Planet Labs Factory Tour: Will Marshall on 700 Satellites and Earth AI

Planet Labs CEO Will Marshall tours the San Francisco factory where the company has built about 700 Earth imaging satellites, for less than one NASA mission. The walkthrough, hosted by Molly O’Shea of Sourcery, goes from the garage-era whiteboard to the clean room where Pelican 10 sits fully integrated, through the shock, vibration and thermal test floor, and up to a new AI interface that lets anyone ask plain-language questions of ten years of daily imagery of the whole planet. Along the way Marshall explains why Planet (NYSE: PL) thinks of itself as a data company with a satellite back end, and why a 100% success rate in orbit worries him.

TLDW

Planet shrank satellites from bus-sized, billion-dollar machines to small, mass-produced units built from consumer electronics, cutting costs by roughly 1,000x and letting about 200 satellites image Earth’s entire landmass every day (about 4 million images daily). Marshall covers the engineering that makes it hard (200 g shock loads, a 60 degree thermal swing every 45 minutes, optics that cannot move more than a thousandth of a millimeter), the 48 ground stations near the poles, the Pelican and upcoming Owl satellites, and why capability per satellite matters more than fleet size. The tour ends with an LLM-driven interface that searches the planet in under a second, demonstrated on Area 51.

Thoughts

The AI demo near the end is the real news in this video, and Marshall frames it with a sharp observation: large language models were trained on the text of the internet, so they know the theory of everything and the current state of almost nothing physical. A chatbot can explain agronomy to a farmer but has no idea what is happening in that farmer’s field this week. Bolt an LLM onto ten years of near-daily imagery of every square meter of land (3,000 layers, by his count) and you get something that can answer questions about the actual world rather than the written one. That is a meaningful shift in who Planet’s customer can be. For most of the company’s history the buyers were people who already knew how to read satellite imagery: NASA, the NRO, governments, analysts. A natural language front end, in any language, turns that into a product a journalist covering a flood or a farmer tracking yield can use without training. The demo itself was rough (the search tool did not fire, and the Area 51 comparison mostly showed that “something changed”), and Marshall admits it is an early beta. But the strategic logic is clear. Planet’s moat is the archive, and the cheapest way to make an archive more valuable is to let far more people ask it questions.

The most counterintuitive line in the tour is Marshall calling Planet’s reliability record a problem. Every one of the roughly 700 satellites that reached orbit on a working launch turned on, and the company has gotten some utility out of all of them, even the ones with an undeployed solar panel or an early reaction wheel failure. Most aerospace executives would put that on a slide. Marshall reads it as evidence that the team is being too careful, that somewhere a newer sensor or a riskier design was left on the table. This is the logical end point of the whole Planet philosophy. If satellites are cheap and you launch a few more than you need, a failure costs you one unit, not the mission. Once the economics change that much, a perfect record stops being a virtue and starts being a sign you are paying for safety you no longer need. It is the same attitude that made the company possible in the first place, applied to itself fifteen years later.

Marshall also pushes back on the way people usually score satellite companies, which is by counting satellites. Planet’s imaging fleet is about ten times larger than any other, but the number he cares about is area: Planet images roughly 100 times more of Earth’s surface per day than any other Western commercial operator, most of whom task a few locations and cover 1% or 2% of the landmass at most. The Owl generation will keep that daily global coverage while moving from 3 meter to 1 meter resolution (about ten times more data per area) and using optical links to get images down ten times faster, within minutes. The satellite count stays about the same or even drops. For anyone following the stock, that is the right lens. Revenue comes from how fresh, sharp and complete the picture of the planet is, not from the length of the launch manifest, and a constellation that delivers ten times more data with the same number of satellites is a better business, not a smaller one.

