Try Free

Google Put an AI Data Center in Space — SpaceX Has a Bigger Plan!

States 2050 November 17, 2026 27m 4,186 words 1 views
▶ Watch original video

About this transcript: This is a full AI-generated transcript of Google Put an AI Data Center in Space — SpaceX Has a Bigger Plan! from States 2050, published November 17, 2026. The transcript contains 4,186 words with timestamps and was generated using Whisper AI.

"Right now, a small Google test satellite is circling the Earth. Onboard are four of the same AI chips Google uses in its giant server buildings. It is part of a project called Suncatcher, and the goal sounds wild. Move the computers behind AI off the planet. The whole machine runs on about one..."

[00:00:00] Speaker 1: Right now, a small Google test satellite is circling the Earth. Onboard are four of the same AI chips Google uses in its giant server buildings. It is part of a project called Suncatcher, and the goal sounds wild. Move the computers behind AI off the planet. The whole machine runs on about one kilowatt of sunlight, roughly what a microwave oven pulls. Its first real job in orbit was writing a poem. That poem came down to Earth over a cluster of icy islands near the North Pole, and Google is not alone up there. SpaceX has unveiled an AI satellite design wider than a Boeing 747, built to work like a full rack of computers in space. Elon Musk says orbit will be the cheapest place to run AI within about three years. A startup with just 25 people beat Google to orbit by almost a year, and it is now valued at $2.3 billion. This is not a drawing on a whiteboard. Sundar Pichai, Google's chief executive, announced the result himself, and Google says the satellite is working as planned. Even Nvidia, the chip giant, has put its own cash into one of the players. But one of Google's own energy bosses just said something about the cost of all this, and it changes the whole story. And the poem was only the first test. Enjoying this so far? The like button is right there, and subscribing takes a second. Both help more than you would think, and they keep these deep dives coming. But to see why anyone would launch a computer into space, we first have to look at what is happening down here on the ground. Earth is running low on power. In 2024, the world's data centers used about 415 terawatt hours of electricity. That is about 1.5% of all the power humans make, and it is climbing fast. The International Energy Agency expects that number to more than double by 2030, to about 945 terawatt hours. That would be roughly as much electricity as the whole country of Japan uses today. The main reason is AI. Every chatbot answer, every AI picture, and every new model needs huge rooms full of chips running day and night. The United States uses the biggest share, about 45% of all data center power on Earth. China comes next with about 25%. The Energy Agency says data centers could drive nearly half of all the growth in American electricity demand between now and 2030. A study from Lawrence Berkeley National Laboratory found that U.S. data centers used about 176 terawatt hours in 2023. That was 4.4% of the country's power. By 2028, it could reach as high as 12%, which is close to one out of every eight units of power in America going to computers. Power is only half of it. Chips get hot, and most data centers cool them with water. One estimate says U.S. data centers used about 17 billion gallons of water directly in 2023, mostly for cooling. That is enough to fill around 25,000 Olympic swimming pools, just to keep machines from overheating. The newest AI buildings are on a totally different scale. The biggest ones being built now can draw up to a gigawatt of power each. That is several thousand times what the solar panels on the International Space Station produce. One of them is a $16 billion campus for the Stargate project from Oracle and OpenAI. It sits in Saline Township, Michigan, and it is expected to need about 1.4 gigawatts. Local people pushed back hard, but the project was approved anyway. Then something interesting happened. At least 19 towns in Michigan passed pauses on new data centers. People did not want giant, power-hungry buildings next door, drinking up water and pushing up bills. That feeling is spreading across the country. One group called Data Center Watch counted 45 projects worth $68 billion that local opposition blocked or delayed. That was in just the second quarter of 2026. So the AI companies face a strange wall. They have the money. And they have the chips. What they are running short of is land, power lines, water, and friendly neighbors. Every new building means a new fight with a town council and a new deal with a power company. Some projects wait years just to get connected to the