How it was counted

They came from X's paid search, bought through the X Developer Console at half a cent a post. On we asked for every original post since Nov 7 with 100 likes or more that mentions Jensen, Huang, "Nvidia CEO", a honeypot, a Chinese spy or the ticker $NVDA. Reposts and replies were left out. X returned posts and charged .

The search is wide on purpose. An earlier, tighter one asked for his name and a scandal word in the same post. Tested against 80 joke posts we had already found by hand, it caught 14. The wide one caught 51. Most jokes do not spell out what they are about.

A wide search brings back a lot that is not a joke, so every post was then sorted:

• A rule first. posts link out to another website. Those are treated as news or promotion. None of the 80 hand-found posts carried a link.
• Then reading. The other were each read by an AI model, Claude, working from a written rubric, and marked as a joke, news, opinion, off topic, or "could be either".
• Only confident calls are kept. posts were marked as jokes. The marked "could be either" are left out.
• A meme needs a picture. Of those jokes, carry an image, a GIF or a video. Those are the memes. The other are jokes in words alone and are not listed.
• Plus the ones search missed. image and video memes featured in news roundups did not match the search words or the 100-like floor. They are included, with their numbers read from X the same way.

The roundups the hand-found posts came from, with how many of each one's picks are memes on this site:

How good is the sorting? Of the 45 hand-found posts that the search returned, the reader marked 39 as jokes, 5 as "could be either" and 1 as off topic. In a random 45 of its "joke" calls that we checked by hand, 43 were plainly jokes. So expect a few percent of the memes here to be something else, and some real memes to be missing.

What is not here: posts with fewer than 100 likes, replies, deleted posts, and jokes that name nobody and use none of the search words. This is a large sample of the popular memes, not a count of every one.

Every X post has an ID number, and the first 41 bits of that number are a clock: the count of milliseconds since X's own starting point in November 2010. Shift the ID right by 22 bits, add the starting point, and you have the moment the post was created.

The API also gives the time directly, and the two agree. All times on the site are UTC. California was eight hours behind that week, New York five.

X's API will count the posts that match a search, hour by hour, without handing over the posts. One count costs half a cent. Using the same words as above, with reposts included, it reported posts in the seven days to .

• It counts words, not meaning. The total includes ordinary talk about Nvidia's shares and anyone else called Jensen or Huang. In the day before the story broke, the same search ran at about posts an hour. Treat that as the background level.
• X rounds these counts. They are close estimates, not an audit.

Each post arrives from the API with its own counts of likes, replies, reposts, quotes, saves and views, and with the type and length of any photo, GIF or video. Together the memes had likes and views when they were read.

• The counts are a snapshot. They keep moving after they are read, and the newest posts have had the least time to collect them.
• Account names are looked up separately, from the public record X serves for embedding a post, and only for the most liked memes. The rest are listed without a name.

• What counts as a meme. A post whose main purpose is a joke about the story and that carries an image, a GIF or a video: edits, reaction clips, captioned pictures, parody videos. One-liners in plain text are jokes but not memes. Straight news, sincere opinion and investing talk are out. The line is drawn by a reader and another reader would move it a little.
• Joke types. Each meme was put into one of groups by the same reader. The hand-found memes were grouped by hand.
• Titles. The most liked memes have a short description written for this site. They are not quotes. The rest are listed by joke type.
• AI-made. A video is marked AI-made only when a news source or the poster says so. That gives videos. The true number is certainly higher.

Nvidia's share price is shown beside the memes so the two can be read on one clock. Prices are half-hourly, from Yahoo Finance, and include trading before and after the regular New York session, which is drawn lighter because few shares change hands then.

The price is context. The site makes no claim that jokes moved it, and nothing here is investment advice.

Nobody publishes what a meme cost to generate, so this is a rough estimate from stated assumptions. It covers the AI-made videos only. Text posts and hand-made edits are counted as zero.

• Tokens. A video model works on a compressed grid. At 720p that is about 80 by 45 patches a frame, and 24 frames a second compress to about 6. That makes 21,600 tokens for each second of video.
• Server time. About 22.5 seconds of one data-centre GPU for each second of video, which is 3 minutes for an 8-second clip.
• Attempts. People rarely post their first try. We assume 6 attempts for every clip posted.
• Power. 0.7 kW per GPU, cooling and overhead included. A US home uses about 29 kWh a day.

The detector figure is from Copyleaks, as reported by TechCrunch on Nov 9: at least 70 AI-made posts about the story on X, with about 8.42 million views between them.

• The site links to posts and never copies or hosts the images and videos.
• The affair is described as reported. The other person involved is not named in our own words.
• A post only loads from X when you ask it to, so X sees nothing about you until then.
• Visits are counted with Google Tag Manager. Nothing else on the site tracks you.