The origin story explains why all of this works. Marshall’s diagnosis of NASA is generous and precise: in the 1960s through the 1980s the agency had to invent its own microprocessors, solar panels and batteries because nothing it needed existed off the shelf, and that built an “invent everything” culture that outlived its reason. Today the dominant electronics research money is spent by Apple, Google and their peers, and the parts in a phone are far better than anything a space agency would design in-house. Planet’s bet was to follow the consumer electronics industry instead of competing with it. Even the failed suborbital phone launch proved the point, since the phone’s electronics survived a crash. Pair that with what Marshall says early in the tour, that Planet has the high barrier to entry of a satellite operator but the margins of a software company, and you get the whole thesis in two sentences: buy what the world already makes, build the systems around it yourself, and sell the data, not the hardware.

Key Takeaways

Planet has built about 700 satellites in 15 years, more than any company except SpaceX, and has launched on more than 40 rockets, four or five launches a year.
All 700 satellites together cost less than a typical NASA satellite, roughly a 1,000x cost reduction driven by miniaturized electronics.
Marshall describes Planet as a data business with a satellite back end: high barriers to entry, with margins that look like software.
About 200 Dove satellites image Earth’s entire landmass every day, producing about 4 million 47 megapixel images (on the order of 100 terapixels) across eight spectral bands.
From 500 km up, about the distance from San Francisco to Los Angeles, the satellites can resolve individual trees and houses.
Satellites must survive 15 g of launch shaking, 200 g stage-separation shocks and a swing from minus 40 to plus 20 degrees Celsius every 45 minutes, while optics stay aligned to within about a thousandth of a millimeter.
Downlink speeds rose from about 1 megabit per second on the first satellites to about 2 gigabits per second, roughly 1,000x more data per satellite.
Planet built its own network of about 48 ground station dishes, clustered at high latitudes in Alaska, northern Canada, northern Norway, New Zealand and Australia to catch its polar-orbit satellites.
The rule for vertical integration: Planet buys the chips and builds the systems, including telescopes, radios and structures.
A Dove takes about eight hours to build and the line can produce up to 40 a week. A new Berlin facility will double manufacturing space and can build up to 60 of the larger Pelicans a year.
The Owl satellites target daily 1 meter imagery, about ten times more data per area, delivered ten times faster, without growing the fleet.
Pelicans use xenon ion engines, which Marshall says are almost ten times more efficient in space than classical rocket engines.
The bigger collision risk is debris, not satellites: about 10,000 working satellites against roughly 100 million debris objects.
Tanager, the hyperspectral satellite built with NASA JPL, spots methane plumes. Shutting off one leak it found equals taking a couple of hundred thousand cars off the road.
Planet’s new AI interface pairs an LLM with ten years of imagery and is in open beta.

Chapters

0:48 700 Satellites for Less Than One NASA Mission

The tour starts in the satellite lab, where Planet had just launched 20 more satellites the week before. Marshall explains that traditional satellites were the size of a double-decker bus and cost a billion dollars or more, while Planet’s entire fleet cost less than one of them. He frames the company as an information services business, with the hard-to-copy capex of a satellite operator and the margins of software, and points out that its new work with Google pushes further into intelligence.

3:45 Seeing Individual Trees From 500 km Away

In front of the whiteboard from the original garage, Marshall walks through an older Dove: the telescope and camera, the star camera that recognizes constellations to work out where it is pointing, the reaction wheels that turn it, and the small radios that push about 2 gigabits per second to the ground. Around 200 of these image the entire planet each day, roughly 4 million images in eight spectral bands, which is why machines and vision language models now do most of the looking.

7:44 Surviving Launch and the $500 Phone Test

Dropping a satellite onto a marble floor roughly matches the 200 g shock of rocket stage separation. The hard part is not the optical design but keeping the telescope in focus through launch and through a thermal swing every 45 minutes. Marshall then shows the remains of the phone Planet’s founders flew on a suborbital rocket to answer a question asked at NASA: why spend a billion dollars when a $500 phone has most of what a satellite needs? The launch failed, but the electronics survived.

