✦ ✦ ✧
Future-Ready · The Compute Stack

The Full Stack

A chatbot answers you in about two seconds — but that answer rides a five-layer tower of energy → chips → data centers → model → app. In this simplified model, pull out any one layer and the tower goes dark — real systems branch, so some AI runs on your phone, cached answers skip layers, and big services keep spares.
🏗️The Compute Stack, Bottom to Top
each layer stands on the one below it 💬 APPLICATION the top floor 5 The app you actually touch — a chatbot or a homework helper. It sends your question down the stack and shows you the answer. 🧠 MODEL the brain 4 A giant web of learned numbers, shaped by TRAINING. Every answer it gives is INFERENCE — chips racing through those numbers in ~2 seconds. 🏢 DATA CENTERS the buildings 3 Warehouse-sized halls of racks, fiber cable and liquid cooling. One big AI data center can draw the power of tens of thousands of homes. 🔲 CHIPS the silicon 2 GPUs printed with light onto silicon: fancy sand! One AI chip holds 80+ billion transistor switches and does thousands of math problems at once. ⚡ ENERGY GRID the basement 1 Power plants and power lines. Nothing above this layer moves without the electricity it delivers.
🎂 It's a layer cake: energy feeds chips · chips fill data centers · data centers run models · models power apps.
🏁One Chip, Three Continents
A relay race nobody can run alone 🏃‍♀️🏃 📐 🇺🇸 California NVIDIA designs the GPU — a chip that does 1000s of math problems at once. 💡 🇳🇱 Netherlands As of 2026, ASML is the only supplier of the EUV machines that print the tiniest features. 🏭 🇹🇼 Taiwan TSMC: >90% of the world's leading-edge logic chips (U.S. Commerce, Apr 2024). 🏁 ONE AI CHIP 80+ billion tiny transistor switches
🌍 Design in one country, the printing machine from another, the factory in a third — that's a supply chain, and as of 2026 essentially every company training large AI models depends on some version of it. Those fabs run in cleanrooms thousands of times cleaner than a hospital operating room.
🏋️Training Day vs. Game Day
⏱ One giant build… then billions of tiny answers, forever 🏋️ 1000s of GPUs weeks to months TRAINING INFERENCE 💬 💬 💬 💬 💬 HAPPENS ONCE GPT-3 used about ~1,287 MWh (2021 estimate) U.S. homes use in a year BILLIONS OF TIMES A DAY Google measured its median text prompt at about 0.24 Wh (Google, median text prompt, May 2025) nine seconds. Now multiply that by billions.
🏋️TRAINING = camp
What happensThousands of GPUs work side by side for weeks or months, studying mountains of text until the model learns the patterns hiding in language.
🗓️ One gigantic push — then it's done.
🏀INFERENCE = game day
What happensEvery time anyone asks the finished model a question, chips in a data center sprint through the model's numbers and generate a reply.
⚡ Tiny per answer × billions of answers.
🔌The Honest Energy Question
🌍 ALL the electricity used on Earth data centers are this thin sliver (2024) 🔍 Zoomed in on the sliver 2024 ~415 TWh · about 1.5% of the world's electricity (IEA) 2030? IEA Base Case: ~945 TWh …to roughly what the entire country of Japan uses today. 🇯🇵
⚡ The efficiency race energy per prompt 33× less …now this thin one year earlier one year later
🏎️ Maybe the most important race in tech
Efficiency = getting the same job done while using less energy, time, or money.
Over about a year, Google reported cutting the energy of its median Gemini Apps text prompt by roughly 33× — smarter chips, software and cooling. It is a ratio between two of Google's own measurements, so don't multiply backwards to a universal "before" number.
And the answers actually got better at the same time.
Efficiency doesn't make the energy question vanish — it changes the math, and the race is nowhere near finished.
🔑Key Terms
🏗️Compute stack All the connected layers — energy, chips, data centers, models, and apps — that work together to make AI run.
🎂Layer One level of a bigger system that does its own job while depending on the levels below it.
🏋️Training Changing a model's learned numbers using data and a goal, with thousands of chips working for weeks or months — repeated across pretraining, fine-tuning and new versions.
🏀Inference The everyday work a finished model does each time it answers a question or makes a prediction.
🧠Model What training produces — a giant web of learned numbers that turns your question into an answer by matching patterns. People say "brain"; literally it is stored numbers and arithmetic, and it can sound fluent while being wrong.
🏢Infrastructure The physical gear — buildings, cables, cooling, and power connections — that technology needs to operate.
🌍Supply chain The worldwide chain of companies and countries that each make one piece of a finished product.
🧰Every Floor Is Hiring
🏗️Data center builds · U.S.
Who's neededElectricians, cooling and HVAC specialists, and power engineers are among the most in-demand workers in tech right now — every new AI data center needs crews to wire in megawatts and keep blazing-hot racks cool.
🎓 Many of these careers start with an apprenticeship instead of a four-year degree.
🧑‍🔬Chip fabs · Taiwan & Arizona
Who's neededCleanroom technicians suit up in head-to-toe "bunny suits" to run the photolithography machines that print billions of transistors. New fabs like TSMC's Arizona site are hiring thousands of technicians.
🎓 Many of them started with a two-year technical degree.
🔦 Also on the payroll: fiber crews splicing hair-thin glass cables · chip designers · model researchers · safety testers · app builders. Next Tuesday at 7:43 a.m., when a kid asks a chatbot about photosynthesis, thousands of these people — on layer after layer, across three continents — answer it together.
🤯Whoa, Really?
☀️ Hotter than the sunTo make the light that prints the tiniest chip features, ASML's EUV machines blast falling droplets of molten tin with a laser tens of thousands of times per second — each flash creates a plasma far hotter than the surface of the sun.
🧵 Thinner than a hairMore than a thousand transistors could line up across the width of one human hair — and a single AI chip about the size of a postage stamp and a half holds over 80 billion of them.
🗺️ A country of serversIf the world's data centers were a country, their electricity use would rank in the top ten on Earth — the IEA found them landing between Saudi Arabia and France.
⭐Remember This
1AI is physical. Every chatbot answer runs on real silicon chips, real warehouse-sized buildings, and real electricity pulled from the energy grid.
2Train again and again; answer constantly. Training changes the model's learned numbers in giant bursts of energy — and labs repeat it through fine-tuning, fixes and new versions. Inference is what a finished version does billions of times a day, so small per-answer savings multiply into enormous ones.
3Nobody builds AI alone. Chip designers, one Dutch machine maker, Taiwanese fabs, cloud providers, model labs, and app makers form one interdependent supply chain — and hiring happens on layer after layer of it, not only at the top.
🗼 Two seconds is never simple. Ride the stack backward from the glowing answer on your screen all the way down to sand — and remember that pulling out any single layer makes the whole tower go dark.
✏️ ClickClass Anchor Chart · The Full Stack: From Power Plant to Chatbot
From ClickClass — hundreds of free printables at clickclassedu.com/printables