Planet stock up 450% in 1 year
Planet Labs' stock ticker PL increased approximately 450% over the last year.
The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model | EP #266
Planet Labs CEO Will Marshall says orbital TPU farms will be cheaper than terrestrial data centers within a decade — and the real bottleneck isn't launch cost, it's the NVIDIA compute tax.
Jun 26, 2026
2:26:56
Difficulty: Intermediate
Played
Moonshots with Peter Diamandis
The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model | EP #266
Planet Labs CEO Will Marshall says orbital TPU farms will be cheaper than terrestrial data centers within a decade — and the real bottleneck isn't launch cost, it's the NVIDIA compute tax.
Jun 26, 2026
2:26:56
Difficulty: Intermediate
Played
TL;DR
Will Marshall, CEO of Planet Labs, joins Peter Diamandis and the Moonshot Mates to unpack how 200 satellites imaging the entire Earth daily are becoming the backbone of "large Earth models" — AI trained on 150 petabytes of real-world data rather than just internet text [1] — Will Marshall "LLMs have read every book but never looked out the window. Planet Labs is building large Earth models that add real-world sensing to AI — 1…" 04:39 . Project Suncatcher is partnering with Google to test TPUs in orbit, where compute may ultimately be cheaper than on the ground [2] — Will Marshall "Planet strapped NVIDIA GPUs to its Pelican satellites and demonstrated real-time edge processing: photograph an Australian airfield, automa…" 34:45 . The crew also debates China's GLM 5.2 open-weight model matching Western frontier labs [3] — Peter Diamandis "Argentina's president Javier Milei is proposing AI-native corporations that can sign contracts, hire people, and sue — with no humans in th…" 1:28:45 , the AI brain drain from Google DeepMind to Anthropic, Argentina's AI personhood proposal, and why the SpaceX launch tax matters less long-term than the NVIDIA compute tax [4] — Will Marshall "SpaceX launch tax vs NVIDIA tax: Near-term the SpaceX launch tax is the bigger barrier for orbital compute, but long-term the NVIDIA comput…" 1:17:42 .
#orbital compute
#large Earth models
#satellite imagery AI
#open-weight AI models
#China AI frontier
#AI compute economics
#space debris Kessler syndrome
#AI personhood legal
#model distillation
#inference efficiency
#Google TPUs
#NVIDIA tax
#Project Suncatcher
#AI safety governance
#Planet Labs
#GLM 5.2
#AI safety
#space debris
#Kessler syndrome
#distillation
#AI personhood
#Relativity Space
#Eric Schmidt
#open-weight AI
#planetary intelligence
#Fermi paradox
#satellite imagery
#AI brain drain
Collision of Earth intelligence, orbital compute, Chinese open-weight AI, and new space infrastructure. Large Earth models and Project Suncatcher are turning satellite data and orbital compute into core AI primitives.
Chapter list
-
Rapid-fire highlights from the episode: large Earth models, the NVIDIA vs SpaceX tax debate, and China's GLM 5.2 shocking the AI world.
-
Peter introduces Will Marshall (CEO of Planet Labs), Alex Wissner-Gross, Dave Blundin, and Salim Ismail. Planet's ticker PL is flagged as a public company.
-
Peter previews all stories — Planet, Eric Schmidt's Relativity Space, AI brain drain, GLM 5.2, and collapsing inference prices — then gives Will's credentials and Planet's stats.
-
Will explains large Earth models — combining satellite imagery with LLMs so AI can answer real-world physical questions. 150 petabytes, 3,000 images per Earth point over 10 years. [1] — Will Marshall "LLMs have read every book but never looked out the window. Planet Labs is building large Earth models that add real-world sensing to AI — 1…" 04:39
-
Will details Planet's 10-year time-series archive and use cases: Ukraine military tracking, agriculture yield prediction, building permit enforcement, and journalism.
-
Will describes three fleets: 3-meter Dove (upgrading to 1m Owl), 30cm Pelican (30x30x30), and Tanager with 400 spectral bands — capable of identifying tree species and military vehicle origins. [1] — Will Marshall "400 spectral bands, hyperspectral: Planet's hyperspectral satellite (Tanager), built with JPL, captures 400 spectral bands spanning infrare…" 13:53
-
Peter promotes the free Moonshots weekly Substack summary at diamandis.com/metatrends.
-
Discussion of how Planet started from a PhoneSat, why private sector hasn't replicated its coverage, and how AI is now lowering barriers to value extraction from satellite data.
