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

Welcome back to Moonshots—home base with the whole crew in studio and a special guest, Will Marshall of Planet. We’re diving into Planet’s large earth models and orbital AI cloud, Eric Schmidt’s big rocket bet at Relativity Space, fresh AI talent moves, and a sharp look at China’s new model; Will, kick us off: what exactly are large earth models?

Think of LLMs that read the web; we’re building models that read the planet. Phase one fuses our global sensor data with language models so you can ask physical questions about Earth; phase two moves compute next to the sensors in orbit for speed and scale.

And every European government wants in, with billion‑dollar deals piling up.

We’ve captured daily imagery for a decade—3,000 shots per spot on land—so you can compare today with the full history, whether you’re a farmer, a journalist, or a defense analyst.

Specs check: what resolution and bands are you flying?

Three fleets: a daily mapper at three meters moving to one meter with eight bands, a high‑res system going from 50 to 30 centimeters with sub‑hour latency, and a hyperspectral imager with roughly 400 bands that can fingerprint materials, species, and emissions.

From phonesats to a rocket‑ship stock—Planet’s growth is wild; how is AI unlocking the value?

Space and AI are finally married: our imagery turns LLM theory into real‑world answers, so a grower can ask how to boost yield on their actual field and a reporter can see flood impact today, not just read about hydrology.

Crystal ball time: can Planet auto‑predict Earth like a future video model?

It’s emerging; we already forecast data‑center build completions by learning from U.S. sites and projecting timelines abroad, and we’re expanding from hindsight to foresight quickly.

We’re also compressing Earth into an embedding space—tile‑to‑text at planetary scale—so you can search objects across time and set up multi‑layer forecasting once thousands of temporal layers are encoded.

Global transparency brings geopolitics; how do you handle sovereignty and risk?

Our mission is accountability through visibility; fewer dark corners means fewer miscalculations, better disaster response, and more enforceable peace.

Do most governments beyond the superpowers actually use this stuff?

Not much yet, but AI shrinks the skill gap so NGOs and small agencies can ask direct questions without teams of imagery analysts.

Revenue mix?

Roughly 60 percent defense and intel, 25 percent civil government, 15 percent commercial, and commercial is accelerating thanks to AI.

How do you price AI training on your data?

We license access and usage; train all you want, and when you need fresh answers you hit our APIs.

Any export filters or area blurring?

We’re licensed under NOAA, follow U.S. and EU blacklists, and we don’t downsample specific sites; at hundreds of kilometers up, you get meaningful transparency without personal privacy exposure, which is why open overflight became the global norm.

You tested Nvidia in orbit; what’s the edge‑compute win?

We detect targets onboard, pass results across satellite links, and return answers in seconds—think fire mapping in minutes instead of hours, where time saves lives and property.

Where does resolution go next—and any aircraft in flight?

Daily scan is heading to one meter with super‑res improvements, tasking to around 30 centimeters; we catch planes in the air often, and despite everyone’s hopes, no UFOs.

Why not bolt cameras on Starlink?

Wrong orbits and lighting for consistent imaging; Starlink is tuned for comms, not daily optical mapping.

Since 2013 our radios jumped from megabits to multi‑gigabits and cameras from a few to dozens of megapixels per frame; the bigger unlock now is AI, which could 100 times the usable value trapped in our archives.

For everyday people, this turns into natural‑language queries of the physical world—permitting checks, field health, flood extent—answered against a decade of daily history.

GDP‑maxing with reinforcement learning on planetary pixels?

We can help markets, but the deeper prize is smarter stewardship—agriculture alone has order‑of‑magnitude gains if we manage inputs with precise, real‑time insight.

How do you cool GPUs in vacuum?

Radiators and thermal design—no magic—radiation scales with temperature to the fourth power, so you point the fins at the dark and dump heat efficiently.

Project Suncatcher: you’re putting TPUs in orbit and nudging toward a Dyson‑swarm‑style compute ring; how does this stack up against Elon’s orbital data centers?

We’ve launched hundreds of satellites, mostly on SpaceX, and we’re building clustered racks in sun‑sync orbits; once launch hits the right band, pure cost tilts in favor of space compute since you beam up questions and beam down answers rather than power, water, and cooling.

Google chose us for early tech demos—TPUs, radiation, cooling, optical interlinks—because compute in space sidesteps local land, water, and grid fights while scaling with sunlight.

Debris risk?

We fly at roughly 400 to 500 kilometers where orbits self‑clean in months to a few years; the real hazard is old fragments higher up, which we can nudge apart from the ground with lasers to prevent cascades.

Up‑mass is compounding; where does this head?

Forecasts get fuzzy far out, but near‑term we should move energy‑hungry compute off‑planet to cool tensions over land use, water, and power.

Okay, the elephant: Elon has rockets and scale—how do you compete?

This will be coopetition; they’ll throw mass, we’ll throw clever systems, and over time the bigger tax isn’t launch, it’s compute efficiency—TPUs’ flops‑per‑watt can decide winners when every watt in orbit drives radiator mass.

So training stays ground‑based while inference goes orbital?

Inference first to space—latency‑light, distributed runs; training follows later, but the bulk of AI cycles already skew to inference.

Launch market update: Relativity Space now led by Eric Schmidt, with a new NASA Mars orbiter on deck; thoughts?

Eric backed Planet early and spots talent and timing well; with fresh capital and focus, Relativity can go far—and we’d love another launcher in the mix.

With Relativity pivoting, is there a 3D‑printing gap for space?

Huge room there—on‑orbit design constraints differ radically from launch constraints, so printing and new architectures can unlock lighter, smarter hardware.

Is size or reuse the real driver on cost?

