24:02
This video is brought to you with IEEE Spectrum, an award-winning engineering magazine. Read their excellent article on the same subject, and subscribe to the magazine with a 20% discount with the links in the description Lately it feels like the entire world has lost its mind chasing the next revolution in computer science. Surging demand for Ai has impacted our lives in many ways. In just a matter of years Ai has grown to demand nearly 2% of total energy generation, with that number expected to double by 2030. Trade agreements have been torn apart and hastily rewritten as countries attempt to create a technological moat around the chips powering this revolution. Driving inflation and energy costs for normal citizens.
And nuclear energy stocks have exploded in growth as energy becomes a limiting factor for growth. Leading to massive data centers illegally powered by portable natural gas generators that destroy air quality in their immediate vicinity. But the craziest ideas have come from the billionaires. They say that reusable rockets will cut down launch costs so much that gigantic solar powered data centers can be placed in Earth Orbit. Free, unyielding solar energy with no land owners to negotiate with, no environmental policies standing in your way, and as we all know, in space no-one can hear you protest. In March 2026, Starcloud received 170 million dollars in funding on the back of this idea.
Billionaires will attempt to pull the rug over your eyes and convince you that this technology makes total sense, but reality is, this technology is dumb. And I'm going to explain to you exactly why. Let's start with the core of a data center. It is filled with server racks like these: the Nvidia NVL-72 sporting seventy-two GB200 GPUs paired with 36 CPUs cores and 17 terabytes of RAM. It provides up to 720,000 teraflops (FP8) of compute at the cost of about 120 kilowatts of electricity. A large house with air conditioning blasting may consume 2 kilowatts.
This one rack consumes the power of 60 suburban homes. Only server racks like these can handle the vast amounts of data and number crunching needed to run large language models. For comparison, this is the Versal System-on-Chip from AMD, found in the latest Starlink satellites. It is a thousand times less powerful while consuming a hundred times less power. That's the 'edge compute' in existing satellites and we can't just stack space grade hardware. This starlink satellite has a peak power generation of 5 kilowatts. No where close to the demands of a single Nvidia NLV72 server. Radically different
space hardware has been imagined in Starcloud's white paper to quote unquote "solve" this problem. Starcloud claims they can put 5 gigawatts' worth of these servers in space, with Nvidia server racks arriving in 40 megawatt containers that dock to the main station That's eight times more powerful than the biggest operation data center today. The goal here is likely to provide on orbit processing of earth observation satellites. Because beaming down terabytes of high resolution imagery and synthetic aperture radar data to the ground is becoming a bottleneck.
There is a useful niche there to process the images in space and just send down the relevant data. Or, if the billionaires are to be believed, it's a satellite running grok to answer your dumb questions on twitter. So let's start building out what a satellite capable of this would actually look like, and compare it to Starclouds figures stated in their white paper. Starting with that 5 gigawatt power requirement. If we had high efficiency panels generating 400 watts per square metre, the 5 gigawatt solar panel would have an area of 12 and a half million square meters. Here Starcloud's numbers
are not completely off. With their rendered concept using 16 million square meters. For comparison, this is nearly 5 thousand times the surface of the ISS solar panels, and the ISS took decades to assemble. But all that power doesn't just vanish, it becomes heat. In starcloud's case 5 gigawatts worth. And this satellite is orbiting in a vacuum with no way to convect that heat away. Without a way to radiate it away, that heat will accumulate and fry this satellite. The only way to get rid of it is with radiator panels, and we can calculate the radiator area with the Stefan-Boltzmann equation.
Emissivity is simply how good a surface is at emitting electromagnetic waves. A perfect blackbody has an emissivity of 1. This is why the SR-71 was black, to help dump as much heat generated from aerodynamic heating as possible. However a black surface is too good at absorbing sunlight in space, overwhelming its cooling capacity, so radiators are generally coated in a white coating like AZ93, which has an emissivity of 0.92, which matches Starcloud's white paper. The temperature of the radiator matters enormously too, because of this fourth power. Tiny increases in temperature lead to huge gains in the radiator performance. We'd want this to be as high as possible.
Starcloud wants to keep their servers at 20°C, the same temperature as a terrestrial server. Here is where I start questioning Starcloud's engineering. At 20 degrees we radiate 380 watts per square meter. Resulting in a radiator four kilometers tall and nearly a kilometer wide, taking up the middle of the satellite. Which is a problem, because inside those panels are tiny coolant channels that circulate a fluid between the server and the radiators, moving heat between them. The difficulty of pumping fluids across such a vast area back and forth is completely ignored by Starcloud's paper. Simply stating that "a workable design is possible without heat pumps".