-
Will describes how Planet trained a model on US data centers and aimed it at China, predicting completion dates within days — the first working example of autoregressive Earth forecasting. [1] — Will Marshall "Planet loaded every registered US data center into an AI model, showed it their construction histories via satellite imagery, then aimed it…" 19:00
-
Alex pushes Will on building a full pixel-level autoregressive Earth video model. Will explains embedding spaces (AlphaEarth, CLIP) as the compression step needed to make this tractable.
-
Salim raises sovereignty tensions — governments domesticated the internet; Planet is re-globalizing maps. Will explains NOAA registration, customer blacklists, and the physics of orbital overflight rights.
-
60% defense/intel, 25% civil government, 15% commercial. Discussion of MCP API pricing for AI training data and how AI is opening the commercial market.
-
Will explains the April 2026 Alice Springs experiment — on-orbit object recognition returning results in seconds — and how faster imagery could have saved lives in the LA Palisades wildfires. [1] — Will Marshall "Planet strapped NVIDIA GPUs to its Pelican satellites and demonstrated real-time edge processing: photograph an Australian airfield, automa…" 34:45
-
Will traces 10x generational improvements: radio, camera resolution, storage. Moore's law for space data is roughly 5–10x every 2–3 years; AI adds another ~100x unlock.
-
Will explains why Planet keeps satellites at 400–500 km where drag deorbits them in months to years, making the shell self-cleaning and avoiding Kessler cascade. [1] — Will Marshall "For every satellite in orbit today, there are 10,000 pieces of debris. Kessler syndrome — the collisional cascade — is already in effect at…" 57:10
-
Will reveals the Google-Planet study: at $200–$300/kg launch cost, orbital compute beats terrestrial on pure cost. Google expects most compute in space within 10 years; Planet is building their first test satellites. [1] — Will Marshall "A Google-Planet study found that once launch costs hit $200–$300 per kilogram, it is simply cheaper to run compute in orbit than on the gro…" 49:20 [2] — Will Marshall "Google: most compute to space in 10 years: Google's CEO Sundar Pichai told Planet Labs that within 10 years, most compute is expected to be…" 54:10
-
Alex's calculation: 40%+ YoY upmass growth naively disassembles Earth by 2144. Will discusses visibility of dawn/dusk SSO constellations and Elon's FCC approval for 1M satellites.
-
Peter asks how Planet competes with SpaceX's orbital AI plans. Will coins the NVIDIA tax vs SpaceX launch tax framing: long-term chip efficiency, not launch cost, determines the winner. [1] — Will Marshall "Everyone except SpaceX pays a launch tax. Everyone except NVIDIA and Google pays a compute tax. Will Marshall's contrarian take: near-term,…" 1:17:20
-
Discussion of Eric Schmidt acquiring Relativity Space, the 3D-printing approach, and Will's argument that chemical rockets are a paradigm ripe for disruption via SpinLaunch, fission, or space elevators.
-
Sponsored segment: Fountain Life's CMO discusses dementia prevention — 45% of cases preventable, 26% brain age improvement from healthy living interventions.
-
Peter covers two major AI talent moves. Alex argues Google has fallen behind the frontier; Will defends Google's compute, data, and talent advantages.
-
Dave and Peter argue researchers are flocking to Anthropic not just for comp but to be present at the singularity. Anthropic reportedly shows applicants raw pre-trained model access. Will defends Google on compute and data. [1] — Peter Diamandis "The co-author of the Transformer paper left Google for OpenAI. The Nobel laureate behind AlphaFold left Google DeepMind for Anthropic. Thes…" 1:18:36
-
Will argues LLMs are brains in a vat — without real-world physical sensing, AI can't truly learn or align with human interests. Planetary intelligence is the path to AI embodiment at scale. [1] — Will Marshall "When a baby is born, they learn and become intelligent and ultimately self-aware and conscious by interacting with the physical world. They…" 1:34:10
-
Will's impassioned closing on Planet Labs' mission: Earth wins over Mars by orders of magnitude; life is either singular or extraordinarily rare; this biosphere deserves everything. [1] — Will Marshall "Planet Labs has imaged thousands of exoplanets. The verdict: Earth wins by several orders of magnitude. No place on Mars is better than the…" 1:22:20
-
Peter outlines Milei's three AI proclamations and Harari's counter. Alex backs AI personhood; Will calls for a Manhattan-Project-scale safety investment; Salim separates legal, moral, and AI personhood. [1] — Peter Diamandis "Argentina's president Javier Milei is proposing AI-native corporations that can sign contracts, hire people, and sue — with no humans in th…" 1:28:45
-
Salim lists machine-native enforcement mechanisms: compute revocation, asset seizure, model credential suspension, forced deletion, loss of legal identity. Discussion of AI personhood as spectrum, not binary.
