air only up to a point. To remove 120 kW with air warming by 15 K takes 6.6 m³ of air every second through one rack, a gale. Air-cooled racks top out at around a few tens of kilowatts, extended a little by rear-door heat exchangers. Water carries about 3,500 times more heat per unit volume per kelvin than air, so the same 120 kW needs under 3 litres of water a second. Hence direct-to-chip liquid cooling: cold plates on the accelerators and CPUs, fed through manifolds in the rack from a coolant distribution unit (CDU) that exchanges heat with the facility's water loop, while fans still cool the remaining components. Immersion cooling submerges whole servers in a dielectric liquid, either single-phase (oils) or two-phase (fluids that boil on the chips); the leading two-phase fluids were fluorinated compounds of the PFAS family, whose main manufacturer, 3M, committed to ending production by the end of 2025, pushing the field towards alternatives. As of 2026, direct-to-chip cooling is the mainstream for AI racks and immersion a niche.

Heat reuse is easier with liquid cooling, because return water at 40–60 °C is useful for district heating directly or through a heat pump. Meta's data centre in Odense, Denmark, has supplied the city's district heating network since 2020, and Stockholm Data Parks connects data centres to Stockholm's network; Germany's 2023 Energy Efficiency Act requires new data centres to reach a PUE of 1.2 and to reuse a growing share of their heat. Reuse needs a heat customer nearby, which many sites, chosen for cheap power and land, do not have.

Water. Evaporating water absorbs about 2.4 MJ per kilogram, so rejecting 1 kWh of heat by evaporation alone takes about 1.5 litres, and a 100 MW campus cooled entirely that way would evaporate millions of litres a day. Many sites use evaporative cooling only on hot days, and closed-loop liquid systems with dry coolers use almost no water — at the price of more electricity for fans and, in hot weather, chillers. The industry metric is water usage effectiveness (WUE), litres per kWh of IT energy. Two caveats matter for an honest picture: the water used to generate the electricity (in thermal power stations' cooling towers) is often larger than the water used on site, and the local question — how stressed is this watershed, in this season? — matters more than the global total.

ExplanationH

Connecting to the grid: queues, on-site power and long contracts

A campus can be built in eighteen months; a new high-voltage transmission line or a large power station typically takes five to ten years. That mismatch is the central problem of 2024–26.

Queues. In many places, a new large load or generator must wait for studies and network upgrades before it can connect. Lawrence Berkeley National Laboratory's Queued Up reports have tracked United States interconnection queues holding more proposed generation and storage capacity than the country's entire installed fleet, with typical waits of several years. Ireland's grid operator effectively paused new data-centre connections in the Dublin area; Amsterdam and Singapore imposed moratoria for some years before allowing growth under efficiency conditions.

On-site generation and storage. To start sooner, some developers build their own power: gas turbines or reciprocating engines on site (sometimes as a "bridge" until the grid connection arrives, sometimes permanently, and in some places drawing local opposition over air quality), fuel cells, and large battery systems. Batteries are also replacing diesel generators for short backup, and can serve the grid when not needed.

Long-term contracts. The largest technology companies are among the world's biggest buyers of wind and solar power through power purchase agreements (PPAs), and some now aim to match their consumption with carbon-free supply hour by hour rather than over a year. Because AI campuses run continuously, they have also turned to firm low-carbon sources:

YearAgreement (as announced)
2024Microsoft and Constellation: a 20-year agreement to restart Unit 1 of the Three Mile Island plant in Pennsylvania (about 835 MW), renamed the Crane Clean Energy Center (Constellation)
2024Google and Kairos Power: up to 500 MW from a fleet of small molten-salt-cooled reactors by 2035 (Kairos Power)
2024Amazon and X-energy, with utilities: development of small modular reactors in Washington state and elsewhere
2025Amazon and Talen Energy: up to 1.92 GW from the Susquehanna nuclear plant in Pennsylvania through 2042, restructured as a grid-connected purchase after US regulators rejected an expanded direct connection in 2024
2025Meta and Constellation: a 20-year agreement for the output of the Clinton nuclear plant in Illinois
2025Google and NextEra: a 25-year agreement to restart the 615 MW Duane Arnold plant in Iowa, targeted for around 2029

