The $10B satellite empire putting AI in orbit, why chips beat rockets, and China’s leading open model

Planet is trying to turn Earth observation into an intelligence layer: daily imagery, a decade of historical archives, and models that let users ask natural-language questions about what is changing on the ground. Will Marshall frames the ambition as a search engine for the planet: not just seeing Earth, but querying it.

Large Earth models

Planet’s moat is not only its current satellite fleet. It is the historical archive. For defense, agriculture, industry, insurance, or government users, today’s image matters because it can be compared with what is normal: past activity around a site, crop changes, data-center construction, or infrastructure movement.

That pushes geospatial data toward a knowledge system rather than an image library. Embeddings can compress massive imagery layers into searchable spaces where objects, anomalies, and long-term patterns become easier to detect and compare.

Why orbit becomes AI infrastructure

Marshall argues that the recent space revolution is not only about lower launch costs. Satellites themselves have improved dramatically in sensors, radios, storage, resolution, and data per kilogram.

That progress makes orbital compute more plausible. Terrestrial data centers consume land, water, and power, and are becoming politically contentious. In orbit, solar energy is abundant; the practical task is to send up questions and return answers. Inference looks like the first realistic workload, with large-scale training likely staying on Earth longer.

The AI race widens

The conversation also turns to Anthropic’s recruiting pull, recursive self-improvement, and the governance questions around increasingly autonomous systems. The speakers argue that technical acceleration needs matching accountability and safety mechanisms.

GLM 5.2 is another strategic signal. A Chinese open-weight model approaching frontier performance in coding, long-horizon agency, and design changes the access question: who controls intelligence, and what happens when a local, open alternative is good enough?

Takeaways

• Planet wants Earth to be searchable, not merely observable.
• Historical imagery is a strategic data moat.
• Orbital compute responds to terrestrial data-center constraints.
• Inference may move to space before training.
• GLM 5.2 suggests Chinese open models can narrow the frontier gap.

Source

• Chaîne: Peter H. Diamandis
• Vidéo source: https://www.youtube.com/watch?v=kPSLLeccrik