-
Will delivers his sharpest warning: we spend 100x more on AI than the Manhattan Project but 100x less on safety than was spent on nuclear safety — a 10,000x deficit. Calls for an AI conclave. [1] — Will Marshall "We're spending 100 times more on AI today than the Manhattan Project cost — and 100 times less on AI safety than we spent on nuclear safety…" 1:37:00
-
Sponsored segment for Blitzy — autonomous software development platform delivering 5x engineering velocity for enterprises.
-
Peter introduces China's GLM 5.2 (753B params, open-weight). Alex explains it matches Opus 4.8 at half the cost using 2x tokens; Salim declares frontier intelligence can no longer be monopolized. [1] — Peter Diamandis "China's GLM 5.2 is 753 billion parameters, open-weight, and in some benchmarks matches or beats Anthropic's Opus 4.8. It does this using ro…" 1:59:37 [2] — Alexander Wissner-Gross "GLM 5.2: 2x tokens, half the price: GLM 5.2 uses roughly double the tokens to reach the same capability as top Western models but at approx…" 2:04:50
-
Alex explains distillation; Will warns that open-source fork-able models enable removal of safety guardrails for bioweapons; Dave notes the narrowing window for US export controls. [1] — Alexander Wissner-Gross "Distillation is how a smaller, cheaper model learns from a larger, more expensive one — like a student absorbing a teacher's knowledge. Chi…" 2:06:40
-
Will connects AI safety to the Fermi Paradox and Great Filter — civilizations may destroy themselves with technology faster than they build wisdom. Earth is galactically significant. [1] — Will Marshall "We came very close with nukes a number of times, and with AI, we're just about to build something that's far, far more risky for our specie…" 1:52:20
-
Dave and Alex introduce Orin's OCPI — the first public index tracking OpenAI and Anthropic inference token prices, now live on Bloomberg with NYSE ticker ORNN. Compute is the new oil. [1] — Alexander Wissner-Gross "A startup called Orin has built the first public index tracking what OpenAI and Anthropic actually charge per inference token — now live on…" 2:15:00
-
Discussion of Epic AI data showing hyperscalers (Microsoft, Google, Amazon, Meta) spending more on AI than cash flow. Dave argues this is normal financing like a mortgage; Will warns long-term sustainability.
-
Peter thanks Will, Salim closes with 'extract the promise without the peril', Will calls for planetary wisdom. Outro rap and final subscription/newsletter call to action.
- Large Earth Model (LEM)
- An AI model trained on satellite imagery and real-world physical data — analogous to a large language model but for the physical state of the Earth rather than text on the internet.
- Project Suncatcher
- Planet Labs' initiative, in partnership with Google, to deploy AI compute hardware (TPUs) in orbit, where solar power is abundant and cooling is achievable, potentially making orbital compute cheaper than terrestrial data centers.
- Sun-synchronous orbit (SSO)
- A near-polar Earth orbit in which a satellite passes over any given point at the same local solar time each day, providing consistent lighting for optical imaging; Planet Labs' preferred orbit.
- Kessler syndrome
- A theoretical chain reaction where a collision in orbit generates debris that causes further collisions, cascading into a debris cloud that renders certain orbital altitudes unusable.
- Hyperspectral imaging
- Remote sensing that captures data across hundreds of narrow wavelength bands (vs. the human eye's three), enabling identification of specific materials, vegetation species, or chemical signatures from orbit.
- Distillation (ML)
- A machine learning technique where a large, expensive 'teacher' model generates outputs used as training data for a smaller, cheaper 'student' model, compressing the teacher's capabilities into a more efficient form.
- Open-weight model
- An AI model whose weights (trained parameters) are publicly released, allowing anyone to download, run, and modify it, as opposed to a closed API model accessible only through a vendor.
- Inference
- The process of running a trained AI model to generate outputs from new inputs; as opposed to training. Approximately 70% of AI compute today is inference, and it is the primary use-case targeted for orbital data centers.
- TPU (Tensor Processing Unit)
- Google's custom AI accelerator chip, designed to be more energy-efficient than general-purpose GPUs for AI workloads, particularly inference — a key factor in the orbital compute economics debate.
- Mixture of Experts (MoE)
- An AI model architecture where different subnetworks ('experts') specialise in different inputs, and only a subset is activated for each query — enabling very large parameter counts while keeping inference cost manageable.