Restarts of recently closed plants and uprates of running ones can add power within a few years. New reactors, including small modular designs, are unlikely to supply significant power before the 2030s, and their costs and schedules remain to be demonstrated. In the meantime, most new data-centre demand is being met by gas, wind, solar and storage, with the mix depending heavily on the country.

ComputationC

Demand flexibility: the data centre as a grid asset

A grid is built for its peak, and the peak lasts only a few hours a year. If a new load could step back during those hours, it could connect without new generation or transmission being built first. A 2025 study from Duke University's Nicholas Institute (Norris et al.) estimated that United States grids could absorb of the order of 100 GW of new load if that load accepted curtailment for well under 1% of its annual energy. The program tests the idea on a synthetic year of hourly demand.

python
import numpy as np

# A synthetic year of hourly demand on a grid whose capacity was built for its own peak.
rng = np.random.default_rng(2026)
h = np.arange(8760)
day = h // 24
season = 0.10 * np.cos(2 * np.pi * (day - 200) / 365)          # summer peak (air conditioning)
daily = 0.12 * np.sin(np.pi * np.clip((h % 24 - 7) / 14, 0, 1)) ** 2   # afternoon hump
weather = np.repeat(rng.normal(0, 0.035, 365), 24)              # hot and cold spells, day by day
load = 0.70 + season + daily + weather + rng.normal(0, 0.01, 8760)
capacity = load.max()                                             # existing system: just meets its peak
print(f"existing load: mean {load.mean() / capacity:.0%} of peak, "
      f"above 90% of peak for {np.sum(load > 0.9 * capacity)} hours a year")

def curtailed_fraction(new):
    """New flat load `new`, cut back whenever the total would exceed capacity."""
    cut = np.clip(load + new - capacity, 0, new)
    return cut.sum() / (new * 8760), np.count_nonzero(cut)

def max_new_load(allowed):
    lo, hi = 0.0, 1.0                                             # bisection on the new load size
    for _ in range(60):
        mid = (lo + hi) / 2
        if curtailed_fraction(mid)[0] <= allowed:
            lo = mid
        else:
            hi = mid
    return lo

print("\ncurtailment allowed   new flat load that fits   hours with some curtailment")
print(f"   none (firm)           {0:6.1%} of system peak")
for allowed in (0.0025, 0.005, 0.01):
    new = max_new_load(allowed)
    frac, hours = curtailed_fraction(new)
    print(f"   {allowed:5.2%} of its energy     {new / capacity:6.1%} of system peak       {hours:5d}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

On this synthetic grid, a new load that could never be curtailed fits nowhere without new capacity, because the existing load already reaches the peak. Accept curtailment of 0.25% of its energy — 77 hours of partial cut-backs in a year — and a flat new load of about 7% of the system peak fits; at 1%, about 11%. The numbers depend on the shape of the load curve, and a real study also has to model transmission limits and reserve margins, but the structure is general: peaks are rare, so a little flexibility buys a lot of capacity.

Can AI data centres actually flex? Training jobs can checkpoint and pause, or slow down, at some cost in time; inference is latency-sensitive but can sometimes be shifted to another region; and on-site batteries can cover short events without touching the computing at all. In 2025 Google announced demand-response agreements with two US utilities for some of its machine-learning workloads, and EPRI's DCFlex collaboration is testing such arrangements with utilities and operators (EPRI).

The grid also needs data centres to behave predictably. Synchronised training steps can swing a large cluster's power by tens of per cent in under a second, which some operators and chip vendors now smooth with energy storage and firmware; and in July 2024 a grid fault in northern Virginia caused about 1.5 GW of data-centre load to switch to backup power at once, an event the North American Electric Reliability Corporation investigated as a new kind of reliability risk.

