From One Token to a Gigawatt
If most jobs used AI every day, the world would need hundreds of gigawatts of new data centres. Chips and capital could keep pace within a few years; grid connections are the slowest part of the build-out.
If 80% of the world's jobs handed whole tasks to AI, how many gigawatts of data centres would that take?
How the answer is built, in four parts
Open the parts in any order. Part 1 calculates the answer to your estimate.
Part 1 · From jobs to gigawatts
Daily AI use at work adds up to gigawatts of data centres
The six steps below each change the chart, and the controls under it update every view. Your estimate stays on the chart as a dashed marker for comparison.
About 3.5 billion people work
The slider sets the share of those workers who use AI in their job every day.
Each task uses tokens
AI use is measured in tokens. Select what a typical worker uses AI for in a day, or pick one of the three presets under the chart.
For comparison, Google processed about 3.2 quadrillion tokens a month across its products in mid-2026, about 74 billion a minute.
Tokens set the number of GPUs, and GPUs set the power
One GB200-class GPU produces about 1,500 tokens a second on a frontier model (the largest current models). Working hours overlap, so peak demand is 1.7 times the average. Each GPU draws about 2.4 kW including cooling and networking.
Faster chips reduce the total in proportion, as the chip speed control shows.
Each gigawatt uses about 6 TWh a year
That assumes the data centres run at 70% of capacity on average, so each GW uses about 6 TWh a year. The circle shows the share of the 32,000 TWh the world uses each year.
Which runs out first: chips, capital or grid power?
Three separate estimates of new AI capacity per year, based on chip sales, capex and grid connections, all come to about 16 GW in 2026. They diverge after that.
The yearly cost per worker
At $50B per GW, the build-out must earn back about 21 cents a year for every dollar of capex. Divided across all AI-using workers, that gives a yearly cost per person.
Part 2 · Inside one request
Memory limits how many people one GPU can serve
A model reads text as tokens. For each one it has read, it keeps notes in the KV cache, so it does not have to reread the conversation. Those notes sit in the GPU's memory next to the model's parameters.
How a model reads text
Type or paste anything. Each colour is one token.
One H100 GPU holds a 70-billion-parameter model. How many people, each with a 20-page report open, can it serve at once?
GPU memory fills up long before processing power runs out
The GPU has 80 GB of memory, and the model takes 70 GB at FP8. Choose a document and how many people have it open.
Memory sets the number of people one GPU can serve at once. Providers spread one model across eight GPUs mainly to make room for these notes.
Where the time goes before the first word appears
Press Send to follow one request (the dot) as it crosses the network, waits for a batch, goes through prefill and decode, and returns. The journey is slowed five times so it can be followed; the answer then streams at the chip's real speed.
Part 3 · The chips
The fastest chips need dozens of racks to hold one model
To produce each token, a chip reads all of the model's weights once. Chips differ mainly in where they store those weights: in HBM stacked beside the chip, or in SRAM on the chip itself, which is much faster and much smaller.
A 1-trillion-parameter model fits in one rack of NVIDIA B200s. How many racks of Groq LPUs does it fill?
Each block is one rack. Change the model size and the number format (bytes per parameter).
One B200 GPU serving a 70B model. Adding people to a batch slows each person's answer slightly and makes each token much cheaper.
Groq (about 280 tokens/s per person) and Cerebras (about 2,100) answer a single person much faster. GPUs serving large batches produce the cheapest tokens. Figures are approximate.
Part 4 · The network
A 100,000-GPU training run stops about every half hour
In training, every GPU waits for all the others before starting the next step. One failed cable, card or chip therefore stops the whole job, which then restarts from the last checkpoint.
How often does a 100,000-GPU training run stop for a hardware fault?
Top: one simulated day, with a red tick for each stop. Bottom: the share of paid GPU time spent training at each checkpoint interval, which falls when saves are too frequent or too rare.
Meta reported 419 unexpected interruptions in 54 days of Llama 3 training on 16,384 GPUs. Each component fails about once in six years, but a 100,000-GPU job has so many components that it stops every half hour.
Each extra switch tier adds about one transceiver (laser module) per GPU, so the count rises in steps. Switches with more ports move each step to larger clusters.
The answer
Grid connections are the slowest part of the build-out
The four parts cover different links in one chain. Tokens set the number of GPUs, memory sets how many people each GPU can serve, and the network sets how much paid GPU time does useful work.
Sources and assumptions: data-centre electricity, $50B per GW, kW per GPU and the three capacity estimates come from the AI build-out analysis (IEA, company filings and analyst estimates, as of 30 October 2026). World electricity (about 32,000 TWh in 2026) extends the IEA’s 2025 figure. Sector electricity, chip speeds and optics prices are rounded public estimates. Google’s token volume is from Alphabet’s Q2 2026 earnings call; the training-run failures are from Meta’s Llama 3 paper.
Model defaults: 1,500 tokens/s per GPU, peak factor 1.7, 2.4 kW per GPU, 70% load, $50B per GW, 0.21 annual capital charge.