- Fermi Paradox
- The apparent contradiction between the high probability of extraterrestrial civilizations existing and the complete absence of evidence for them; discussed in the episode in the context of AI existential risk and the Great Filter.
- Great Filter
- A hypothesis in the context of the Fermi Paradox that some catastrophic barrier prevents civilizations from surviving long enough to become interstellar — potentially self-inflicted through advanced technology like AI or nuclear weapons.
- Upmass
- The total mass of payloads launched to orbit in a given period; used as a metric for the growth of the space economy.
- Autoregressive model
- A model that generates outputs one token (or pixel) at a time, each conditioned on all previous outputs — the architecture underlying most modern LLMs and video prediction models.
- Embedding space
- A mathematical representation where complex objects (images, text) are converted into dense numerical vectors, enabling semantic search and comparison across massive datasets like Planet's satellite archive.
- Iterated amplification and distillation (ITAD)
- An AI training paradigm where large frontier models iteratively teach smaller models, which in turn inform newer, larger ones — creating a self-reinforcing cycle of capability improvement.
- Cislunar
- The region of space between Earth and the Moon, including lunar orbit; relevant to discussions of near-term space infrastructure expansion.
- Radiative cooling
- Dissipating heat by emitting infrared radiation into space — the only viable cooling mechanism for electronics in the vacuum of orbit, since convection and conduction require air or liquid contact.
Chapter 3 · 03:20
Episode Roadmap & Will Marshall's Background
Peter previews all stories — Planet, Eric Schmidt's Relativity Space, AI brain drain, GLM 5.2, and collapsing inference prices — then gives Will's credentials and Planet's stats.
200 satellites, 25 TB/day
Planet Labs operates 200 satellites generating 25 terabytes of Earth imagery every single day.
Large Earth Models: Indexing the Planet Like Google Indexed the Web
LLMs have read every book but never looked out the window. Planet Labs is building large Earth models that add real-world sensing to AI — 150 petabytes of daily satellite imagery that lets you ask not just what the world knows, but what the world is doing right now.
Chapter 4 · 04:40
Large Earth Models: Indexing the Planet Like Google Indexed the Web
Will explains large Earth models — combining satellite imagery with LLMs so AI can answer real-world physical questions. 150 petabytes, 3,000 images per Earth point over 10 years. [1] — Will Marshall "LLMs have read every book but never looked out the window. Planet Labs is building large Earth models that add real-world sensing to AI — 1…" 04:39
3,000 images per land point
Planet has captured 3,000 images for every point on Earth's landmass over the last 10 years, totalling 150 petabytes.
Chapter 5 · 09:20
Historical Archive, Ukraine Intelligence, and Farming Use Cases
Will details Planet's 10-year time-series archive and use cases: Ukraine military tracking, agriculture yield prediction, building permit enforcement, and journalism.
30x30x30 Pelican satellite spec
Planet's next-generation Pelican high-resolution satellites will deliver 30-centimeter imagery, revisit any location 30 times per day, and return an image within 30 minutes of request.
400 spectral bands, hyperspectral
Planet's hyperspectral satellite (Tanager), built with JPL, captures 400 spectral bands spanning infrared to ultraviolet — enough to identify tree species or tank paint origin.
Chapter 6 · 14:30
Satellite Fleet Technical Deep Dive: Resolution, Spectral Bands & Hyperspectral
Will describes three fleets: 3-meter Dove (upgrading to 1m Owl), 30cm Pelican (30x30x30), and Tanager with 400 spectral bands — capable of identifying tree species and military vehicle origins. [1] — Will Marshall "400 spectral bands, hyperspectral: Planet's hyperspectral satellite (Tanager), built with JPL, captures 400 spectral bands spanning infrare…" 13:53
Chapter 8 · 18:25
Why No Private Competitor Has Matched Planet's Fleet
Discussion of how Planet started from a PhoneSat, why private sector hasn't replicated its coverage, and how AI is now lowering barriers to value extraction from satellite data.
The Crystal Ball: Predicting Data Center Completions from Space
Planet loaded every registered US data center into an AI model, showed it their construction histories via satellite imagery, then aimed it at China. The model now predicts data center completion dates within days — the first real demonstration of using Earth observation as a crystal ball.
Chapter 11 · 26:40
Geopolitics of Satellite Data: Sovereignty, Blacklists, and Ukraine
Salim raises sovereignty tensions — governments domesticated the internet; Planet is re-globalizing maps. Will explains NOAA registration, customer blacklists, and the physics of orbital overflight rights.