12:05 1,000x More Data Than the First Satellites

The biggest change from the first satellites to today is how much data each one gets to the ground. Early units sent about a megabit per second, far too slow for 47 megapixel images. Current satellites send about 2 gigabits per second, roughly a thousand times more per satellite per unit of time, and that is what made Planet’s first mission, imaging the whole Earth every day, possible.

14:47 48 Ground Stations and a Flock of Bird Names

Because nobody sold the ground network it needed, Planet built roughly 48 parabolic dishes, often four or eight per site, as close to the poles as an internet connection allows. Marshall also explains the naming: Doves for peace and security, Tanagers for the colorful tropical birds whose ecosystems the hyperspectral satellite tracks, Pelicans because they fly low for lift the way the high-resolution satellites fly low for sharper images, and Owls for big eyes and onboard compute.

18:19 Inside Pelican 10: Buy the Chips, Build the Systems

In the clean room, Pelican 10 sits fully integrated with its propulsion tanks, enclosed star cameras and a telescope wrapped inside the main body. Planet buys telescope blanks and chips from suppliers but builds the systems itself. Doves take about eight hours each, against roughly eight years for a satellite when Marshall was at NASA. A Berlin facility will double manufacturing space, and an export-controlled Tanager sits behind black curtains.

22:47 Why Fleet Size Is Not What Matters

Building now mostly replenishes the fleet. Marshall says the measure that counts is area imaged per day, where Planet leads by about 100x, and notes that a Chinese system copied even the bird naming. The Owl upgrade moves from 3 meter to 1 meter daily resolution and adds optical links so data arrives within minutes, all without adding satellites.

24:30 Ion Engines, Space Debris and Methane Plumes

The roughly 50 cm telescope and the orbit altitude set how small an object Planet can see. The Pelicans run xenon ion engines, too weak to leave Earth but far more efficient once in space. On collisions, Marshall supports sharing space domain awareness data but says debris is the real risk. A Tanager poster shows a methane plume that the gas company shut off once Planet flagged it.

27:40 Testing Satellites to the Breaking Point

The test floor holds the three main tests: a hammer rig for the 200 g shock, a vibration table that simulates launch, and a thermal vacuum chamber (nicknamed the crockpot) for reaction wheels. The telescope is the most delicate part, built from special materials so it does not drift out of focus as it heats and cools. Every satellite that reached orbit worked, which Marshall says means Planet is not pushing hard enough.

32:19 Asking an AI What Changed at Area 51

Upstairs, Marshall shows the global imagery map, which scrubs through ten years of near-daily layers, and the LLM interface on top of it. Asked to compare Area 51 between 2021 and 2026, it zooms in and flags structural changes and planes that came and went. The tour closes with the “dove nest” of finished satellites waiting for launch and an RF chamber where radios are tested without reflections skewing the power readings.

Notable Quotes

“We’re talking about launching all 700 of our satellites cost less than a typical satellite at NASA. So, it’s about a 1,000x reduction in the cost.”
Will Marshall, on what miniaturization did to satellite economics

“So, we have a high barrier to entry, but our margins look like a software business. And so, it has the best of both worlds.”
Will Marshall, on Planet as a data business

“It didn’t need to cost a billion dollars. Phones work just fine in space.”
Will Marshall, on the question that started Planet

“The chips we buy, the systems we build.”
Will Marshall, on how vertically integrated Planet really is

“But generally, we’ve got utility out of 100% of our satellites, which I actually think is a problem.”
Will Marshall, on why a perfect record means the team is not taking enough risk

“So LLMs really only trained on the text of the internet. So they didn’t know much about the real world.”
Will Marshall, introducing Planet’s AI interface

“Now anyone can go and play with this because all they need to do is be able to speak and by the way you can speak in any language into it.”
Will Marshall, on opening geospatial data beyond imagery experts

Watch the full Planet Labs factory tour here.