grid. That wall is exactly why Google started looking straight up. Space has no neighbors to upset, no water bills, and, in the right spot, almost no night. Blue Origin, the rocket company started by Jeff Bezos, makes the same pitch. It says moving hungry computers off Earth would take pressure off American towns. SpaceX says something similar, only louder. Space has its own set of problems, and they are big ones. But the idea only starts to make sense once you see the numbers somebody at Google sat down and worked out. Google's math for catching the sun. The sun puts out more than 100 trillion times as much energy as all the electricity humans make. Google put that fact near the top of a research project it announced on November 4th, 2025. The project was called Suncatcher, and the blog post was written by Travis Beals. He is a senior director at Google in a group called Paradigms of Intelligence. The paper behind it made a bold claim. It said, "Space may one day be the best place to grow AI computing," and that nothing in physics rules it out. The trick starts with a special orbit. Google picked a path called a dawn-dusk orbit. The satellite flies along the line where day meets night on Earth, so it almost never falls into shadow. That means the solar panels see sunlight nearly all the time. There are no clouds and no weather. Google says a panel there can make up to eight times more power than the same panel on the ground. It also needs far fewer heavy batteries, because it rarely has to save power for the dark. The next part is the clever one. A big AI job is never done by one chip. Thousands of chips split the work and talk to each other constantly, like a team passing notes at lightning speed. On Earth, they are wired together with cables. In space, Google wants to use lasers fired through empty space between satellites. Lasers work best over short distances, so the satellites would have to fly very close together. Google sketched an example cluster of 81 satellites flying about 650 kilometers up, which is around 400 miles. The whole group would fit inside a circle with a 1 kilometer radius. Neighbors would sit only about 100 to 200 meters apart. That is one or two football fields apart, while racing around the planet at full orbital speed. The gaps would shrink and grow as they circle. But Google expects only small engine burns to keep them in line. Can lasers really carry that much data? In a test on the ground, one pair of laser devices sent 800 gigabits per second each way. That is 1.6 terabits in total. And Google says tens of terabits per second should be possible. Beals put it this way: The bandwidth and the latency between TPUs really, really matters when you're trying to run a multi-rack workload. TPU stands for Tensor Processing Unit. It is Google's own AI chip and its answer to the graphics chip's NVIDIA cells. Then came the scary question: Space is full of radiation that can scramble computer chips. So Google took its Trillium chips to a particle accelerator and blasted them with a beam of protons, set up to copy life in low orbit. The weakest part was the memory. It only started acting oddly at a dose nearly three times what a shielded chip would face over a five-year mission. Nothing broke for good, even at about 20 times that mission dose. Later, Google ran the tests again, because the first setup gave more shielding than real orbit would. The chips showed slightly more errors this time. Even so, Google says they can still handle answering questions for a full five-year satellite life. Beals said the error rate for that kind of work is like one in a million. Training brand new models is another story, and radiation could be a real problem there. Pichai called the whole thing a moonshot that will need a lot of hard engineering. He also guessed that in a decade or so, it could look like a fairly normal way to build data centers. A decade is a long time though, and a paper is not a satellite. The beam tests could only copy space, never fully match it. The only way to know for sure was to stop testing on the ground and try the real thing. A poem beamed down from orbit. On October 1st, 2026, a SpaceX Falcon 9 lifted off from California, carrying more than 100 payloads at once. The mission was called Transporter 18, a rideshare trip where many small satellites share one rocket. Tucked in among them was Google's first Suncatcher test satellite. Google did not build the body from scratch. It teamed up with Planet, the company known for taking pictures of Earth, and used one of Planet's standard satellite frames. Inside sat four of Google's AI chips. That sounds like a lot, but there is a catch. Only one of the four can run at any moment, because the satellite cannot supply enough power or shed enough