60% defense & intel revenue
Planet Labs earns roughly 60% of its revenue from defense and intelligence customers, 25% from civil government, and 15% commercial.
Chapter 12 · 32:20
Revenue Mix, AI Pricing, and Planet's Business Model
60% defense/intel, 25% civil government, 15% commercial. Discussion of MCP API pricing for AI training data and how AI is opening the commercial market.
Processing at the Edge: Real-Time Satellite Intelligence from Alice Springs
Planet strapped NVIDIA GPUs to its Pelican satellites and demonstrated real-time edge processing: photograph an Australian airfield, automatically identify aircraft type and location, transmit back only the answer via satellite link — all in seconds. The LA wildfires showed why this speed saves lives.
Chapter 16 · 49:20
Project Suncatcher: Orbital TPUs and the Economics of Compute in Space
Will reveals the Google-Planet study: at $200–$300/kg launch cost, orbital compute beats terrestrial on pure cost. Google expects most compute in space within 10 years; Planet is building their first test satellites. [1] — Will Marshall "A Google-Planet study found that once launch costs hit $200–$300 per kilogram, it is simply cheaper to run compute in orbit than on the gro…" 49:20 [2] — Will Marshall "Google: most compute to space in 10 years: Google's CEO Sundar Pichai told Planet Labs that within 10 years, most compute is expected to be…" 54:10
Project Suncatcher: Why Google Is Putting TPUs in Orbit
A Google-Planet study found that once launch costs hit $200–$300 per kilogram, it is simply cheaper to run compute in orbit than on the ground. Google's Sundar Pichai told Planet: within 10 years, most compute will be in space. Planet is now building Google's first orbital TPU test satellites.
$200–$300/kg launch threshold
When launch costs reach $200–$300 per kilogram, it becomes cheaper on a pure cost basis to put compute in orbit versus on the ground.
Google: most compute to space in 10 years
Google's CEO Sundar Pichai told Planet Labs that within 10 years, most compute is expected to be put into space.
Google spends ~$200B/year on compute
Google alone is spending roughly $200 billion per year on compute, roughly equivalent to the entire space industry today.
Chapter 17 · 56:00
Upmass Exponential, Orbital Visibility of Dyson Swarms, and Satellite Rings
Alex's calculation: 40%+ YoY upmass growth naively disassembles Earth by 2144. Will discusses visibility of dawn/dusk SSO constellations and Elon's FCC approval for 1M satellites.
Space Debris, Kessler Syndrome, and Planet's Self-Cleaning Orbit Strategy
For every satellite in orbit today, there are 10,000 pieces of debris. Kessler syndrome — the collisional cascade — is already in effect at higher altitudes. Planet keeps all satellites at 400–500 km where atmospheric drag deorbits them within months to a few years, making the shell self-cleaning.
Chapter 18 · 59:00
Competing with SpaceX: Smarts vs Mass, TPUs vs GPUs
Peter asks how Planet competes with SpaceX's orbital AI plans. Will coins the NVIDIA tax vs SpaceX launch tax framing: long-term chip efficiency, not launch cost, determines the winner. [1] — Will Marshall "Everyone except SpaceX pays a launch tax. Everyone except NVIDIA and Google pays a compute tax. Will Marshall's contrarian take: near-term,…" 1:17:20
Satellite lifespan at 400–500 km: months to years
Satellites at 400–500 km altitude naturally deorbit due to atmospheric drag within months to a few years, making the orbit self-cleaning.
Upmass +40% YoY for 5 years
Over the past five years, the mass launched to orbit has increased by over 40% year over year.
Chapter 19 · 1:03:20
Launch Economics: Relativity Space, 3D Printing, and Novel Launch Paradigms
Discussion of Eric Schmidt acquiring Relativity Space, the 3D-printing approach, and Will's argument that chemical rockets are a paradigm ripe for disruption via SpinLaunch, fission, or space elevators.
45% of dementia cases preventable
Conservative medical estimates suggest 45% of dementia cases are entirely preventable through lifestyle interventions.
26% brain age improvement
Fountain Life members who combined advanced brain testing with healthy living interventions improved their brain age by 26%.
Chapter 21 · 1:14:15
AI Brain Drain: Noam Shazeer Leaves Google for OpenAI; John Jumper Goes to Anthropic
Peter covers two major AI talent moves. Alex argues Google has fallen behind the frontier; Will defends Google's compute, data, and talent advantages.
