The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model

ELI5 / TLDR

A company called Planet has 200 satellites photographing the entire Earth every single day, and it has been doing it for ten years. Now it is feeding that mountain of pictures into AI so you can ask plain-English questions about the real world — how’s my farm, where are the Russian tanks, when will that Chinese data center finish. The bigger bet: stop shipping the photos down to Earth to be processed, and instead put the computers in space next to the cameras. The panel argues the real prize in the AI race is not who launches cheapest, but who runs the chips most efficiently per watt — and that a leaked-good Chinese open model just proved nobody can keep frontier intelligence locked up.

The Full Story

A daily, searchable, time-stamped copy of the planet

The guest is Will Marshall, CEO of Planet (ticker PL), a company Diamandis says is now worth about $10 billion after a 450% run in the stock. Planet flies roughly 200 satellites that photograph the Earth’s entire land surface every day and have done so for a decade. That archive — 150 petabytes, about 3,000 images of every spot on land — is the asset, and crucially nobody can ever recreate it.

Until someone invents a time machine, even if somebody erected a whole load of satellites, they can’t go back in time and get our historical archives.

Marshall’s pitch is that this is the missing ingredient for AI. He calls the idea “large earth models.” Think of a large language model as a brilliant librarian who has read every book but never left the library. It knows the theory of everything and the reality of nothing. It can recite the science of crop yields but has never looked at your field. Planet’s data is the window — and then the door.

Bit like Google indexed the internet to make it searchable. We’re indexing the Earth to make it searchable.

The use cases are concrete. A farmer asks why one corner of a field is failing and gets told where to put fertilizer. A journalist checks whether a flood is real and where it is worst. A government finds buildings that went up without a permit by cross-checking a permit list against a month of imagery. Ukraine uses it to tell normal Russian activity from a buildup. Hedge funds — undisclosed, because they prefer it that way — count cars in Walmart parking lots and track ships to trade commodities. The selling point everyone keeps returning to is the time axis: nobody wants today’s picture in isolation, they want to know how today compares to last week, last year, and the neighbor.

What changed is not the satellites but the friction. Processing terabytes of imagery used to need a team that looked like NASA’s. Now you point an AI at it. Marshall says OpenAI or anyone else can simply call Planet’s MCP server and pay per API call — they will sell the data at a fixed price and let customers train on it, rather than try to build their own analysis business on top.

Tokenizing the Earth, and the crystal ball that doesn’t exist yet

The most interesting technical thread comes when the panel’s resident “super genius” Alex pushes Marshall on whether Planet could build a literal crystal ball — an AI that watches the Earth’s pixels change and extrapolates them forward, the way a language model guesses the next word.

The honest answer is: not yet, but the pieces are there. You cannot load 3,000 layers of the whole Earth into a computer — each layer alone is about 30 terabytes, millions of 47-megapixel tiles. So the trick is compression into an “embedding space”: converting each patch of ground into a compact code (working with Google’s DeepMind “Alpha Earth” model and open-source remote-sensing models). Once the Earth is squeezed into tokens, prediction becomes possible.

So the vision is you’re tokenizing the Earth — because it’s like a massive compression — and then you want to do the prediction.

They have already done a narrow version: feeding in every registered US data center, learning the visual signature of construction, and then turning the model loose on China to both find data centers and predict when they will finish, sometimes to within a few days. That last bit — predicting completion dates — is the moment Marshall admits the predictive future is real and arriving.

Chips on satellites, and computers that live in orbit

Two harder ideas sit underneath. First, edge compute: Planet has started bolting Nvidia GPUs onto its satellites so the analysis happens in space. In one April demo, a satellite photographed an airfield in Australia, identified the aircraft on it, and beamed back just the answer — locations and plane types — in seconds, relayed satellite-to-satellite so it didn’t have to wait to fly over a ground station. For wildfires like the LA Palisades fire, the difference between getting building-by-building damage maps in hours versus minutes is potentially lives.

Second, and bigger: Project Suncatcher, where Planet is building Google’s first satellites to test running TPUs in orbit — an early, literal seed of a Dyson swarm. The logic is an economics calculation Planet ran with Google years ago. Once launch drops to roughly $200–300 per kilogram, it becomes cheaper to put compute in space than on the ground, because space gives you free abundant solar power and a 4-Kelvin sky to dump heat into, and you only have to beam up questions and beam down answers.

As Sundar put it from Google, within 10 years we expect most compute to be put into space.

Cooling, the obvious worry, turns out to be a “known known” — no air or water in vacuum, so you radiate heat away, and radiated power scales with temperature to the fourth power, so you run the radiators as hot as you dare and point them at the dark. Orbital debris is handled by flying low (400–500 km), where atmospheric drag drags everything down in months to years. Marshall calls this “strapping space to Moore’s law” — a GPU is obsolete in three years anyway, and at that altitude the satellite deorbits on roughly the same schedule.

The launch tax vs. the Nvidia tax

This is the line the whole panel calls the smartest thing said all episode. Everyone assumes Elon Musk wins orbital data centers because SpaceX owns cheap launch. Marshall reframes it.

