The argument about AI spend is usually one number pretending to be three. What it costs to stand the racks up. What a lab might earn on that power. What the tokens would be worth if you sold them at a list price. Those are not the same claim. We built a page that keeps them apart.
The live model is here: What Is One Megawatt of AI Worth?. The math is TypeScript. An LLM does not compute it. You can change the assumptions. The engine still does the arithmetic.
I used to sell this kind of tool for a living. The failure mode is always the same: a blended figure that cannot tell you which input moved. So the page labels every bar.
The 1 GW picture
Default is owned, 1,000 megawatts. Capex is $12 million per MW, the middle of Dylan Patel's $10–15 million all-in for GPU plus facility. That is $12 billion to stand up a gigawatt. It is a one-time check. If you switch the page to rented, the investment bar becomes annual rent at the same $12 million per MW. The model will not add rent and capex in the same scenario. Do not do it in your head either.
That $12 billion is not revenue.
Three bars, not a mix
The hero is three labeled pictures at 1 GW.
Investment / capex. $12 billion owned, as above. Estimate. Dylan, 25 August 2026, on Dwarkesh. Range $10–15 billion if you take the ends of his band.
SemiAnalysis lab API revenue. $50–100 billion a year at 1 GW, which is $50–100 million per MW. Anthropic is cited as high as $50 million per MW, with a path to $100 million. The line they use is that 1 GW produces $100B of revenue. That is an estimate of what a frontier lab might earn on that power. It is not a 10-Q. It is not the token engine.
Our BASE token engine. About $306 billion a year at 1 GW. Throughput 2.8 million tokens per second per MW, 70% serving hours, DeepSeek-class mix that blends to about $4.95 per million tokens. 2.8 million times 70% times a year times $4.95, times 1,000 MW. That is list-price token math on an InferenceMAX snapshot (B200, all-in utility MW, cache off, 11.1% input / 88.9% output). Epoch's TCO is why utilization is 70%, not SemiAnalysis. PUE is not applied on top of 2.8 million. The benchmark is already utility megawatts.
$306 billion is not Patel's number. It is not a lab print. It is what those tokens would bill at those prices if the mix and the uptime held. If you think list is the wrong price, move the slider. The engine will move the bar. The lab range will not.
SPCX is on the page as a filing, not a third revenue theory. Q2 2026 AI revenue $2.561 billion, AI capex $15.828 billion, nameplate 1.4 GW, 8-K Exhibit 99.1 / 10-Q filed 4 August 2026. Implied at 1 GW scale that is about $1.8 billion of quarterly revenue and $11 billion of quarterly capex. A quarter. Not annualized. Not Dylan's $100 million per MW. Hyperscalers and NVIDIA are omitted because those filings do not give us an AI dollar per megawatt to put on the same table.
The $11 trillion path
Separate block. Dylan on Dwarkesh, 25 August 2026, and SemiAnalysis, "Nvidia GPU Debt Backstop," July 2026.
Cumulative AI capex 2024 through 2029: about $11 trillion. The SemiAnalysis note prints $11.1 trillion for the same window. Funded $6 trillion cash and $5 trillion of new credit. $7.1 trillion is the stock of AI debt outstanding in 2029. It is not the $5 trillion. Do not add them.
GW he named: about 30 this year, 50 next, 70 in 2028, 90–100 in 2029. Annual dollars he named: 2026 "a little bit over a trillion dollars of CapEx," 2028 "more than $2 trillion" of IT, with an all-in callout of $3–4 trillion. We do not draw bars for 2024, 2025, or 2027. He did not give those years as a dollar figure, so the page does not invent one.
A 100 GW year is a different thought: IT about $5 trillion, with plants and shells built ahead $7–10 trillion. That is a 2030-ish run-rate, not the 2024–2029 stack.
All of that is an estimate. None of it is a filing. None of it is the $306 billion token bar.
What the page is for
You can argue the lab will never see $100 million per MW. You can argue list-price tokens are a fantasy. You can argue $12 million per MW is light once the substation is in the number. The page will let you say that with the inputs in the open. It will not let you stack capex on rent, or fold Patel's credit line into his debt stock, or replace a lab estimate with token math and call it one print.
I am not going to scrape 10-Qs live into this model. When a filer gives us an AI dollar per megawatt, it goes on the table as a filing, the way SPCX does. Until then the three pictures stay three pictures.
Open the One Megawatt model. Move one input. Watch which bar moves.
Stay sharp.
— Stock Alarm