The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model | #266
Planet CEO Will Marshall joins the Moonshots crew to unpack 'large earth models' (LLMs fused with Planet's 150-petabyte, 10-year satellite imagery archive) and Project Suncatcher, Google's early effort to put TPU compute clusters in orbit. The panel debates whether launch cost (SpaceX) or compute efficiency (Nvidia/Google TPUs) will ultimately govern the economics of space-based AI, and digs into orbital debris, satellite miniaturization, and the compute-vs-inference balance for space data centers. The back half covers Eric Schmidt's Relativity Space pivot toward a Mars mission, a wave of AI-talent defections (Noam Shazeer to OpenAI, John Jumper to Anthropic), Javier Milei's proposal for AI corporate personhood in Argentina and Yuval Noah Harari's rebuttal, China's GLM 5.2 open-weight model rivaling Western frontier models, and new financial infrastructure (Orin's OPTI index) for pricing AI compute like a commodity.
Large earth models: fusing LLMs with satellite imageryAI▶ 4:31
Will Marshall explains 'planetary intelligence' — combining Planet's continuous Earth imagery with LLMs so users can ask natural-language questions about the physical world (crop yields, floods, construction), not just text-trained theory.
Planet's satellite fleet and resolution roadmapSpace▶ 12:00
Planet runs three fleets: a daily-scan fleet (3m resolution moving to 1m), a high-resolution fleet (40cm moving to 30cm), and a 400-band hyperspectral imager. Government defense satellites have far higher resolution but cover under 1% of what Planet covers daily.
Historical imagery archive and commercial use casesEconomy▶ 8:00
Planet's unique differentiator is a 10-year historical archive (3,000 images per point on Earth) letting customers compare current conditions to historical norms — used by Ukraine defense, US intelligence, farmers, hedge funds, and journalists.
Government revenue mix and export-control blacklistsGeopolitics▶ 29:01
Planet's revenue is roughly 60% defense/intelligence, 25% civil government, 15% commercial. As a US remote-sensing company it follows a NOAA-mandated customer blacklist (Iran, North Korea, terrorist groups) that roughly mirrors the EU's own list.
Edge computing on satellites (Nvidia GPUs in orbit)Compute▶ 34:27
Planet now flies Nvidia GPUs on its Pelican satellites to process imagery in orbit and relay results satellite-to-satellite, cutting response time from hours to seconds/minutes — demonstrated with real-time aircraft detection and LA wildfire response.
Project Suncatcher: Google's orbital TPU computeCompute▶ 49:20
Planet is building and testing Google's first orbital TPU compute satellites (radiation management, cooling, inter-satellite optical links) as an early step toward moving energy-intensive compute off Earth, an idea Google and Planet studied together roughly eight to nine years ago.
Orbital debris, Kessler risk, and satellite lifespanSpace▶ 56:35
Marshall argues satellite mega-constellations aren't the main debris risk (debris outnumbers satellites ~10,000 to 1); Planet flies at 400-500km specifically so dead hardware deorbits within months to a few years.
Launch tax vs compute tax: what really governs space economicsCompute▶ 1:17:59
Marshall's framing: everyone except SpaceX pays a 'SpaceX launch tax' and everyone except Nvidia/Google pays an 'Nvidia tax' — near-term, launch cost dominates; long-term, compute efficiency (flops per watt) determines who wins orbital AI compute.
Relativity Space and the rocket-launch cost raceSpace▶ 1:06:41
Eric Schmidt, an early Planet investor, has taken over as CEO of Relativity Space (pivoting from 3D-printed rockets to the heavy-lift Terran R) and just landed a NASA Mars orbiter mission (ELIS); the panel debates alternative launch paradigms (mass-manufactured rockets, space elevators, lunar rail launch) beyond SpaceX-style reusability.
AI talent brain drain: Google loses Shazeer and JumperAI▶ 1:25:37
Noam Shazeer (Transformer co-author) leaves Google for OpenAI a second time, and AlphaFold co-creator John Jumper leaves Google DeepMind for Anthropic; the panel debates whether this reflects Google falling behind the frontier or simply smaller-team agency advantages.
AI corporate personhood: Milei vs HarariGeopolitics▶ 1:44:26
Argentina's Javier Milei proposes a new legal category letting AI-run corporations incorporate, sign contracts, and be sued without humans in the loop; Yuval Noah Harari publicly rebuts it, warning it lets humans hide behind a non-human liability shield.
GLM 5.2: China's open-weight model shocks the frontierAI▶ 1:59:17
Zhipu AI's (Z.AI) GLM 5.2, a 753B-parameter open-weight MoE model with a 1M-token context window, matches or beats top Western models on reasoning-heavy benchmarks, reigniting debate over whether China's open-weight models are still 6-8 months behind and whether frontier AI can be contained at all.