Common misconceptionsH

Common misconceptions

Misconception

"AI will soon use most of the world's electricity." The IEA's base case has data centres at a little under 3% of world electricity in 2030, with a wide range of uncertainty. The real pressure is local and about timing: gigawatts arriving at particular places faster than grids can be extended.

Misconception

"A PUE of 1.1 means the data centre is efficient." It means the facility overhead is small. It says nothing about the efficiency of the servers, how much useful work they do, the carbon intensity of the electricity, or the water used.

Misconception

"Higher distribution voltage is about saving a few per cent of energy." Efficiency matters, but at megawatt racks the binding constraints are physical: current, copper weight, connector heating and space. Raising the voltage from 54 V to 800 V cuts the current by fifteen times and the resistive loss in the same conductor by over two hundred times.

Misconception

"Liquid cooling uses more water." Closed-loop liquid cooling recirculates the same coolant; whether a site evaporates water depends on how it finally rejects heat to the outside (evaporative towers or dry coolers), not on whether the chips are liquid- or air-cooled. Warm-water liquid cooling often makes dry cooling easier.

Exercises

Exercises

1. Powering the AI data centre

Warm-upHuman

Over a year, a data centre's IT equipment averages 20 MW. Cooling averages 3 MW, losses in transformers, UPS and distribution 1.2 MW, and lighting and offices 0.2 MW. What is its PUE? Give 100 × PUE as an integer.

2. Powering the AI data centre

Warm-upHuman

A server tray draws 1.5 kW through a cable-and-connector path of 2 mΩ. By what factor does the resistive loss in that path fall if the distribution voltage is raised from 12 V to 48 V, delivering the same power?

3. Powering the AI data centre

CoreHuman

A campus has a grid connection of 300 MW and runs at an average of 75% of it through the year (8,760 hours). How much electricity does it use in a year, in GWh?

4. Powering the AI data centre

CoreComputational

Write coolant_flow(heat_w, delta_t, cp, density) returning the volume flow in m³/s needed to carry heat_w watts with a temperature rise of delta_t kelvin, for a fluid of specific heat cp (J/(kg·K)) and density (kg/m³). Then set WATER_LPM to the water flow, in litres per minute, for a 120 kW rack with a 10 K rise (cp = 4186, density = 1000), and AIR_M3S to the air flow in m³/s for the same rack with a 15 K rise (cp = 1005, density = 1.2).

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

5. Powering the AI data centre

StretchComputational

Reproduce the lesson's flexibility calculation. Write curtailed_fraction(load, capacity, new): a constant new load new is added to an hourly load array; in any hour where the total exceeds capacity, the new load is cut back by the excess (but by no more than new). Return the curtailed energy as a fraction of the new load's energy. Then write max_new_load(load, capacity, allowed), the largest new (between 0 and capacity) whose curtailed fraction is at most allowed, found by bisection to within 10−610^{-6}.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

6. Powering the AI data centre

CoreHuman

Why do servers distribute power at 48 V (or higher) and convert to the chip's 0.8 V only a few centimetres — or millimetres — from the processor?

Choose one

7. Powering the AI data centre

CoreHuman

Which of these would reduce the water a data centre evaporates on site? Enter the letters.

  • A. Rejecting heat with dry coolers (air-cooled heat exchangers) instead of evaporative cooling towers
  • B. Running evaporative cooling towers for more hours of the year
  • C. Closed-loop direct-to-chip liquid cooling whose facility loop is cooled by dry coolers
  • D. Allowing warmer coolant temperatures, so that outdoor air can cool it on more days without evaporation
  • E. Moving the cooling fans from the building into the servers

Write elements separated by commas. Order and repeats don’t matter — just like in a set!

Further reading and resources