deadrisk

Analysis

After Scarcity, Who Gets Paid

A post on X says compute will stop being scarce and capital will matter more than ever. Run both through DeadRisk's forecast and the capital loses: the chip makers are paid, the data-centre owners and the labs are not. Keep power short and the owners are paid again, which makes the post's line about compute in space the one that matters most to them.

SPCXNVDA
Coverage intelligence, not investment advice — methodology.

The post

On 26 October Jacob (@jsnnsa) posted a list of views on where AI takes the world. Most of it no market can price: new religions, a pope, living until the heat death of the universe. Four lines bear on who gets paid for the AI build:

  • “work ends”
  • “you can throw money at problems that used to require time and careful talent picking so capital matters more than ever”
  • “specialist models get used while compute is scarce. compute will mostly not stay scarce”
  • “99.999% of compute is in space by 2040”

The second and third pull against each other for anyone who owns the build. Capital buys the compute. If compute stops being scarce, its price falls toward what it costs to run, and the capital that bought it earns less than it cost. DeadRisk’s forecast scores every layer on that test: what it earned from 2026 to 2035 against its costs and its cost of capital.

His world, as inputs

The forecast’s “Situational Awareness: the race” world comes closest to the first two lines. It was built from Leopold Aschenbrenner’s 2024 essay. AI does half its eventual work by 2028 and nearly all of it in the end. Investors put $500 billion into the labs in 2027. Chip, memory and grid supply double. Power can reach 15% of the world’s electricity. In that world compute stays scarce: the best models can charge up to five times their compute cost.

The third line takes two sliders. Set the frontier markup (“…but at most this many times its compute cost”) to 1.5× rather than 5×, so rivals and open models undercut any price above that. Set “Compute rents: pricing discipline by 2030” to 0.50, so rents recover half the fleet’s full cost when supply meets demand, not all of it. Both values come from the forecast’s “Commodity intelligence” world. Load the race, move those two, and the page shows the second row below.

World Fleet owners Labs Chip makers
The race, as published +$5.0T −$10.3T +$10.0T
The race, compute plentiful −$8.7T −$14.2T +$14.4T
…and power binds at 5% +$3.6T −$10.3T +$5.0T

Figures are 2026–35 results over each layer’s cost of capital, read from the model on 26 October. The third row starts from “…and power binds” and moves the same two sliders.

The owners go from $5.0 trillion ahead to $8.7 trillion behind. More demand does not save them. In “Commodity intelligence”, which has the same markup cap and rent discipline but slower demand and a smaller frontier share, they finish $3.9 trillion behind. With the race’s demand they build more at prices that do not cover it, and the extra fleet loses money. At a 1.5× markup the owners stop covering their cost of capital once rents recover between eight- and nine-tenths of the fleet’s cost.

The labs lose more too, because they cannot charge much over their compute. The challenger lab now runs out of money in 2033 rather than 2030: cheap rent keeps it alive longer. The chip makers take more, on $35.6 trillion of hardware sales against $27.8 trillion in the race. The model has them earning their cost in every world it names, so their gain tells us less than the others’ losses.

Part of the owners’ loss is built in: telling rents to fall is telling the owners to earn less. The model adds two things the inputs do not say: how large the loss grows with the race’s demand, and the scarcity that undoes it.

17 Nov 2025
The date of the first story on computing in orbit in TEXXR's record: the Wall Street Journal on Musk, Bezos and Pichai
None before it in a record that reaches back to 2006; 27 records to 22 August 2026, some of them repeats; one in the two months since, on Google's test satellite

Why the space line matters most

Keep power short and the owners are paid again: $3.6 trillion ahead on the same prices, against $9.9 trillion at the race’s own. When power binds, whoever holds powered capacity collects a scarcity rent, even at the lower prices. In this world, power is the one scarcity left that pays the owners. “99.999% of compute is in space by 2040” is the claim that removes it, since power in orbit comes from the sun, not from a grid queue. For owners of powered sites on the ground, orbit is the threat in the post, not the promise. The model runs to 2035 and cannot test his date.

TEXXR’s record picks the idea up late. It holds no story on computing in orbit before the Journal’s 17 November piece. Starcloud then ran a Google model on a satellite with an Nvidia H100. Blue Origin asked the FCC to deploy nearly 52,000 satellites for its project. SpaceX told investors orbital AI data centres use “unproven technologies” and may not achieve “commercial viability”. Reuters’ sources said it aims for orbital tests by late 2027 and has asked to launch up to a million data-centre satellites. In September the Times reported that Google planned to launch a test satellite with four TPUs on a SpaceX rocket on 1 October, to run for a year.

Sam Altman called the idea “ridiculous” for now and said it would not matter at scale this decade. Masayoshi Son noted, in the Journal’s account, that electricity is just 7% of costs. That fits the model: the price of power matters little, and the question is whether a builder can get it at all.

SpaceX’s third-quarter 10-Q names data centres 13 times, never in orbit: its ability to “obtain sufficient power, GPUs”, the gas turbines at Colossus II, and a $34.2 billion rise in nine-month capital spending, on data centres and launch sites. Every orbit it mentions belongs to a launch or a Starlink satellite.

Scarce for now

On the record, memory and finished sites are getting scarcer, not less scarce. SK Hynix’s chief executive expects the memory industry’s worst-ever shortage in 2027, with demand ahead of supply beyond 2030. SynMax found almost 40% of US data centres due in 2026 facing delays. The owners’ side is already strained: on 26 October Musk, whose SpaceX has closed $40 billion of GPU debt, said debt markets had shut off credit for AI data-centre companies and that 200-plus neoclouds face “death and consolidation”. He is an interested party. The post may be right about the end. The model’s point is about who is paid along the way: while something stays scarce, its owner collects, and when nothing does, the money that built the fleet does not earn its cost.

What to watch

  • SpaceX’s orbital tests, aimed for late 2027 by Reuters’ account. A working test is the first evidence about the post’s 2040 date.
  • Rents against cost. At a 1.5× markup the owners’ result turns on whether rents recover more than eight- to nine-tenths of the fleet’s cost once supply catches up.
  • Memory supply in 2027. SK Hynix’s shortage call is the nearest test of “compute will mostly not stay scarce”.
  • Google’s test satellite. Whether it flew on 1 October and what a year of four TPUs in orbit shows. TEXXR has no report since.

Sources

Dates are the publisher’s where the page shows one, otherwise TEXXR’s.

Disclosure: this piece was drafted with Claude, a model made by Anthropic. The post it answers names Claude.