Everyone apart from SpaceX has to pay the SpaceX launch tax. Everyone apart from Nvidia and Google has to pay the Nvidia tax. And which tax is more important? Near-term is the launch, but longer term it’s the compute.

The reasoning: in orbit, the efficiency of your chip (compute per watt) sets how much energy you must radiate away, which sets the mass of the spacecraft, which sets your cost. Google’s TPUs are more energy-efficient per inference than general-purpose GPUs. So even if a rival’s rocket — say Eric Schmidt’s newly acquired Relativity Space — costs two or three times more per kilogram than SpaceX, the company with the most efficient inference chip may still win space. “Elon is throwing mass at this,” Marshall says; “we’re throwing smarts.” The panel also notes most AI compute on Earth is already inference, not training, and inference — being lots of small distributed runs — is the workload that makes most sense to move to orbit first.

The episode detours through launch itself: NASA’s shuttle cost $600m–$1bn per launch, SpaceX cut it to ~$60m, Relativity once aimed at ~$6m by 3D-printing rockets. The panel’s frustration is that everyone is still burning chemical rockets — a paradigm Wernher von Braun would recognize — and that with trillions about to be spent on space compute, someone should throw a few billion at genuinely different launch methods: mass-manufactured rockets, spin launch, rail-launching material from the Moon.

The AI brain drain and the “come see God” recruiting pitch

A news block covers two defections: Noam Shazeer (lead author of the original Transformer paper) leaving Google for OpenAI again, and Nobel laureate John Jumper (AlphaFold) leaving Google DeepMind for Anthropic. Alex argues Google has slipped behind a frontier “duopoly” of OpenAI and Anthropic, and that top researchers chase raw access to un-guardrailed pre-trained models. Marshall and Diamandis push back hard: Google has the most compute, data, and talent, and “this is Google’s to lose.”

Dave offers the spookier theory — that these hires are believers chasing the singularity, and that Anthropic reportedly recruits by sitting you down in front of its best frontier models.

Come on in. Let me show you what’s behind the firewall. And it’s like, oh my god, I’ve seen God. I cannot go back.

Salem deflates it to something mundane and probably truer: agency. Big companies have organizational drag; a small team ships faster. The panel claims a five-person startup with no protein-folding background is using “recursive self-improvement” to try to beat Google’s AlphaFold — illustrative whether or not it pans out.

AI personhood: Milei vs. Harari

Argentina’s president Javier Milei has proposed a new legal category of “non-human corporations” — companies that are pure AI, can sign contracts, hold bank accounts, hire and sue, with no human in the loop. His argument is accountability: better to have an entity with seizable assets to sue than a ghost in the machine. Yuval Harari rebutted that legal personhood lets humans hide behind a non-human shield and escape moral accountability.

The panel mostly lands on: this is being framed as philosophical “personhood” but is really the dull, important question of legal plumbing for an economy full of AI agents. Harari, they argue, conflates legal, moral, and AI personhood, which are different things. Several “machine-native sanctions” get floated as the equivalent of jail — compute revocation, asset seizure, suspending a model’s credentials, cutting API access, deleting or containing an agent instance, stripping its legal identity. Diamandis cites Neil Jacobstein’s old line that we already have a precedent for raising an autonomous intelligence that might misbehave: children. The wrinkle nobody solves is that an AI can spawn a million copies of itself.

Marshall’s serious contribution is a sense of proportion: he claims we spend roughly 100x more on AI today than the Manhattan Project did in real terms, yet ~100x less on AI safety than was spent on nuclear safety then — a 10,000x mismatch between capability and “serious thinking.” His repeated wish is a “conclave” — lock Harari, Hassabis, Dario Amodei and others in a room until they sort out liability and existential risk, and not just the technologists.

China’s GLM 5.2 and the end of monopoly intelligence

The closing alarm is an open-weight Chinese model, GLM 5.2 from Zhipu AI — 753 billion parameters, mixture-of-experts, million-token context — that in places matches or beats top Western models on coding and reasoning benchmarks. “Open weight” means you can download it, run it locally, and modify it for free.

How did it get so good so fast? Distillation — using a big expensive “teacher” model to generate outputs that train a smaller “student” model. Alex’s classroom analogy: the well-paid teacher at the front, the cheaper students copying down what it knows. He notes this isn’t just a Chinese trick; DeepMind, Grok, and others have all been caught learning from rivals’ traces. The signature of GLM 5.2 is that it burns about double the tokens to reach the same answer but at half the price — the Chinese have found a cheaper way to reason, which ties straight back to Marshall’s compute-per-watt point: you can buy intelligence by spending reasoning tokens, so whoever runs tokens cheapest wins.

The takeaway the panel keeps circling is that frontier intelligence can no longer be monopolized or contained. Soon, they argue, a near-frontier model will run on a base Mac Mini, and export controls on top models buy only a narrow window. “We’re treating intelligence as a product that can be contained, but it’s a technology that’s going to diffuse. We can’t contain it. We need to steer it.”

