THE POST
@zephyr_z9 Silly analysis. Robot RAM needs will exceed everything else on Earth, as you need the robots to be on physical substates in order to shape atoms. Super intelligence needs will vastly exceed that, but ~100% of SI will be deep space compute. Obviously.
REPLIED TO
Zephyr@zephyr_z9
Interesting So they are assuming annual production of 550k humanoids and total DRAM demand of 21.5 PB Even if annual production was 20X (11M), total DRAM demand will be 430 PB or 0.43 EB DRAM production in 2026 will be around 48EB So, 21.5PB is around 0.04% of 2026 DRAM production
https://x.com/zephyr_z9/status/2114662531709178306This post itself quotes another post, which is not copied here.
ENGAGEMENT
- likes8,256
- reposts590
- replies386
- quotes217
- bookmarks1,875
- views1,243,196
AS RETURNED BY THE X API · FETCHED 2026-10-27 06:19 UTC
WHAT IT SAYS
Replying to an analyst who calculated that humanoid robots would use a tiny share of world DRAM output, Musk calls the analysis silly. He says robot memory needs will exceed everything else on Earth, because robots are what physically shape atoms, and that superintelligence will need far more again, with about 100% of it running as compute in deep space.
CONTEXT
SpaceX's Starmind page describes its first AI satellite as a solar-powered compute node in sun-synchronous orbit, with up to 250 kW peak and 175 kW average of compute payload, sending results back through Starlink lasers. It says a Gigasat Factory in Bastrop is planned to build thousands of these satellites starting as soon as late 2027, and names mass to orbit, power and AI chips as the constraints to solve. That is near-Earth orbit, not deep space.
The FCC accepted SpaceX's application for an orbital data centre system of up to one million satellites at 500 to 2,000 km for filing in February 2026 and asked for comment. Accepting an application for filing is not an approval.
SpaceX's CFO said in September that the company is targeting its first orbital compute satellites next year, with large amounts of compute going up into 2028. US Brief notes that none of the power or size figures have been tested in orbit, and that cooling, radiation, repairs and launch cost remain open problems.
WHAT IT MEANS FOR K
If compute moves off-planet onto sunlight that never reaches Earth's surface, it is the most direct path past Type I toward harvesting a real share of the Sun's output. Today it is a forecast with a first test planned for 2027, not harnessed power.
CLAIMS
- Robot memory needs will exceed everything else on Earth.
- Superintelligence will need vastly more than that.
- About 100% of superintelligence will be deep space compute.
EVIDENCE
- SpaceX publishes a design for a 250 kW-peak orbital compute satellite and plans a factory to build thousands from late 2027.
- The FCC has accepted, not approved, a filing for up to one million orbital data centre satellites.
- No orbital compute satellite has flown yet; SpaceX targets 2027.
- The reply gives no numbers for robot memory demand; the analyst's figures were not rebutted with data.
REASONING
- The claims are about the far future and come without supporting figures, so the verdict is speculative.
- The near-term SpaceX plan is real and published, but it is low Earth orbit, not deep space.
- If it works, solar-powered orbital compute adds harnessed power without touching Earth's grid, which would speed the climb.
SEE ON THIS SITE
TAGS
- COMPUTE
- SPACE
- ORBITAL DATA CENTRES
- ROBOTS
- MEMORY