heat for more. Even that single chip runs for only about 15 to 20 minutes at a time. Then it has to stop and cool down before it can work again. Reports do not agree on the exact number of minutes, but the pattern is clear: work hard, then rest. Before any of that, the team spent about three weeks to a month just waking the satellite up. They powered on the main computer, then the trays holding the AI chips, then ran early machine learning tests. Beals was honest about the stakes. We've done testing on the ground, but you know, there's no test that's completely as good as the real thing. Then came the moment Google calls its "Hello World" test. During a pass over a ground station, engineers sent a short message up to the satellite. It was a prompt for Gemma 4, Google's Open AI model, which had been loaded on board before launch. The request was simple: write a haiku about being in space. The answer came back as the satellite flew over Svalbard, the Arctic islands in Norway that host a major satellite ground station. Here is the poem the satellite wrote: Blue Marble Below/Data Streams in Silent Grace: Starlight Feeds My Core. On October 30th, Pichai shared the news on X. He said Google had booted up the TPU, ran the prompt, and received our first complete response back on Earth. The team also asked for space limericks, and for reports on the satellite's own health. Now a word about the title of this video. Google calls this the first-ever AI response from a data center class TPU in orbit. The key words are "Data Center Class". The chips are the same kind Google puts in its giant server buildings. The satellite itself is not a data center. It is a single test box running one chip at a time on solar power. A big AI data center on the ground can draw up to a gigawatt, about a million times more power than this prototype. So what did the haiku actually prove? It showed the chips survived the shaking of launch. It showed they could boot up, run real AI software, and send an answer back down. That whole chain, from a message going up to a reply coming down, worked on the first real try in orbit. What it did not prove is that the chips can last for years in heat and radiation. Google has not yet shared any results on cooling or radiation from orbit. Those tests are still running. One of the big ones looks for bit flips, tiny errors that can quietly change data without setting off any alarms. That is the honest size of Google's first step. It is small, but it is real, and it is the first time Google's own AI chip has worked off the planet. It is also a step toward the bigger plan, because every lesson from this box feeds into the next satellites. Still, the headlines left out one important detail about who really got there first. The tiny startup that got there first. About 11 months before Google's Haiku, a 60kg satellite, about as heavy as an adult person, was already running a top AI chip in orbit. It belonged to Starcloud, a startup that did not exist until January 2024. Its satellite, Starcloud 1, reached orbit on November 2, 2025. It rode a SpaceX rideshare called Bandwagon 4 from Cape Canaveral. On board was one NVIDIA H100, the same graphics chip that fills data centers all over the world. NVIDIA said that was 100 times more GPU computing power than had ever been in orbit before. Then Starcloud did something nobody had done. In December 2025, it trained an AI model in space. The model was Nano-GPT, a small teaching model made by the well-known AI researcher Andrei Karpathy. Starcloud fed it the complete works of Shakespeare, and the chip learned from them while flying around Earth. The satellite also ran Google's Open Gemma model to answer questions. So, in a funny twist, Google's own AI software was working in orbit on someone else's satellite, almost a year before Google's chip arrived. That is why Google's claim is worded so carefully. Google was the first to run its own data center class, TPU, in orbit. Starcloud was the first to train a model on a data center chip in space. TechCrunch reported that Starcloud is still the only company known to be operating an H100 up there. The man behind it is Philip Johnston. He was born in Guildford, England in December 1986 and spent five years of his early childhood in South Africa. He studied applied math at Nottingham in Columbia. Then he earned an MBA at Wharton and a public policy degree at Harvard. He started out writing code for high-speed stock trading. Later, at the consulting firm McKinsey, he worked with national space agencies from 2019 to 2021. He even co-founded a non-profit that hopes to send a spacecraft to Alpha Centauri, the closest star system to our sun. Johnston started the company with Ezra Fylden and Addy Olteon under the name Lumen Orbit. A legal