"The $10B Satellite Empire Putting AI in Orbit..." — Peter Diamandis (Moonshots)

Why this is in the vault

Planet Labs' "large earth models" thesis reframes satellite imagery from a niche government product into a core AI primitive — the same pattern (vast proprietary dataset + AI unlock = new market) that drives RDCO's data infrastructure advisory work. The episode's treatment of orbital compute (Project Suncatcher), the SpaceX/Nvidia tax framing, and China's GLM 5.2 open-weight model are all conversation-ready in client settings where AI infrastructure choices and geopolitical supply chain risk come up. For Ray as a Deal Solutions Architect at phData, the semiconductor/chip constraint angle maps directly to the capital-cycle thesis being tracked in the Markov investing pipeline.

Episode summary

Will Marshall, co-founder and CEO of Planet Labs, joins Peter Diamandis to explain how Planet's 200-satellite constellation — imaging the entire Earth daily for 10 years — has accumulated a 150-petabyte archive that AI is only now making commercially accessible at scale. The conversation covers the concept of "large earth models" as a physical-world counterpart to LLMs, on-orbit AI processing (Nvidia GPUs on Pelican satellites), the competitive dynamics of the SpaceX launch tax versus the Nvidia compute tax, and the geopolitical stakes of making Earth's surface fully legible in near-real-time. The episode then pivots to China's GLM 5.2 open-weight model — described as matching or exceeding top US closed models — before zooming out to the orbital compute thesis (Project Suncatcher) as the next infrastructure bet after the current ground-based GPU buildout.

Key arguments / segments

• [00:00] Cold open — three themes stacked — Pull quotes establish the episode's three pillars: large earth models, the SpaceX-vs-Nvidia tax framing, and China's GLM 5.2. Planet's stock is up ~450% in the past year; Will Marshall is framed as building "humanity's orbital AI and data layer."

• [00:03] Large earth models defined — LLMs know all the text of the internet but have never "looked out the window." Planet's large earth models add real-world physical sensing as a layer — satellites answer questions about the actual current state of the planet (crop yields, military movements, disaster extent) in ways no text corpus can.

• [00:06] The archive moat — 150 petabytes, 3,000 images per point on Earth's land mass across 10 years. A competitor launching satellites today cannot go back in time; the historical archive is structurally irreproducible. Revenue is roughly 60% defense/intelligence, 25% civil government, 15% commercial — with commercial accelerating as AI lowers the access barrier from Python-fluent analysts to natural language queries.

• [00:10] Three fleet architecture and Tanager hyperspectral — Dove/Owl fleet at 1–3m resolution with sub-hour latency; Pelican high-res fleet targeting 30cm resolution / 30x daily / 30-minute tasking window; Tanager hyperspectral with 400 spectral bands (built with JPL) enabling tree-species identification, methane plume detection, and vehicle paint signature classification.

• [00:21] Tokenizing the Earth + geopolitical transparency — Work with Google DeepMind's "Alpha Earth" and open CLIP models converts planetary imagery into embedding spaces — effectively a compression step before predictive modeling. Marshall argues full Earth transparency deters war by eliminating the information asymmetry that historically precedes conflicts; the US deliberately chose satellite overflight rights after the Gary Powers U2 incident.

• [00:34] On-orbit AI — April 2025 Alice Springs experiment — Nvidia GPUs installed on Pelican satellites; in a live demo, the satellite autonomously identified aircraft type and location over Alice Springs and returned structured data in seconds via satellite-to-satellite RF links. Moore's Law equivalent for Planet hardware: 5–10x improvement every 2–3 years. Radio throughput went from 1 Mbps (2013) to 10 Gbps today.

• [00:56] SpaceX launch tax vs. Nvidia compute tax — Marshall's key competitive framing: everyone except SpaceX pays the launch tax; everyone except Nvidia and Google pays the compute tax. Near-term, launch cost dominates; long-term, compute is the binding constraint. This reframes the orbital compute race as fundamentally a semiconductor infrastructure question, not a rocket question.

• [01:09] China's GLM 5.2 and the open-weight shock — GLM 5.2 from Zhipu AI matches or exceeds top US closed models from OpenAI and Anthropic in certain benchmarks, released as an open-weight model. Framed as "absolutely shocking" — China has cracked how to burn tokens for intelligence gains at lower cost, and open-weighting it collapses the US labs' moat.