Reuse gets us to threshold, but long‑run economics hinge on compute watts per inference because cooling mass dominates; Google’s systems and interconnects are an overlooked superpower here.

One last space question: bigger prize in leaving Earth or making Earth smarter?

The best world by orders of magnitude is the one we’re standing on; our lane is space for Earth—observe it daily and shift heavy compute upstairs so life can flourish down here.

AI brain‑drain update: Noam Shazeer heads to OpenAI and John Jumper to Anthropic; is the frontier consolidating?

Right now the sharpest edge looks like a two‑horse race with raw access to frontier pre‑trains as the lure; Google’s great at cheap, fast models for search, but it didn’t show a new top‑end at IO.

Talent flows are noisy, and I wouldn’t bet against Google—Gemini’s strong, TPUs are real leverage, and they’ve got billions of users ready for AI‑powered products.

Some researchers want to be at ground zero of self‑improving systems and will trade comp for unfettered model access; seeing the unfiltered frontier is a hard pitch to turn down.

I’ve seen that movie—show someone what’s behind the firewall and they can’t unsee it; still, I’m bullish on where this all goes and want us aiming at a positive future.

Before we get heavy, quick book corner: my go-to sci‑fi is Accelerando; what’s your favorite future on the page?

I read more science than sci‑fi, but the early cyberpunk canon nailed near‑term futures; these days I’m drawn to what Nature reveals each week.

On why small beats big: agency wins; a tight team ships without permission layers, like Facebook out-iterating Google during the Google Plus era.

Intelligence needs a body; space sensing and on‑orbit compute can complete the real‑time loop, pushing us from planetary data toward planetary intelligence and, with luck, wiser alignment.

Case in point: a five‑person team claims a recursive, self‑improving system that could surpass AlphaFold, showing what agency plus RSI can do.

If today’s frontier models already learn from images, video, and world models, why insist orbital data is essential to ground physical understanding?

Embodiment matters; watching a million first‑person clips isn’t living in the world, and I think that difference becomes pivotal for higher‑order learning.

Big policy turn: Argentina’s president pitches an AI haven with no AI regs, a non‑human corporate class, and ultra‑low taxes, even proposing AI agents that can incorporate, contract, hire, and sue—Alex, where do you land?

I’m for AI personhood via non‑human corporations; we can sanction misbehavior by pausing, throttling, or otherwise penalizing agents, and broader plural personhood will matter as new forms of minds emerge.

It’s promising and perilous; we’re racing on capability while underinvesting in governance and safety by orders of magnitude, so we need a serious, inclusive process—more Manhattan Project‑level foresight than hot takes.

Don’t conflate moral personhood with legal personhood; build machine‑native accountability—think compute revocation, credential suspensions, network bans, containment, or even identity loss, like time‑outs for machine ‘children.’

For clarity, Argentina’s not giving AIs civil rights; they’re allowing AI‑run corporations to bank, contract, and operate, which is the right debate to have.

Next up, China’s GLM 5.2 just topped open‑weight leaderboards with a million‑token window and strong results—how big a deal is this?

It challenges the neat story that China lags six to eight months; in practice it’s close enough on coding and long‑range reasoning that teams are swapping it in locally over top Western APIs.

Expect heavy use of distillation—compressing teacher outputs into smaller students—which accelerates catch‑up and raises biosecurity worries when guardrails can be removed.

Distillation is teacher‑student learning at scale; big, expensive teachers generate traces that train cheaper students, now a core loop for progress.

The striking part is reasoning‑centric benchmarks and cheaper long‑thought; they spend more tokens to think step by step, yet still hit attractive cost per answer.

That’s why inference per watt is strategic; whoever wins TPUs and power efficiency wins long‑thought AI and, by extension, space compute.

This dwarfs nukes in stakes; alignment under recursive self‑improvement is the pivotal question for our species.

Quick Fermi check: why assume it’s a paradox at all?

Maybe intelligence finishes the puzzle, goes mostly digital, and stops expanding; or we hit the great filter—tech outpaces social systems and we self‑terminate, so muddling through is not an option.

Another take: long‑stable oceans on Earth created a rare runway for life; that alone could thin the odds elsewhere.

On financing, Orin is packaging the compute economy so capital can flow into GPUs, data centers, and beyond with liquid instruments.

Disclosure: I advise Orin; compute is the new oil, and their compute price index is already on Bloomberg, with a NYSE symbol to catalyze hedging and capex planning.

Hyperscalers are outspending cash flow on AI; if sentiment wobbles, does the build‑out stall?

It’s like a mortgage on a thirty‑year house; they can finance far beyond today’s cash, and the world’s capital wants in.

Long‑term control is the real question—who commands the system at scale, firms or AIs—and sustainability depends on that answer.

Also, they can raise prices to bolster operating cash, and we’re already seeing successful hikes.

Intelligence is getting cheap; manufacturing intelligence is getting very expensive.

Wildest part: a handful of companies are profitable enough to midwife an entire new industry out of their own cash flow.

Closing us out with Ekram Alam’s Moonshots theme: stay optimistic; if we align it, AI can help us rise above our ancient wiring and build abundance—moonshot mates, thanks for riding with us.

Highlights for me were low Earth orbit, the Kessler risk, and TPUs as the key driver.

Tech is the engine of progress; our job is to pull out the promise while fencing the peril.

We’re wiring up a planetary nervous system on the way to planetary intelligence, which we’ll need to earn planetary wisdom.

Parting thought for the Fermi fans: don’t sleep on the galactic zoo hypothesis.

Third‑gen biosphere or not, we’re grateful you’re here—Alex, Dave, Salim, love you all, and to every moonshot mate listening, thank you.