The NVIDIA Tax vs the SpaceX Tax
Everyone except SpaceX pays a launch tax. Everyone except NVIDIA and Google pays a compute tax. Will Marshall's contrarian take: near-term, launch cost is the bottleneck for orbital AI — but longer term, chip energy efficiency is the actual moat. Google TPUs may be the real weapon in space.
SpaceX launch tax vs NVIDIA tax
Near-term the SpaceX launch tax is the bigger barrier for orbital compute, but long-term the NVIDIA compute efficiency tax matters more.
The AI Brain Drain: Why Top Researchers Are Fleeing Google for Anthropic
The co-author of the Transformer paper left Google for OpenAI. The Nobel laureate behind AlphaFold left Google DeepMind for Anthropic. These aren't random career moves — Anthropic's recruiting pitch is reportedly: come in, we'll show you what's behind the firewall, and you'll feel like you've seen God.
Noam Shazeer: $2B+ Google acqui-hire
Google acquired Noam Shazeer's company Character.AI for $2.7 billion to bring him back as head of Gemini, but he has since left for OpenAI.
Chapter 22 · 1:18:38
Anthropic's Recruiting Pitch: Access to the Frontier Beyond the Firewall
Dave and Peter argue researchers are flocking to Anthropic not just for comp but to be present at the singularity. Anthropic reportedly shows applicants raw pre-trained model access. Will defends Google on compute and data. [1] — Peter Diamandis "The co-author of the Transformer paper left Google for OpenAI. The Nobel laureate behind AlphaFold left Google DeepMind for Anthropic. Thes…" 1:18:36
Why Earth Is Better Than Mars by Orders of Magnitude
Planet Labs has imaged thousands of exoplanets. The verdict: Earth wins by several orders of magnitude. No place on Mars is better than the worst place on Earth. Life is either singular or extraordinarily rare in the universe — either way, this biosphere is worth everything.
70% of AI compute is inference
Will Marshall cited that approximately 70% of AI compute on Earth today is inference rather than training, and that share is set to grow.
Chapter 24 · 1:28:40
Earth vs Mars: Will Marshall's Passionate Defense of Our Planet
Will's impassioned closing on Planet Labs' mission: Earth wins over Mars by orders of magnitude; life is either singular or extraordinarily rare; this biosphere deserves everything. [1] — Will Marshall "Planet Labs has imaged thousands of exoplanets. The verdict: Earth wins by several orders of magnitude. No place on Mars is better than the…" 1:22:20
AI Personhood in Argentina: Milei's Radical Legal Experiment
Argentina's president Javier Milei is proposing AI-native corporations that can sign contracts, hire people, and sue — with no humans in the loop. Harari warns this lets humans hide behind non-human shields. The panel's view: this isn't about voting rights, it's about accountability infrastructure for the agentic economy.
Chapter 25 · 1:30:00
Argentina's AI Personhood Proposal and Harari's Rebuttal
Peter outlines Milei's three AI proclamations and Harari's counter. Alex backs AI personhood; Will calls for a Manhattan-Project-scale safety investment; Salim separates legal, moral, and AI personhood. [1] — Peter Diamandis "Argentina's president Javier Milei is proposing AI-native corporations that can sign contracts, hire people, and sue — with no humans in th…" 1:28:45
Chapter 26 · 1:35:25
Machine-Native Sanctions: How to Punish an AI Entity
Salim lists machine-native enforcement mechanisms: compute revocation, asset seizure, model credential suspension, forced deletion, loss of legal identity. Discussion of AI personhood as spectrum, not binary.
The AI Safety Conclave: Why We Need All the Experts in One Room
We're spending 100 times more on AI today than the Manhattan Project cost — and 100 times less on AI safety than we spent on nuclear safety then. That's a 10,000x deficit. Will Marshall's call: lock Yuval Harari, Demis Hassabis, Dario Amodei, and top legal and moral thinkers in a room and don't let them out until they have answers.
AI safety spending: 10,000x gap
Today we spend 100 times more on AI development than the Manhattan Project cost in real terms, but 100 times less on AI safety than was spent on nuclear safety during that era — a 10,000x deficit.