• [01:21] Project Suncatcher — orbital compute thesis — The episode's second major thesis: as GPU demand outstrips Earth-based data center capacity, moving AI compute to orbit makes physical sense (infinite solar power, radiative cooling, lower latency to other satellites). Project Suncatcher is Planet's or an affiliated venture's bet on this. Sets up a direct answer to the "how do you compete with Elon's orbital AI data centers?" tease from the cold open.

• [01:47] $100B market framing and the commercial unlock — Will estimates the retroactive-analysis market alone (answering historical questions about Earth's state) is a $100B opportunity, before counting predictive models. The commercial inflection happens when AI reduces the access barrier from specialized satellite analysts to any product team with a natural language query interface.

• [02:01] Broader AI infrastructure and the collapsing cost of intelligence — Conversation zooms out to Kling/Vidu/Wan video models, Bidance's video generation, and the pattern of Chinese AI catching up to US closed models in successive waves. The frame is that intelligence cost curves are collapsing globally, making the physical-world sensing layer (Planet's archive) more valuable, not less — because the bottleneck shifts from model quality to data quality.

• [02:14] Closing — Earth as an intelligent system — Will closes with the "index the Earth to make it searchable" frame: the same way Google indexed text to create a new economy, Planet is indexing physical reality. The endgame is Earth itself as a queryable, predictive, intelligent system — with humans as "smart stewards of our planet."

Notable claims

• Planet is currently valued at ~$10 billion; stock up ~450% in the past year (ticker: PL on NYSE)
• 200 active satellites imaging the entire Earth land mass every single day; 25 terabytes of new imagery generated daily
• 150-petabyte 10-year archive with ~3,000 images per point on Earth's land mass — structurally irreproducible by any new entrant
• US defense/intelligence community has roughly half a dozen high-res satellites covering less than 1% of Earth daily — Planet's daily global scan is something no government has achieved
• Tanager hyperspectral imager has 400 spectral bands, enabling tree-species identification, methane plume mapping, and vehicle paint signature classification by manufacture location
• On-orbit AI experiment (April 2025): Pelican satellite with Nvidia GPU identified aircraft type/location over Alice Springs autonomously in seconds via satellite-to-satellite RF
• Radio throughput on Planet satellites: 1 Mbps in 2013 → 10 Gbps today; cameras: 2MP → 47MP; storage: ~100MB → ~2TB per satellite
• GLM 5.2 (Zhipu AI, China) matches or exceeds OpenAI and Anthropic top models in certain benchmarks — released open-weight
• Estimated retroactive Earth-analysis market: $100B+ before predictive modeling is counted

Guests

• Will Marshall — Co-Founder & CEO, Planet Labs

Mapping against Ray Data Co

Medium-strong. Three direct connections:

1. Data infrastructure thesis: The "150 petabytes becomes commercially useful only when AI lowers the access barrier" pattern is exactly the value proposition RDCO articulates to clients building on large proprietary datasets. Will's framing of moving from Python-fluent analysts to natural language query interfaces is a client-ready story for any enterprise data platform conversation.

2. Semiconductor/chip capital cycle: The SpaceX-launch-tax vs. Nvidia-compute-tax frame is the most useful take-away for the Markov investing pipeline — it argues that long-term, orbital infrastructure is constrained by silicon availability, not launch economics. This sharpens the chip-fab capital cycle thesis.

3. China AI advisory: GLM 5.2 matching closed US models as an open-weight release is a development any client building on top of US closed-model APIs should know about. As a DSA at phData, Ray will have conversations about AI vendor selection; this is fresh competitive context.

The satellite observation/space vertical itself is not a current RDCO client vertical, so the mapping is relevance-by-analogy rather than direct client applicability.

Related

• [[06-reference/2024-03-14-moonshots-ep90-will-marshall-satellites]]
• [[06-reference/2025-11-04-moonshots-ep205-openai-ipo-china-sp500-jobs]]
• [[06-reference/2025-01-29-moonshots-ep146-deepseek-openai-emad-mostaque]]
• [[06-reference/2026-03-26-moonshots-ep242-elon-chip-race-sp500-human-drivers]]
• [[06-reference/2026-04-15-dwarkesh-jensen-fires-back-china-chip-ban]]
• [[01-projects/investing/2026-05-27-markov-equities-pipeline-spec]]
• [[06-reference/transcripts/2026-06-26-moonshots-satellite-empire-ai-orbit-266-transcript]]