# Your Worth In GPUs AI-generated; not human-verified. Timestamps are approximate. A presenter asks what human labor is worth when AI can do the same work. ## Timestamped description [00:00–00:01] A banana-headed character wearing a business suit speaks before rows of Nvidia graphics cards. Synchronized subtitles appear at the bottom. [00:01–00:05] Microchips with Texas Instruments branding rain down behind the character, followed by shifting screens showing digital code and electronic components. [00:05–00:10] The speaker gestures against a backdrop of automated factory machinery and circuit boards. [00:10–00:13] The scene switches to colorful maze-like abstract patterns with an economic equation overlaying the screen: LA = phi(Kc). [00:13–00:15] Behind the character and the formula, a cleanroom semiconductor laboratory with workers in bunny suits appears. [00:15–00:18] A banana-headed man in a suit speaks in front of cleanroom workers and a formula for agent labor output. [00:18–00:22] The speaker gestures beside economic labor market supply and demand charts. [00:22–00:25] The speaker talks with split cleanroom and fab backgrounds behind him. [00:25–00:28] The speaker appears in front of tech campus facilities, including a Texas Instruments building. [00:28–00:30] The speaker finishes his statement inside a chip manufacturing cleanroom facility. [00:30–00:32] A banana-headed man in a suit speaks in front of server racks and a microchip background with animated word captions appearing below. [00:32–00:36] A 3D economic graph displaying rental rates and compute curves appears next to the speaker against tech manufacturing backdrops. [00:36–00:39] The speaker duplicates into two identical figures standing in front of an automated assembly conveyor line. [00:39–00:45] A mathematical inequality appears above multiple animated banana men, with labels identifying compute intensity, productivity ratio, and rental rate. [00:45–00:47] A banana-headed animated man in a business suit speaks before technical diagrams, formulas, and digital imagery. [00:47–00:52] Banana-headed figures present an equation calculating hourly compute cost ($2/hour) in front of semiconductor manufacturing equipment. [00:52–00:56] The speaker appears before factory machinery and cleanroom technicians in white bunny suits. [00:56–01:00] The speaker gestures dynamically against shifting backgrounds of semiconductor chips, waveforms, and technical schematics. [01:00–01:01] A banana-headed humanoid in a suit speaks against an abstract tech background with vertical circuitry elements. Large animated text captions appear at the bottom. [01:01–01:03] The speaker appears centered against an amber-colored microchip circuit board background while speaking and gesturing. [01:03–01:05] The background switches to an automated industrial factory floor with machinery and workers in protective gear. [01:05–01:07] Glitchy transitions show purple circuitry and electronic circuit boards as the banana-headed man continues his explanation. [01:07–01:09] Multiple clones of the banana-headed man appear alongside him against an industrial factory background. [01:09–01:11] Glitchy neon background shifts to robotic machinery in a lab setting as the speaker references legal roles. [01:11–01:15] An economic supply and demand graph appears over a legal office background featuring a statue of Lady Justice and legal documents, with the speaker explaining substitutable work. [01:15–01:17] A banana-headed man in a suit speaks in front of an economics chart with supply and demand curves. Dynamic text overlays words in sync with speech. [01:17–01:19] The speaker stands in an office while clones of himself appear behind him amid digital glitch effects. [01:19–01:21] Rapid flashing color inversions and camera pans show the banana man speaking with animated hand gestures. [01:21–01:23] The speaker gestures broadly in an office lobby with glitching digital background monitors. [01:23–01:26] The background switches to an industrial semiconductor cleanroom interface display. [01:26–01:28] The backdrop shifts to a wall of tech screens, followed by circuit boards flying past a dark background. ## Speech transcript [00:00–00:02] Speaker 1: How many GPUs does it take to get rid of you? [00:02–00:04] Speaker 1: If you can spin up new AI agents for almost nothing, [00:04–00:06] Speaker 1: just the cost of running a GPU, [00:06–00:08] Speaker 1: then wages for cognitive work collapse to zero. [00:08–00:09] Speaker 1: That's the conclusion most people are drawing. [00:09–00:11] Speaker 1: And it's kind of right, [00:11–00:13] Speaker 1: but the reasoning to get there is backwards. [00:13–00:14] Speaker 1: So what's an agent? [00:14–00:15] Speaker 1: You can model that [00:15–00:18] Banana Man: could work, comes out as tech and not workers. That distinction is the important part. [00:18–00:22] Banana Man: If you call agents a form of labor, that puts the pricing question in the labor market, where infinite supply would clear wages at zero. [00:22–00:25] Banana Man: But agents inherit their supply from compute. That means they need fabs and energy, [00:25–00:28] Banana Man: as well as water, land, and the right passport to buy chips. [00:28–00:30] Banana Man: All of this is finite and slow to expand. [00:30–00:32] Speaker: So the thing that can change is not the amount of labor, but the amount of compute. [00:32–00:36] Speaker: Wages on cognitive tasks that can be automated now get priced into the compute market, scaled by how good the model is. [00:36–00:39] Speaker: The wage ceiling for these substitutable tasks comes out to three things multiplied together: [00:39–00:41] Speaker: 1. How productive you are relative to an agent. [00:41–00:43] Speaker: 2. How much compute one hour of agent work burns. [00:43–00:45] Speaker: 3. The rental rate of that compute. [00:45–00:46] Speaker: Let's plug in today's numbers. [00:46–00:47] Speaker: H100s rent for about $2 an hour. [00:47–00:51] Speaker: A frontier model burns roughly 1 GPU-hour per hour of senior knowledge work. [00:51–00:53] Speaker: If you're as productive as the agent, your ceiling lands around $2 an hour. [00:53–00:56] Speaker: But for small distilled models doing more trivial work like classification or summarization, [00:56–00:57] Speaker: 5 to 20 cents an hour. [00:57–00:59] Speaker: That ceiling already binds on high-volume work, [00:59–01:00] Speaker: and every algorithmic... [01:00–01:00] Speaker: pushes the ceiling down. [01:00–01:01] Speaker: So, here's the thing. [01:01–01:04] Speaker: Old theory said high-skill cognitive workers win, while routine, unskilled workers lose. [01:04–01:07] Speaker: But with AI agents, the split actually runs through cognitive work itself. [01:07–01:09] Speaker: The dividing line moved inside the white-collar category. [01:09–01:11] Speaker: Take a junior paralegal and a senior litigation partner. [01:11–01:14] Speaker: A paralegal does clause extraction, redlining, summary memos. [01:14–01:15] Speaker: That's substitutable work, [01:15–01:17] Speaker: So wages get squeezed toward the compute ceiling. [01:17–01:19] Speaker: The partner does strategy, client management, courtroom work. That's complementary work. [01:19–01:21] Speaker: And agents make the partner sharper and higher leverage, so her wage rises. [01:21–01:23] Speaker: Same industry on paper, but a world of difference between them. [01:23–01:26] Speaker: So stop thinking about the labor market and start thinking about the compute market. [01:26–01:28] Speaker: It's really about how many GPUs it takes to do your job. Generated with google/gemini-3.8-flash. Video: https://pleometric.net/videos/your-worth-in-gpus/