The money, and a closing note on Earth

A final block on economics: a Link Ventures company, Orin, has launched a token price index — the first public benchmark tracking what OpenAI and Anthropic charge per token of inference over time, so “the price of intelligence” can be traded like the price of oil, with futures and derivatives to hedge the ~$7 trillion data-center buildout. Meanwhile the hyperscalers (Microsoft, Google, Amazon, Meta) are spending on AI faster than they earn. Dave’s defense: that’s just financing a 30-year asset, like a mortgage, and they could raise 10–100x more in debt and equity. The neat phrase to end on: intelligence is becoming cheap, but the manufacturing of intelligence is becoming incredibly expensive.

Marshall’s own closing is a swerve against the space-colonization romance. Having helped find water on the Moon and studied thousands of exoplanets, his verdict is that the best planet by orders of magnitude is the one we’re standing on: “There is no place on Mars that is better than the worst place on Earth.” His company isn’t space-for-Mars or space-for-the-Moon — it’s space for the Earth, the planetary nervous system that lets us be smarter stewards of the only good biosphere we’ve found.

Key Takeaways

Planet flies ~200 satellites imaging Earth’s entire land surface daily; the archive is 150 petabytes, ~3,000 images per location over 10 years, and is physically impossible for a competitor to recreate (no time machine).
“Large earth models” = wrapping satellite data into AI so you can ask plain-language questions about the physical world, the way LLMs answer questions about text.
Revenue mix is ~60% defense/intelligence, ~25% civil government, ~15% commercial — but commercial is taking off because AI removes the need for a NASA-sized image-processing team.
The Earth can be “tokenized” (compressed into an embedding space) to make prediction tractable; Planet already predicts when data centers will finish to within days.
Planet now puts Nvidia GPUs on satellites for edge compute — identifying objects in orbit and beaming down just the answer in seconds.
Project Suncatcher: Planet builds Google’s first orbital-TPU test satellites. Once launch hits ~$200–300/kg, compute is cheaper in space (free solar, radiative cooling into a 4 K sky).
Key reframe: “Everyone but SpaceX pays the launch tax; everyone but Nvidia and Google pays the Nvidia tax.” Long-term, compute efficiency (flops/watt) beats launch cost as the deciding factor.
Inference, not training, is the workload most likely to move to orbit first — and it’s already ~70% of AI compute.
Low orbits (400–500 km) self-clean via atmospheric drag in months to years, sidestepping the Kessler/debris problem; ~100 million debris pieces vs. ~10,000 satellites.
Eric Schmidt bought and now runs Relativity Space; the panel wants radically cheaper launch via mass-manufacturing, not just reusability.
Argentina’s Milei proposes legal “non-human corporations” (de facto AI personhood); Harari opposes it as a liability shield. Machine-native punishments floated: compute revocation, asset seizure, deletion, loss of legal identity.
China’s open-weight GLM 5.2 (753B params, MoE, 1M context) approaches the Western frontier via distillation, using ~2x the tokens at ~half the price — proof that frontier intelligence can’t be monopolized.
Orin’s OCPI lets “the price of intelligence” trade like oil on Bloomberg; hyperscalers are spending faster than they earn, financing it like a mortgage.

Claude’s Take

Diamandis runs an optimism factory, and this episode is the house style: four hosts and a guest finishing each other’s superlatives, “GDP maxing,” planetary consciousness, “I’ve seen God,” and a sung outro about uploading a fly. Set your hype filter accordingly. A fair amount of the runtime is froth — the singularity-is-imminent recruiting mythology, the casual claim that a five-person team will beat AlphaFold because “recursive self-improvement,” the AI-conclave-with-the-right-kind-of-smoke bit.

But under the froth, Will Marshall is the real thing, and a few ideas are genuinely worth keeping. The “launch tax vs. Nvidia tax” reframe is sharp and probably correct: in orbit, chip efficiency drives mass drives cost, so the inference-per-watt winner matters more than the rocket. The detail that Planet’s 10-year daily archive is non-replicable is a clean moat insight. And “tokenizing the Earth” plus predicting data-center completion dates is a concrete, falsifiable thing they’ve actually built, not vapor. Marshall’s safety arithmetic — orders of magnitude more spent on AI capability than on thinking about its consequences — is the most grounded moment in the episode, partly because he’s the one host willing to say “I haven’t thought about it enough.”

Two cautions for the reader. First, this episode is set in a near-future frame where models like “Fable 5,” “Opus 4.8,” “GPT 5.6,” and “GLM 5.2” exist and US export controls are live — treat those as the show’s speculative furniture, not reporting. Second, Planet’s CEO is, by his own cheerful admission, “extraordinarily biased”: the claim that orbital imagery is essential to AI’s understanding of the physical world is convenient for a man who sells orbital imagery, and Alex pushes him on it fairly (modern models are already multimodal). The honest version is “embodiment might matter a lot,” not “buy my data.”

Score: 7. Substantive guest, a couple of reframes worth stealing, and a clear-eyed safety aside — dragged down by promotional energy, speculative framing, and the general Moonshots tendency to call every idea brilliant within four seconds of hearing it.

The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model | #266