Pricing intelligence: Orin's OPTI index and AI capex vs cash flowEconomy▶ 2:15:01
Orin launches OPTI, a public index tracking OpenAI/Anthropic per-token inference prices (like an oil price index), while Epoch AI data shows hyperscalers' AI capex now outpacing their cash flow, funded by debt and equity rather than revenue.
EP #? · · due: within 10 years of the statement (~mid-2030s) · ▶ watch
“within 10 years we expect most compute to be put into space”
Your call:
openAlex Wissner-Gross: Sun-synchronous-orbit satellite swarms (an early Dyson-swarm-style ring) will become faintly visible from Earth around dawn/dusk.
EP #? · · due: 18 months from the post (~late 2027) · ▶ watch
“There will be an open-source fable level model that runs on a base Mac mini or equivalent. Uh he gave it 18 months.”
Your call:
openDave Blundin: Anthropic's most advanced model ('Fable 5') will return to availability, but as a reduced/limited version of what was originally released.
EP #? · · due: unspecified (Polymarket odds tracked through end of the current month) · ▶ watch
“Fable 5 will come back, but it'll be a reduced version of what it was the first time it was out.”
Your call:
openAlex Wissner-Gross: Humans will begin being uploaded (mind uploading becomes real).
“We're going to get uploaded humans sometime, I think in the next 10 years.”
Your call:
Numbers that matter
Planet is a $10 billion companyPeter Diamandis's opening valuation of Planet (ticker PL).
450% stock price increase over the past yearPlanet's (PL) share price growth cited by Dave Blundin.
200 satellites in Earth orbitPlanet's active satellite fleet size.
25 terabytes of imagery generated per dayPlanet's daily data output (later cited elsewhere in the episode as 40 terabytes/day).
3,000 images per point on Earth's land mass over 10 yearsSize of Planet's historical imagery archive per location.
~150 petabytes of total imagery dataTotal size of Planet's historical data stack.
Scanning fleet resolution improving from 3m to 1m; high-res fleet from 40-50cm to 30cmPlanet's next-generation resolution roadmap (Owl and Pelican satellites).
Hyperspectral imager captures 400 spectral bands vs. the human eye's 3 (RGB)Planet's Tanager hyperspectral satellite, built with JPL.
~half a dozen classified US defense satellites, 10x+ higher resolution but covering under 1% of Planet's daily coverageComparison of government spy satellites to Planet's commercial fleet.
Planet images ~200 million sq km/day vs. Earth's total land mass of ~150 million sq kmScale of Planet's daily global coverage.
Satellite radio downlink speed grew from 1 megabit/s to 10 gigabits/s over 13 yearsTechnology improvement since Planet's first Dove satellite launched in 2013.
Expect roughly 100x more usable data/capability unlocked in the next couple of yearsWill Marshall's estimate of AI unlocking latent value in existing satellite imagery.
Space-based compute becomes cheaper than ground-based once launch costs hit $200-300/kgFindings from Planet and Google's joint feasibility study (~8-9 years prior).
Google spends roughly $200 billion/year on compute, roughly the size of the entire space industry todayScale comparison motivating Project Suncatcher.
Worth digging into
🕳️ Project Suncatcher's actual tech-demo results
Whether Google's TPU efficiency really beats SpaceX-style GPU compute in orbit could decide the long-term winner of space-based AI compute, per Will Marshall's 'launch tax vs compute tax' framing.
🕳️ GLM 5.2 independent benchmark verification
The claim that a Chinese open-weight model now rivals Opus 4.8/GPT-5.5 on reasoning benchmarks challenges the assumption of a persistent 6-8 month Western lead.
🕳️ Planet + Google DeepMind's AlphaEarth embedding project
This 'tokenizing the Earth' embedding-space work is the technical bridge toward Alex Wissner-Gross's proposed predictive 'crystal ball' Earth video model.
🕳️ Javier Milei's AI corporate-personhood proposal
This is a live, real-world legal experiment in granting non-human entities corporate rights — a template other jurisdictions may follow.
🕳️ Orin's OPTI index and NYSE early-stage ticker program
A public commodity-style index for AI inference token prices, plus a novel NYSE ticker program for early-stage startups, could reshape how AI compute capex gets financed and hedged.
🕳️ Relativity Space under Eric Schmidt
Schmidt was an early Planet investor before AI-in-space was an obvious thesis; his pivot of Relativity from 3D-printed rockets to heavy-lift and a NASA Mars mission (ELIS) suggests a long-held conviction about orbital compute.