Chris Collins

Tools · ai

AI cost and energy economics

What a token, a training run and a gigawatt datacenter cost: cost per million tokens, 6·N·D training cost, datacenter TCO against GPU life and power price, generation mix and overbuild, PPA prices and energy bills, embodied vs operational carbon, GPU depreciation, and an educational export-control and EU AI Act check.

Added

How to use it

Eight tabs, each a small calculator with the book’s worked example as a preset:

  • Cost per token — all-in $ per GPU-hour ÷ (tokens per second × 3,600) × 106, with a grid of prices and throughputs and, for reasoning models, the cost of one long answer.

  • Training cost — C ≈ 6·N·D, GPU-hours = C ÷ (dense peak × MFU × 3,600), times a $ range. Give the reported GPU-hours instead and it shows the MFU they imply.

  • Datacenter TCO — annual cost = hardware capex ÷ GPU life + energy (TWh × $/MWh) + facility overheads, as stacked bars for lives of two to eight years with the energy share on each, and a table at other power prices.

  • Generation mix — on-site turbine capacity (units × MW), the nameplate a source needs to supply an average load at its capacity factor, and the book’s menu of sources with time to power.

  • PPAs & bills — the price a power-purchase agreement implies (value ÷ MWh delivered) and a site’s annual energy bill across a price range, against the ≈ $80/MWh siting rule of thumb.

  • Embodied carbon — one accelerator’s manufacturing emissions against its running emissions on a chosen grid: the break-even time, the embodied share over its life, and a table across grids.

  • Depreciation — book value year by year on straight-line, double-declining or sum-of-digits schedules, comparing the life on the books with a shorter economic life.

  • Compliance (educational, not legal advice) — training compute against the EU AI Act’s 1025 FLOP presumption, and the US export-control Total Processing Performance (TPP) and performance density (PD) of an accelerator against the October 2023 ECCN 3A090 thresholds.

Numbers accept 1e9, 1,000 and 1_000.

Where the book’s figures differ

The tool computes; where the book prints a different number it says so on the tab.

  • Epoch’s 1 GW site: $30B of accelerators over five years is $6B a year, not $5B, so the annual total is $7.1B (with $0.6B of energy and $0.5B of facility), not "~$8B". Three years gives $11.1B; seven years $5.4B with energy at about 11%, not "past 20%". Depreciation is 10× energy.

  • Colossus: 35 turbines × 16.5 MW is 577.5 MW, not "roughly 400 MW".

  • TPP is the highest single-precision figure (2 × MacTOPS × bit length, dense), not a sum across precisions; non-datacenter parts at TPP ≥ 4,800 are NAC-eligible rather than licensed; and 3A090.b has a second band (TPP ≥ 1,600 with PD 3.2–5.92).

Limits

  • Every result is an estimate from simple formulas: no cost of capital, taxes, or utilisation ramps; straight-line depreciation in the TCO.

  • Throughput, MFU, prices and grid intensities are inputs, not predictions. Grid intensities are approximate yearly averages (Ember; EIA for the US). Hardware figures are dense (NVIDIA’s headline FP8/INT8 numbers assume 2:4 sparsity and are twice these).

  • The implied PPA price assumes full delivery for the whole term; Talen–AWS ramps to 1.92 GW by 2032, so its true average price is higher than the ≈ $63/MWh shown.

  • The compliance tab is a simplified, educational reading of ECCN 3A090 (October 2023, per CSET’s summary) and Article 51 of the EU AI Act. Later rules — the 2024 HBM controls, the rescinded 2025 diffusion rule, the April 2025 H20 licence requirement, case-by-case H200 licensing in 2026 — sit on top. It is not legal advice. Die areas for MI300X and B200 are estimates.

Sources

Related: AI hardware calculator (memory, roofline, training time and tokens per second — the throughput this tool prices), yield, chiplets and weights-in-silicon (what a good die costs), solar ROI (payback and LCOE of a single solar system).

Companion to The Economics: Cost per Token and the Depreciation Debate, the generation menu and the cost curve of the energy layer, Embodied versus operational carbon and AI Regulation and Export Controls of _AI Acceleration: From Transistors to Trillion-Parameter Models.

Linking to an example

Links can be written by hand. Add the tab after #: token, training, tco, power, energy, carbon, depreciation or compliance. Settings not in the link take their defaults.