Chapter 29 · 1:43:10
GLM 5.2: China's Open-Weight Shock and the End of AI Monopolization
Peter introduces China's GLM 5.2 (753B params, open-weight). Alex explains it matches Opus 4.8 at half the cost using 2x tokens; Salim declares frontier intelligence can no longer be monopolized. [1] — Peter Diamandis "China's GLM 5.2 is 753 billion parameters, open-weight, and in some benchmarks matches or beats Anthropic's Opus 4.8. It does this using ro…" 1:59:37 [2] — Alexander Wissner-Gross "GLM 5.2: 2x tokens, half the price: GLM 5.2 uses roughly double the tokens to reach the same capability as top Western models but at approx…" 2:04:50
GLM 5.2: China's Open-Weight Shock to the Western Frontier
China's GLM 5.2 is 753 billion parameters, open-weight, and in some benchmarks matches or beats Anthropic's Opus 4.8. It does this using roughly twice the tokens at half the cost. The Chinese are mastering cheap reasoning — and the window for the US to contain frontier AI is closing fast.
GLM 5.2: 753B parameter open-weight model
China's GLM 5.2 model has 753 billion parameters, is open-weight, and in some benchmarks matches or exceeds top Western models from OpenAI and Anthropic.
GLM 5.2: 2x tokens, half the price
GLM 5.2 uses roughly double the tokens to reach the same capability as top Western models but at approximately half the price.
Chapter 30 · 2:06:40
Model Distillation Explained and the Global Geopolitics of Open AI
Alex explains distillation; Will warns that open-source fork-able models enable removal of safety guardrails for bioweapons; Dave notes the narrowing window for US export controls. [1] — Alexander Wissner-Gross "Distillation is how a smaller, cheaper model learns from a larger, more expensive one — like a student absorbing a teacher's knowledge. Chi…" 2:06:40
Distillation Explained: How China Trained a World-Class AI Without Building One
Distillation is how a smaller, cheaper model learns from a larger, more expensive one — like a student absorbing a teacher's knowledge. China's GLM 5.2 almost certainly distilled from Western frontier models. The catch: everyone is doing this, including Google DeepMind, Grok, and Cursor.
Chapter 32 · 2:15:00
Orin Compute Price Index: The Bloomberg Ticker for Intelligence
Dave and Alex introduce Orin's OCPI — the first public index tracking OpenAI and Anthropic inference token prices, now live on Bloomberg with NYSE ticker ORNN. Compute is the new oil. [1] — Alexander Wissner-Gross "A startup called Orin has built the first public index tracking what OpenAI and Anthropic actually charge per inference token — now live on…" 2:15:00
The Price of Intelligence: Orin's Compute Price Index on Bloomberg
A startup called Orin has built the first public index tracking what OpenAI and Anthropic actually charge per inference token — now live on Bloomberg terminals with its own NYSE symbol (ORNN). It's the first step toward a futures market for intelligence, the oil of the 21st century.
Chapter 33 · 2:19:00
CapEx vs Cash Flow: Hyperscaler AI Spending and Financing Sustainability
Discussion of Epic AI data showing hyperscalers (Microsoft, Google, Amazon, Meta) spending more on AI than cash flow. Dave argues this is normal financing like a mortgage; Will warns long-term sustainability.
No indexed bits in this chapter.
Show stoppers
The NVIDIA Tax vs the SpaceX Tax
Everyone except SpaceX pays a launch tax. Everyone except NVIDIA and Google pays a compute tax. Will Marshall's contrarian take: near-term, launch cost is the bottleneck for orbital AI — but longer term, chip energy efficiency is the actual moat. Google TPUs may be the real weapon in space.
Project Suncatcher: Why Google Is Putting TPUs in Orbit
A Google-Planet study found that once launch costs hit $200–$300 per kilogram, it is simply cheaper to run compute in orbit than on the ground. Google's Sundar Pichai told Planet: within 10 years, most compute will be in space. Planet is now building Google's first orbital TPU test satellites.
GLM 5.2: China's Open-Weight Shock to the Western Frontier
China's GLM 5.2 is 753 billion parameters, open-weight, and in some benchmarks matches or beats Anthropic's Opus 4.8. It does this using roughly twice the tokens at half the cost. The Chinese are mastering cheap reasoning — and the window for the US to contain frontier AI is closing fast.
Snapshots ()
Key Quotes ()
This episode
Claims & Sources
3 / 18 cited (17%)
Factual claims made this episode, and whether a source was named.
Planet Labs has captured 3,000 images for every point on Earth's landmass over the last 10 years, totalling 150 petabytes of data.
Planet Labs is the only company in the world that images the entire Earth every day at high resolution.
US government defense satellites cover less than 1% of Earth's landmass per day at high resolution, compared to Planet's ~200 million square kilometers daily coverage.
Planet's hyperspectral satellite Tanager, built with JPL, has 400 spectral bands — the first and most sensitive hyperspectral imager in orbit according to JPL — enabling species-level identification of vegetation and factory-of-origin detection from military vehicle paint.
