How Data Centers Convert Electrons to Tokens
Every AI response is a product manufactured at a token factory — a data center. But data centers aren't just buildings; they are advanced machines that use electrical energy to power IT equipment that produces tokens and generates a lot of heat in the process. Understanding how a data center works is analogous to understanding the global supply chain that underpins the AI boom. Join us as we dissect the most powerful factories humans have ever built!
"You can't create an industry without energy."— Jensen Huang
Every token begins with an electron. OpenAI's Stargate data center in Abilene, Texas, will draw 1 gigawatt of power, enough to power the entire city of Seattle. Transmission lines need to transmit power over long distances, so they operate at very high voltages — typically between 115 and 500,000 Volts.
The request to connect data centers to a grid can take 3 to 5 years to be approved.
The first bottleneck is in transforming electricity.
High voltage is efficient for transmission, useless for servers. High-voltage transformers step the grid power down to 20 kV; medium- and low-voltage transformers closer to the data halls step it further to roughly 400V.
Lead times have blown out for high-voltage transformers and are higher for medium- and low-voltage units too. Data centers use so much power that they are effectively building their own versions of electrical substations, which you often see in cities or next to a cluster of large buildings.
The switchboard.
Medium-voltage switchgear is the building's electrical control center — it routes power to each individual data hall and isolates faults before they cascade.
Unlike the transformers upstream, switchgear is the rare piece of the chain that isn't a multi-year wait: lead times run 3–4 months.
Grid backup and uninterrupted runtimes.
Data centers support real-time digital infrastructure. It's not just your chatbot; hospital systems also run "on the cloud" (read: data centers). A wall of UPS batteries bridges the first 90 seconds, and rows of diesel generators take over until utility power returns, so that data centers can run without any interruption.
While generators are facing longer lead times, they are large enough to power the entire data center (i.e., not just backup), eliminating the need to wait multiple years for grid connections. xAI reportedly shipped an entire power plant from abroad to power its Colossus data center in Tennessee.
An invisible switch.
The Automatic Transfer Switch (ATS) decides between the grid and backup power. When utility power fails, it switches over in milliseconds — fast enough that the racks downstream never notice.
These also now have a lead time longer than a year. Noticing a trend?
Down to the rack.
The Power Distribution Unit is the last electrical step before the chips. It splits the building's main supply into per-rack circuits — typically at 400 volts — and meters each circuit's draw.
From here, every electron's destination is a server.
Time to do the work.
"Compute" means mathematical work — trillions of operations per second, run by GPUs and CPUs. AI workloads require 10–100× more compute per rack than traditional servers running search and streaming. Inside the rack, electrons stream across the GPUs in lane after lane, performing trillions of operations per second as they pass through.
One rack pulls up to 120 kilowatts today — enough to power 120 average homes. By 2028, top racks are projected to require 1,000 kW — a 100× jump in just eight years.
A super computer.
Each rack contains individual server trays packed with GPUs, CPUs, memory, and high-speed networking hardware that enable them to communicate with their neighbors to run calculations in parallel.
A frontier training run uses 10,000+ GPUs at once, all working in lockstep.
Where electrons become tokens.
Inside the GPU, transistors doing math at roughly 1 volt transform the electrical work that took all these pieces of industrial equipment to deliver into mathematical work — the next token in your reply.
Dealing with the heat.
Every watt of electricity that goes in becomes a watt of heat — 100% of it. Cooling accounts for up to 23% of the facility's total electricity use, and the latest AI racks require liquid cooling — air cooling simply cannot keep up.
Coolant Distribution Unit (CDU) or Computer Room Air Handler (CRAH) moves hot air/liquid away from the racks. They also return cool liquid or air to the racks. Chillers cool the hot coolant/air from the racks by releasing the heat into the atmosphere.
The AI commodity.
Just as oil output is measured in barrels, AI output is measured in tokens. Tokens are the new industrial commodity sold by the millions, billed in cents, and generating billions in revenue.
AI-driven compute demand could more than triple by 2030.
Email, Cloud Storage, Streaming.
47% of today's data center capacity still serves traditional workloads — search, streaming, online games, and productivity software. Compute-light, retrieval-dominant.
This is the workload that data centers were originally built for and which the internet runs on. It's still the majority… for now.
Teaching the model.
A frontier LLM model training run takes 10,000+ GPUs working in concert for weeks, costing as much as $390 million today — up from ~$2 million for GPT-3 released in November 2022. Training produces exactly one thing: a new AI model.
The capability inside every AI product comes from this step. It's where electrons become intelligence.
Using the model.
Using the model. Every AI response requires inference and creates recurring token usage for every user's query, every day. Inference is the fastest-growing workload. As models get better and the world reaches for them, this workload will compound.
McKinsey's 3× projections by 2030 could be conservative.
Tokens need more than electrons.
Three constraints sit outside the data hall — and each could become its own bottleneck.
With all the electrical, computing, and cooling equipment, data centers can take up a lot of space. Stargate will take up as much space as 450 soccer fields.
Data centers can "drink" up to 334 million gallons per year, primarily to stay cool (enough to fill a little over 500 Olympic-sized pools). Operators are chasing the cold in cooler climates, underwater, and even in...space!
AI chips in data centers are built from imported minerals — gallium, germanium, indium, and tantalum, to name a few. The U.S. imports 100% of many of these.