How Many GPUs
Would It Take?

A single GPU could read every federal statute and regulation in less than a day, and keeping up with new law costs almost nothing. The U.S. Code takes 1.3 GPU-hours to read; adding the Code of Federal Regulations brings the job to 6.9 GPU-hours.

The U.S. Code runs 24.4 million words, which a modern tokenizer turns into about 37 million tokens, and the regulations add more than 100 million words. GPUs read in parallel, and reading is the cheap half of their work. The recurring cost is the daily Federal Register and new public laws: a few hundred thousand tokens a day, or less than 0.01 GPUs.

Most legal text is case law. Every published American opinion comes to about 19 billion tokens, and the world’s published opinions to roughly 350 billion, a one-time job of 12,200 GPU-hours.

Reading isn’t legal analysis. Checking every provision against every other multiplies the bill many times over. A sensible system reads once, builds an index, and pulls the relevant sections for each question.

Watching one livestream all day takes a small slice of one GPU. Watching every surveillance camera on Earth takes millions. One stream at default resolution needs 0.02 GPUs around the clock, and catching up on its last 30 days takes 0.5 GPUs to do this in one day.

Video models don’t watch the way people do. Gemini samples one frame per second and turns each into 70 tokens by default, or 280 at high resolution, plus 32 tokens for every second of audio. A day of one stream is about 8.8 million tokens at default resolution and 27 million at high resolution.

At scale, the totals jump. Every Twitch, YouTube and Kick stream plus 4,000 satellite channels needs 2,400 GPUs. The world’s roughly one billion surveillance cameras need 13 million GPUs, and indexing the whole YouTube library once takes 1.3 million GPUs to do this in one day.

Two caveats apply. One frame per second misses fast motion, and naming strangers or reconstructing a day’s events is harder than describing a scene. Those are accuracy problems, not just more tokens. Video priced at text-serving rates is also an approximation: a real pipeline runs vision and audio encoders too.

Defending 1,000 servers takes about 5.7 GPUs, as long as ordinary software condenses the logs before a model reads them. Re-checking the last 90 days once takes 405 GPUs to do this in one day.

A thousand busy servers produce around 10,000 security events a second. Send every one straight to a language model and it reads about 216 billion tokens a day, using 255 GPUs before it writes a single verdict.

Real defense tools condense first. The planning design has each server send one 2,000-token evidence packet every 100 seconds and runs 1,000 deep investigations a day, which is where the number above comes from. Elastic reports its triage agents settle at 7 to 9 model calls per investigation after tuning.

Compute isn’t the limit here. No amount of reading can find an attacker the logs never recorded, and a wrong containment decision is expensive at any GPU count.

Flying 20,000 drones takes almost no AI. Asking a language model to fly each one takes hundreds to thousands of GPUs, and it still wouldn’t work.

Autopilots already fly drones with plain arithmetic. In February 2026, 22,580 drones flew at once from a single computer, on paths calculated in advance. Planning such a show is a one-time job: a model writes 31 formations and the show software fills in the paths, which takes 1.5 GPU-hours.

Put a model in charge of 200 squads, briefing each every 10 seconds, and it needs 8 GPUs. Give every drone its own decision each second and it needs 800 GPUs. Let each one reason for 1,000 tokens first and it needs 10,600 GPUs.

The GPU count is the smaller problem. In a published benchmark with 16,384 simultaneous requests, a large reasoning model took almost six seconds to start answering and then wrote about 27 tokens a second per request. A drone at 10 meters a second covers 60 meters on stale orders in that time. Robot models avoid this by thinking a few times a second and passing faster motor commands to a small controller.

An AI can label everything a 100,000-person company writes in a day on about 6.3 GPUs. Thinking hard about each message takes 178 GPUs. Reading the kept archive once, about two years of de-duplicated email and chat, takes 1,040 GPUs to do this in one day.

The average Microsoft 365 worker receives 117 emails and 153 Teams messages a day, but most are copies of something sent to many people, and Teams chats are copied into every participant’s mailbox. Counted once, that is about 100 messages per person. With quoted replies removed, the median email reply is 43 words.

A short label per message, such as “payroll” or “customer complaint”, adds little. Writing 1,000 tokens of reasoning about every message multiplies the output a hundredfold, and writing is the expensive half. Scaled to all 450 million paid Microsoft 365 seats, even the light version needs 9,380 GPUs.

Reading every person’s medical record once is a bounded job; the daily bill depends on whether each workup re-reads it. Reading all 8.3 billion records and writing a summary of each takes 211,000 GPUs to do this in one day. After that, a daily workup for everyone in a care episode, using the summary plus new notes, takes 26,900 GPUs around the clock.

Who counts is a choice, not a measurement. On an average day about 1.5% of humanity sees a clinician. A classic study found that each month, about a third of people consider seeking care, which works out to about 5% a day if each episode lasts five days. We plan for 5% and show the other choices as separate bars.

Record length varies enormously. The median patient arriving at a U.S. emergency room in 2022 had 98,000 tokens of notes on file, and one in five had more text than Moby-Dick. Most of the world has far shorter or paper records, so a global average is closer to 12,000 tokens.

