Why now?
Change one assumption and calculate the budget again. Do not present the optimistic scenario as an achieved price.
Links containing ?t= open the video at a specific second.
The ideas to retain
1. The explosion in compute demand
Generative AI has been the most visible catalyst for accelerating compute demand, but it is not the only one. Training large models, inference at scale for hundreds of millions of users,…
2. Terrestrial bottlenecks
Building more data centers on Earth is not infinitely scalable. Six main problems slow the expansion of terrestrial infrastructure, and unlike analyses from five years ago, there are now…
Power grid
Large-scale data centers consume power at rates that local grids cannot always absorb. Google reported in January 2026 that grid-connection lead times had become its biggest operational…
Jump directly to a section
Read the reviewed transcript
Transcript of the visual text. This video has no narration.
00:00 — More requests. The same infrastructure.
Training and inference require compute and physical capacity.
When jobs arrive faster than they are served, the queue grows.
Adding chips is not enough when the power supply remains limited.
Illustrative example; assumptions, not operational measurements.
00:22 — Six conditions. All of them matter.
Power, water, land, permits, heat and latency constrain projects differently.
A completed building cannot resolve a pending grid connection by itself.
Check each condition; do not add resources with different units.
Illustrative example; assumptions, not operational measurements.
00:44 — A lower launch cost does not erase the rest.
Launch cost depends on mass and price per kilogram.
Lower that price in the example: only part of the budget changes.
Panels, radiators, links and operations still have a cost.
Illustrative example; assumptions, not operational measurements.
01:08 — You also launch what supports the chip.
The processor is only part of the mass sent into orbit.
Add power generation, thermal control, structure and communications.
Evaluate the mass of the whole system, not just the accelerator.
Illustrative example; assumptions, not operational measurements.
01:30 — Where the data originates matters.
When data originates in orbit, processing there may reduce downlink traffic.
When it originates on Earth, moving the job adds an uplink and a downlink.
Compare both paths for the same user need.
01:52 — A projection is not an observed result.
An economic forecast depends on launch, mass and lifetime assumptions.
Change one assumption and calculate the budget again.
Do not present the optimistic scenario as an achieved price.
Illustrative example; assumptions, not operational measurements.