challenge from a company called Lumen Technologies forced a new name in early 2025, and Starcloud was born. The team had already gone through Y Combinator, the famous startup school, in the summer of 2024. The money came fast. In March 2026, Starcloud raised $170 million at a value of $1.1 billion. That made it the fastest Y Combinator company ever to pass $1 billion, just 17 months after the program. Then, in August 2026, it raised another $250 million, which put its value at $2.3 billion. Nvidia reportedly put in $25 million, and Cisco joined too. All of this went to a company with just 25 employees. Now Starcloud is building bigger. It has production lines in a 100,000-square-foot building in Woodinville, Washington, near where SpaceX and Amazon build satellites. Starcloud 2 should be two satellites with 8 kilowatts of computing each, flying in 2027. They are meant to serve paying customers, including U.S. government agencies. The company has also asked the FCC, the U.S. agency that licenses satellites, for permission to fly up to 88,000 of them. That is a request, not an approval. Johnston's biggest worry is not the chips. It is getting rides to space. "We can see what's coming," he said. "We're going to need to book an enormous amount of launch." He also admitted that if he could not book any SpaceX launches in 2029, that would be hard for the company. Falcon 9 is set to be retired around 2028, and most other big rockets are not flying often yet. So the whole race to put AI in orbit depends on a very short list of rockets. And that is where things get a lot bigger. SpaceX draws up a rack bigger than a 747. SpaceX's AI satellite is designed to make about 150 kilowatts of power at its peak. That is roughly 150 times the power of Google's little test satellite. SpaceX calls it AI-1, and Elon Musk showed it off in a video on X around June 8th and 9th, 2026. That was just days before SpaceX's stock market debut. His pitch was short. Think of it as a rack of compute in space. A rack is the tall metal cabinet of chips you see lined up in a data center, and Musk compared AI-1 to one rack of NVIDIA's GB300 chips. The size is the shocking part. With its solar wings open, AI-1 would stretch 70 meters across, about 230 feet. That is wider than a Boeing 747, and it would stand 20 meters tall. In the middle sits one compute module. On the sides are huge solar wings, and behind it is a folding liquid radiator. The radiator alone covers about 110 square meters, roughly the floor of a small apartment. It has backup pumps and armor against tiny space rocks. AI-1 would fly about 600 to 800 kilometers up. Every one of these numbers is a company claim, because AI-1 has not flown. Musk insists this is the easy part. He says it borrows a lot of technology from SpaceX's Starlink V3 internet satellites. The AI satellite is much simpler than a Starlink satellite, he said. He called it essentially a lot of solar cells, without the complex antennas Starlink needs. There's not some magic that's necessary that doesn't exist, he added. It's a very achievable goal. The chips inside can be swapped. The design supports Nvidia's GB300 and its newer Vera Rubin chips, and SpaceX plans a version for Google's TPUs too. That last part matters. Google is a major investor in SpaceX, and it also pays SpaceX for launches. Its own test satellite rode up on a SpaceX rocket. SpaceX is also planning a giant factory for this, called Gigasat, in Bastrop, Texas. The site covers more than 1,000 acres, with room for over 11 million square feet of buildings. It is meant to make its own wafers, solar cells, and finished satellites. Production is expected to begin by the end of 2027. So when does AI-1 actually fly? Gwynne Shotwell, SpaceX's chief operating officer, told CNBC, "I think we'll be launching the full AI-1 satellites late next year." First, she said, SpaceX will put some computing on Starlink satellites as early canary tests. That puts the first full AI-1 launches in late 2027, about a year from the day you are watching this. Musk's goals go much further. By the end of 2027, he wants to be adding 1 gigawatt of orbital computing per year. After that, he talks about 10 gigawatts, then 100, then a terawatt. In January 2026, SpaceX asked the FCC for permission to fly up to 1 million data center satellites. That is a filing, not a green light. On a podcast with Dwarkesh Patel in February 2026, Musk said space will be by far the cheapest place to run AI within 30 to 36 months. "It's not even close," he said. Patel guessed that 100 gigawatts in orbit would take about 10,000 Starship launches. But SpaceX's own stock market filing tells a much quieter story. It says orbital AI rests on unproven technologies, or technologies that do not exist. It warns the business