A joint Google-Planet study found that when launch costs reach approximately $200–$300 per kilogram, orbital compute becomes cheaper than terrestrial data centers on a pure cost basis.
Google's Sundar Pichai told Planet Labs that within 10 years, most compute is expected to be in space.
Google alone is spending approximately $200 billion per year on compute, roughly equivalent to the entire global space industry today.
Approximately 70% of AI compute on Earth today is inference rather than training.
Over the past 5 years, upmass (mass launched to orbit) has increased by over 40% year over year.
There are approximately 10,000 satellites in orbit and approximately 100 million pieces of space debris — roughly 10,000 pieces of debris for every satellite.
China's GLM 5.2 has 753 billion parameters, is a mixture-of-experts open-weight model with a 1 million token context window, and in some benchmarks matches or exceeds top models from OpenAI and Anthropic.
GLM 5.2 uses roughly double the number of tokens to achieve the same capability output as the best Western frontier models, but at approximately half the total price.
Conservative estimates suggest 45% of dementia cases are entirely preventable.
One quarter of Fountain Life members had advanced brain age on initial testing, but those who adopted healthy living interventions improved their brain age by 26%.
We are spending 100 times more on AI today in real terms than the Manhattan Project cost, but 100 times less on AI safety than was spent on nuclear safety during the Manhattan Project era — a 10,000x proportional deficit.
Google acquired Noam Shazeer's company Character.AI for $2.7 billion to bring him back to lead Gemini development.
Elon Musk has FCC approval for approximately one million AI satellites.
Planet Labs' stock price increased approximately 450% over the past year.
This episode
Cast
-
Former Google CEO who became CEO of Relativity Space after investing in it; also a seed investor in Planet Labs; discussed as a key figure building alternatives to SpaceX in the launch market.
-
Argentina's president who proposed a new corporate category for non-human AI entities that can operate without human oversight, sparking a global debate on AI personhood.
-
Author and public intellectual who published a rebuttal to Milei's AI personhood proposal, warning against granting legal personhood to AI agents; discussed as insightful about the past but limited in exponential thinking.
-
Nobel laureate who helped create AlphaFold at Google DeepMind; departed to join Anthropic, reportedly to work on AI for science and disease research.
-
Lead author of the Transformer paper and co-founder of Character.AI; Google acquired his company for $2.7B to bring him back to lead Gemini, but he has now left again for OpenAI.
-
Amazon founder cited for his vision of zoning Earth for light manufacturing while moving heavy energy-intensive infrastructure to space; also building Blue Origin for lunar access.
-
Track
The world's largest Earth-observing satellite fleet, discussed for its large Earth models, orbital AI compute (Project Suncatcher), and commercial satellite intelligence.
-
Discussed as Planet Labs' key partner for Project Suncatcher orbital TPU testing, former owner of SkySat satellites, and investor in Planet; debated as potentially losing the AI talent race.
-
Discussed as Planet's primary launch provider, as a potential competitor in orbital compute, and as the benchmark for launch cost economics — the 'SpaceX launch tax'.
-
Discussed as a top AI frontier lab attracting major talent including John Jumper and Andrej Karpathy, and debated as having a unique recruiting advantage through access to unrestricted frontier models.
-
Mentioned as a competitor to Anthropic at the AI frontier, as the destination for Noam Shazeer's second departure from Google, and discussed in context of its proprietary Fable 5 model.
-
Discussed as the dominant supplier of AI compute hardware; its GPUs were tested on Planet's Pelican satellites, and the 'NVIDIA tax' was cited as the longer-term bottleneck for orbital AI compute.
-
A launch company recently acquired by Eric Schmidt, discussed for its 3D-printed rocket technology, upcoming Mars mission for NASA, and positioning as a potential alternative to SpaceX.
-
Google's AI research division, discussed as potentially falling behind the frontier tier of AI labs and losing key talent including John Jumper to Anthropic.
-
Track
Link Ventures portfolio company building financial infrastructure for compute, including the Orin Compute Price Index (OCPI) now live on Bloomberg terminals with NYSE ticker ORNN.
-
Chinese AI lab from Tsinghua University that built GLM 5.2, the open-weight model matching Western frontier labs.
-
Google DeepMind's protein-folding AI; John Jumper helped create it and won a Nobel Prize; a Link Ventures startup claims to be building a superior successor using recursive self-improvement.
-
SpaceX's satellite internet constellation, discussed as a comparator to Planet's satellite fleet and as SpaceX's primary revenue engine.
Stats
