# Artificial-Intelligence Ambitions Face Physical Constraints in Orbit and on the Ground

SpaceX proposes launching millions of satellites to host computing, while two researchers argue models should learn physics rather than language.

- Published: 2026-08-26T05:17:19.611Z
- Canonical: https://polylog.news/2026-08-26/artificial-intelligence-ambitions-face-physical-constraints
- Publisher: Polylog (Global desk)
- Section: tech
- Sources: [Financial Times](https://www.ft.com/content/41e60b41-5f13-4d72-8dfc-9bab83f57fb3?syn-25a6b1a6=1), [The Japan Times](https://www.japantimes.co.jp/business/2026/08/26/tech/ai-founders-bezos-prometheus-universe/)

Two stories from opposite ends of the artificial-intelligence build-out describe the same constraint. SpaceX intends to place data centres in orbit to supply computing power, a plan the [Financial Times reports](https://www.ft.com/content/41e60b41-5f13-4d72-8dfc-9bab83f57fb3?syn-25a6b1a6=1) faces formidable obstacles in launching millions of satellites, with one industry figure quoted in the piece calling the idea apparently crazy.

The proposal exists because ground-based data centre capacity is reaching its limits. Data centres compete for electricity, cooling water, land and grid connections, and those inputs cannot be scaled at the pace of chip production. Orbit removes the land constraint and the cooling problem in exchange for launch cost, radiation and servicing difficulties that have no precedent at that scale.

At the other end, Anima Anandkumar and Benedikt Jenik, researchers once connected to a project backed by Jeff Bezos, have built a model that predicts physical phenomena across space and time rather than generating text, [The Japan Times reported](https://www.japantimes.co.jp/business/2026/08/26/tech/ai-founders-bezos-prometheus-universe/). Their argument is that modelling the physical world requires a different architecture from modelling language, and by implication a different cost structure.

Both efforts respond to the same economics. The current approach consumes capital and energy at a rate that assumes continued cheap financing, and the search for alternatives shows that assumption is now in question.

## What this means

The binding constraint on artificial-intelligence capacity is shifting from chip supply to power, land and cooling, which redirects capital toward electricity generation, transmission equipment and, in the more speculative case, launch services. Utilities and grid-equipment manufacturers gain from that shift. Investors in companies whose valuations assume compute grows without limit face the opposite exposure, because physical bottlenecks show up as delayed revenue rather than as cancelled orders.

## What to watch

- Whether any hyperscale operator commits capital to orbital computing rather than studying it, the line between a research proposal and an industry direction.
- Grid connection queues and power purchase agreements for new data centres in the United States and Europe, the most direct measure of how tight the physical constraint has become.
- Whether physics-based models attract commercial customers in weather, energy or industrial simulation, which would show a second architecture is economically viable alongside language models.