Kyle Corbitt has named a price at which he will believe artificial general intelligence has arrived: about $30 an hour to rent one H100, a standard Nvidia chip for running AI. Writing on his personal site, corbt.com (the post is credited “with Claude Opus”), he sets the bar at an agent that can “outperform humans at most economically valuable work,” and says renting an H100 costs $3.52 an hour as of October 2026, up 18% over the year. His claim is that the rent would need to rise roughly tenfold, because supply is capped by chips that already exist.

It is a prediction with a number, a mechanism and a date to check it against, which is rare in AGI talk. Most forecasts in this field cannot lose. This one can, and that makes it worth working through.

The $30 comes from a short chain, and the post shows every link. Corbitt takes Claude Opus 5.5, which he calls the most capable public model, and prices it as a full-time agent at list rates: $8.21 an hour. He assumes Anthropic spends half of that on compute, which at $3.52 per chip-hour means 1.17 chip-hours for each agent-hour. One chip can therefore run about 0.86 agents around the clock. If each of those agents is worth the $35 an hour that the 95 millionth best-paid worker earns, a chip-hour is worth 0.86 times $35, or $30.01.

The $35 floor is itself derived. Corbitt estimates the world’s compute (about 25 million H100-equivalents serving customers, after labs keep half for research) could run agents standing in for roughly 95 million full-time workers, since an agent works 8,760 hours a year against 2,000 for a person. He then asks what the 95 millionth top earner makes. US census-based figures are measured; for the rest of the world he says he could not find a count and estimates. That puts the floor at $31 to $40.

Several inputs carry the result. One is that an AGI agent needs no more hardware than Opus 5.5 does today, which he justifies by costs of reaching a given capability falling 5 to 10 times a year. Another is the continuous token-rate assumed for an agent. A third is that Anthropic’s compute is bought near market price, though he notes its reported xAI deal may work out to $6 to $7 per chip-hour. Loosening the assumptions (higher margins, a cheaper model, employer costs beyond wages) gets him as high as $75. By his own arithmetic the conservative case is about 8.5 times today’s rent, so “tenfold” is rounding. At today’s price he says agents would want about 28 times the compute that exists.

The weak points are where the post is quiet. Corbitt names two outcomes if demand outruns supply: rationing, where labs stop selling compute externally, or higher prices. He does not address the other way rent could jump, a shock to supply such as export controls, a fab problem or a power shortage, none of which requires any change in what AI can do. He also offers no explanation for the 18% rise already recorded this year, which he calls surprising. An indicator that moves for reasons unrelated to capability is not clean, and the post does not say how to separate them. The unit is a stand-in too: Anthropic also runs on TPUs, Trainium and newer Nvidia hardware, so Corbitt allows that newer silicon could climb faster than the original H100.

“Most economically valuable work” is borrowed wording, and Corbitt does not define it. His model quietly defines it for him: any job paying $35 an hour or more, with every earner treated as a white-collar stand-in.

Compute prices are also the hinge of other arguments in today’s issue, including an investor’s case that cheaper models raise compute demand as long as someone funds training, and a measurement of whether model routers actually cut costs. Anyone holding a view on those can test it against the same line. If the rent for these chips is still near $3.52 next autumn, Corbitt’s own logic says agents are not yet doing the work of 95 million people.

Based on Kyle Corbitt’s post on corbt.com, October 2026 (credited as written with Claude Opus)