Parameter Meaning Example

h

All-in $ per GPU-hour (Cost per token)

h=3

tps

Output tokens per second per GPU

tps=2000

ans

Tokens per answer (optional)

ans=20000

n, d

Parameters and training tokens, in billions (Training cost)

n=405&d=15600

gpu

GPU and precision: h100, h100fp8, a100, b200, mi300x, custom

gpu=h100

peak

Dense peak TFLOPS

peak=989.5

mfu

Model FLOPs utilisation, %

mfu=40

lo, hi

$ per GPU-hour range

lo=2&hi=4

gh

Reported GPU-hours, millions (0 = estimate from MFU)

gh=30.84

capex

Hardware capex, $B (Datacenter TCO)

capex=30

life

GPU life, years

life=5

twh

Energy per year, TWh

twh=8

mwh

Power price, $/MWh

mwh=75

fac

Facility overheads, $B per year

fac=0.5

units, unit

Turbines on site and MW each (Generation mix)

units=35&unit=16.5

load, duty

Load in MW and duty cycle %

load=1000&duty=80

src

Source: gasAero, ccgt, nuclearRestart, smr, solar, wind, diesel

src=solar

cf

Capacity factor %

cf=27.5

ppa, yrs, gw, lf

PPA value $B, term in years, capacity GW, delivered load factor % (PPAs & bills)

ppa=18&yrs=17&gw=1.92

bt, p1, p2

Site energy TWh per year and a low and high $/MWh

bt=8&p1=30&p2=120

emb

Embodied kgCO2e per GPU (Embodied carbon)

emb=175

w, util, pue

Board power W, utilisation %, PUE

w=700&util=60&pue=1.2

grid

hydro, france, eu, us, gas, coal, custom

grid=us

ci

Grid intensity, gCO2e/kWh

ci=370

cy

Service life, years

cy=5

cost

Fleet capex, $B (Depreciation)

cost=100

book, econ

Life on the books and economic life, years

book=6&econ=3

meth

sl, ddb or syd

meth=ddb

salv

Salvage, % of cost

salv=10

en, ed

Parameters and tokens in billions for the EU check (Compliance)

en=400&ed=4000

fl

Or total training FLOP (0 = use 6·N·D)

fl=3.8e25

chip

h100, h800, h20, a100, a800, l40s, rtx4090, mi300x, b200, custom

chip=h20

tops, bits, area, dc

Custom part: dense TOPS, bit length, logic-die mm², designed for datacenters (1/0)

chip=custom&tops=1000&bits=8&area=600&dc=1

example

One of the presets below; parameters given with it override the preset’s

example=p19-train-405b

In a URL + means a space: write 1e25 or 1e%2B25. A hand-typed fl=1e+25 also works — the tool reads 1e 25 back as 1e+25.

Presets:

example= Shows

p19-cost-per-token

$3 per GPU-hour at 2,000 tokens/s ≈ $0.42 per million tokens (open)

p19-reasoning-answer

The same rate priced per 20,000-token reasoning answer (open)

p19-train-405b

Llama 3.1 405B: 3.8 × 1025 FLOP, 30.84M H100-hours at $2–4 ≈ $62–123M, implied MFU ≈ 35% (open)

p19-mfu-20-vs-40

The same run estimated at 20% MFU: twice the GPU-hours of 40% (open)

p19-depreciation

$100B of accelerators on a 6-year schedule against a 3-year economic life (open)

energy-tco-epoch

Epoch’s 1 GW site: $6B a year of depreciation, $7.1B in all (open)

energy-tco-3yr

The same on a 3-year GPU life: $11.1B a year (open)

energy-tco-7yr

A 7-year life: $5.4B a year, energy ≈ 11% (open)

energy-colossus

35 SMT-130 turbines × 16.5 MW = 577.5 MW (open)

energy-solar-firming

A 1 GW cluster at 80% duty needs ≈ 2.9 GW of solar at a 27.5% capacity factor (open)

energy-susquehanna-ppa

Talen–AWS: $18B over 17 years for 1.92 GW ≈ $63/MWh (open)

energy-power-bill

8 TWh a year at $30–120/MWh = $240–960M (open)

sust-embodied-clean

One H100, ~175 kgCO2e embodied, on a ~30 g/kWh hydro grid (open)

sust-embodied-us

The same card on the US-average grid (open)

reg-eu-threshold

400B parameters × 4T tokens = 9.6 × 1024 FLOP, just under 1025 (open)

reg-eu-405b

Llama 3.1 405B at 3.8 × 1025 FLOP (open)

reg-tpp-h100

H100: TPP 15,832, PD 19.4 — 3A090.a (open)

reg-tpp-h20

H20: TPP 2,368, PD 2.9 — under the October 2023 lines (open)

reg-tpp-a800

A800: TPP 4,992 — caught by the October 2023 rewrite (open)

Runs entirely in your browser; nothing you type leaves this page.