may not achieve commercial viability. It even says the idea has not been proven at commercial scale, or at all. So it is the same company speaking in two very different voices. Musk is selling the dream, while the legal paperwork lists the risks. SpaceX has also not shared what each AI-1 costs, or named any paying customers in orbit. Those missing numbers are the real test of whether any of this can work, and the most honest answer about them came from a place few people expected. The bill behind the poem. Space data centers could cost up to 100 times as much as the ones on Earth. That warning did not come from a rival. It came from Amanda Peterson Corio, Google's own head of data center energy, at the information's WTF summit in Napa Valley. Her advice was blunt. "Let's explore all the terrestrial options before we go extraterrestrial." She also said the money does not work today, with enormous launch costs and maintenance challenges. "This was a remark at a conference, not a figure from a Google study. "Up to" means it is the high-end, but it still came from inside Google. "Why so expensive? Start with heat. Space feels cold, but it is a vacuum, so there is no air or water to carry heat away. The only way out is to send it off as invisible light through radiators, and that is much slower. That is why Google's chip has to rest after each burst. It is also why SpaceX needs that apartment-sized radiator. Then there are repairs, or the lack of them. SpaceX's filing says it does not anticipate servicing or repairing processors in space. A chip that dies up there stays dead, and whole satellites wear out and need replacing. Getting answers back down is its own puzzle. Alan George of the University of Pittsburgh said, "We don't have the convenient, easy way to get that data back to the ground." MIT professor Kerry Cahoy pointed to weight instead. "Launching enough mass to generate the power is huge." Google's paper does have an answer, and it is all about launch prices. If launch costs fall to about $200 per kilogram, about $90 a pound, the math starts to change. At that price, launching a satellite, spread over its life, could roughly match the power bill of a ground data center. The latest version of Google's paper hopes prices get close to that by 2035. The forecast leans on a trend. Because launch prices have dropped about 20% a year since SpaceX's first rocket, Falcon 1. The catch is hidden inside that trend. To keep prices sliding that fast, Starship would need to lift 370,000 tons into orbit over 10 years. At 200 tons a flight, that is about 1,800 launches, or 180 every year. That is not how many launches Google needs for its own satellites. It is how much flying the industry needs to push prices down to that target. So far, Starship has never flown more than five times in a single year. Musk has floated hourly Starship flights by 2029, and TechCrunch's reply was dry. Musk says a lot of things. Even Google's researchers stress that their math is not a full business plan. Others see a longer road. Jeff Bezos said in October 2025 that gigawatt-scale data centers in orbit are realistic in 10 to 20 years. He called two or three years a little ambitious. China has its own entry too, the 3-body computing constellation, which launched its first 12 satellites in May 2025. So the clock is now ticking toward late 2027. That is when Google hopes to fly two stronger satellites that talk to each other by laser. It is also when SpaceX hopes to send up its first full AI-1, and when Starcloud's next pair should fly. For the first time, we will see whether chips can work together in orbit, not just wake up one at a time. Cahoy had a warning for everyone in the race. It's usually the late entrants who finally figured out how to do it effectively. Which brings us back to that little poem over Svalbard. "Starlight feeds my core," the satellite wrote, and in a way it was right. The starlight really is the cheap part, free and nearly endless up there. Everything else is what Google, SpaceX, and Starcloud now have to pay for. That means the heat, the radiation, the rockets, and the repairs. Just 17 syllables proved a chip can think in orbit. The next few years will decide whether the sky can carry the bill. So a tiny Google satellite proved its chips can wake up and write poetry in orbit, while SpaceX and Starcloud raced to build something far bigger. The price is the part still waiting for an answer. If this one taught you something new, a like and a subscribe mean a lot. And tapping the bell lets you know when the next deep dive is ready. Do you think AI data centers will really move to space? Tell me in the comments.

Transcribe Any Video or Podcast — Free

Paste a URL and get a full AI-powered transcript in minutes. Try ScribeHawk →