That is why the design matters more than the population. Reading a full U.S.-sized 200,000-token chart at every workup takes 117,000 GPUs. Re-sending it with each of eight questions takes 789,000 GPUs. A full assessment for every person, every day, takes 5.9 million GPUs, and the most wasteful version needs 150 million GPUs, more AI chips than exist.

One warning applies to every medical number here. The serving speeds were measured on 2,000-token prompts, and long records cost more per token as attention grows with length. A real system would read records in windows of 32,000 to 64,000 tokens and retrieve what each question needs, so these figures are reference work rather than a tested design.

Arguing every new court case in the world, each at the depth its file needs, takes about 202 GPUs running all year. Clearing the existing backlog once is as much reading and writing as a full year of new cases. Arguing the roughly 280 million pending cases takes 74,300 GPUs to do this in one day.

The world opens roughly 280 million court cases a year. U.S. state courts alone receive 70 million filings, most of them traffic tickets. Chinese courts accepted 37.5 million cases in 2025, about 29% of them enforcement stages of earlier disputes, and India has a backlog of 54 million.

Most files are short. A traffic case is a page or two, and even a federal appeal brief is capped at 13,000 words. The expensive American habit is discovery, the exchange of internal documents before trial: a large corporate production reviews about 100 GB, nearly half a billion tokens per case. Give that to every case and the bill becomes 466,000 GPUs.

Token budgets are the biggest unknown. A far more generous docket, with two million tokens for every contested case and ten billion for each of the largest discovery fights, needs 2,790 GPUs.

Watching every televised game and updating win probability on every play takes about 16 GPUs at the busiest hour of a Saturday. Sports produce few decision points: a handful per second, worldwide.

Sportradar streamed over 525,000 matches in 2025. Games average under two hours, so roughly 135 are live at any moment and about 500 at a Saturday peak, when European football overlaps with American college football. A 2.5-hour game at high resolution is about 2.8 million tokens.

Production systems compute probabilities with small statistical models; NFL Next Gen Stats runs 75 of them per play in under a second. Having a language model re-reason every play instead raises the peak to 20 GPUs. Reading decades of play-by-play once to prime the captions takes 258 GPU-hours.

As a stress test, watching every data-covered match with 3,000 feeds live at once and a detailed update every 45 seconds would need 124 GPUs.

Rewriting every app in both stores is a one-time job of about 6,220 GPUs to do this in one day. Keeping them all current afterward takes about 5.6 GPUs.

The App Store lists about 2.6 million apps and Google Play about 2.6 million, but 37% of apps are on both, so there are roughly 3.4 million distinct products. Most are small, and a third haven’t been updated in two years. A few giants pull the average up: Uber’s Android codebase grew from 10,000 lines in 2008 to 10 million in 2021.

Code runs about 7 to 10 tokens per line. A coding agent reads far more than it writes, but most of those reads are cached: in 731 benchmark sessions, 97.6% of input tokens were cache hits, which cost almost nothing to serve again. We count new tokens only.

A heavy loop that explores, rewrites, tests and reviews every app takes 22,100 GPUs to do this in one day, and review is a large share: in one study of agent traces, code review used 59.4% of all tokens.

A model that checks every order a person places, worldwide, needs about 2,360 GPUs. Reading every filing and a decade of financial news first takes 564 GPUs to do this in one day.

Owning stock isn’t trading. China has 251 million securities investors and India 238 million demat accounts, but only 8% of India’s retail traders traded on more than 50 days last year. Counting orders people place themselves gives about 120 million a day, led by China, where Shanghai and Shenzhen logged 68.8 billion two-sided transfers in 2025.

An always-on agent that briefs each of 40 million active traders every 15 minutes during their market hours needs 8,890 GPUs. Calling a model on every order and cancel, including algorithmic flow, needs 39,500 GPUs. Market ticks themselves aren’t read as text.

These are averages across the day. Load peaks at about three times the average when Asian markets open, so a real deployment would provision for the peak.

A robot helper in every home, with its planning done in a datacenter, needs about 85 million GPUs running around the clock. Even on OpenAI’s faster Jalapeño chips at 12,000 tokens a second, that is 49% of all the AI chips projected for 2030. It is the only job here on that scale, and a robot for every person needs 1.7 billion GPUs.

There are about 2.2 billion households and 8.3 billion people. A robot that listens all waking day hears 32 audio tokens a second. Robot models plan a few times a second: Gemini Robotics takes about 250 milliseconds from camera images to an action chunk, and each second of motion encodes to about 30 tokens per arm.

The one-time cost is scanning each home and learning the family: 3.4 million GPUs to do this in one day. The recurring cost is what matters, and it depends on where the robot thinks. If robots run their models on board and call the cloud only for hard questions, datacenter demand drops to 624,000 GPUs. Four 30-frames-per-second cameras with nonstop reasoning for every person would need 40 billion GPUs.

Bodies, not tokens, limit this decade. Goldman Sachs projects 890,000 humanoid robots shipped in 2030, a tiny fraction of 2.2 billion homes. If the robots arrive, their datacenter demand would exceed the planning estimates of every other job on this page combined, many times over.