Even if this could be powered passively this cooling rate would require immense fluid flows. If they use a coolant fluid like glycol, and have it enter the radiator at 35°C to return chilled to 5°C, then they'll need to circulate 68870 kilograms of it. Per second. That's emptying an Olympic swimming pool in 40 seconds, or pumping rates we'd see on a hydroelectric dam fed by gravity. For reference: the Space Shuttle's RS-25 engine has a mass flow rate of 514 kilograms per second. So this thing would need the equivalent pumping power of 134 rocket engine turbopumps. And somehow not require constant maintenance in
orbit. None of which their white paper accounts for. Nor does it account for the possibility of micrometeorites piercing one of these high mass flow rate lines, or any of the other delicate structures needed to launch something this massive. Like when a micro meteorite struck the ROSA solar panels of the ISS. That required a space walk to fix. They have also completely ignored another engineering problem. Massive flat panels hanging off a satellite are a nightmare for maintaining a stable orbit.
The ISS regularly needs to perform boost burns to compensate for the aerodynamic drag its solar panels create. But it gets even more difficult for something this massive as Earth's lumpy gravity keeps pulling its far-flung edges unevenly, adding to the forces pulling it off-course. We need to keep the station facing the Sun but that's easier said than done. Satellites typically use inertia wheels to control orientation. But inertia wheels are scaled according to the moment of inertia of the satellite, and moment of inertia is determined by the weight of the satellite
and its distance from the rotational axis. This satellite is massive and has hundreds of thousands of tonnes of fluid flowing to its far extremities, its moment of inertia will be accordingly massive, and of course the whitepaper glosses over this too. Just stating that the system is in development. Orbiting this way exposes one side of the radiators to the Sun constantly, and all that heat being absorbed from the sun will decrease the efficiency of our radiators. A more sensible design would use an X shape, with the solar panels face on to the sun and the radiators edge on. Their renders show everything just facing the sun.
It really seems like anyone with some renders and a whitepaper written by someone being gassed up by an overly agreeable AI can get VC funding these days. The next big challenge is degradation of materials in orbit. "Regular maintenance in space is difficult, redundancy has to be built in at launch, and cost estimates have to account for efficiency degradation over time." And space degrades things really quickly. Atomic oxygen from the ionosphere chemically attacks surfaces, and the van allen belts trap high energy particles that damage both radiators and solar panels over time.
There's ultraviolet radiation, solar flares and cosmic rays. There's orbital debris and there's good old fashioned bad luck. All of these factors attack exposed surfaces chemically and physically, degrading the output of solar panels and reducing the emissivity of radiators. The ISS has been testing many of these factors since 2001 and to date have tested over 1500 samples of different materials. Allowing them to identify the more resilient materials for space. [REF] Like AZ-93, the white coating on their radiators.
AZ93 is highly resistant to this, but not immune, its emissivity can drop from 0.92 to 0.90. [REF] However, for a crewed station, the degradation of the paint is a minor concern compared to the threat of orbital debris strikes. Something that has plagued the international space station for years, with countless stories of leaks that have resulted in entire coolant loops being drained. Ionizing radiation is even more complicated. Ionizing particles passing through satellites will burn out a transistor or flip a bit of information stored inside.
This would result in the mother of all ai hallucinations without a software constantly checking results. Hewlett Packard Edge Computer Servers aboard the ISS have to run three instances of the exact same calculation on three different nodes, to compare answers to weed out corrupted data. Tripling the power draw and tripling the mass needed to be launched, when they could just keep it on the ground and avoid all these problems. Space-rated chips are so much slower than commercial hardware because of the hardening needed. The nanometer scale transistors and the tiny voltage they hold are just no match for a cosmic ray.
The Starcloud white paper, once again, glosses over this issue, simply explaining that "logic devices are resistant to radiation" and that their dockable servers will have a larger volume to surface ratio than smaller satellites and therefore need comparatively less shielding. So a constellation of smaller data center satellites would struggle even more by their own admission. However, the most ridiculous part of this white paper may be the cost estimates. Which are over optimistic on the launch weights and launch costs. Starcloud claims that a single 100 tonne launch can deliver a 40 megawatt server container. That's 400 watts of compute per kilogram launched. That's way too high.
The Nvidia servers they reference weigh around 1400 kilograms alone. With a power draw of 120 kilowatts. They achieve 88 watts per kilogram. Starcloud's figures don't even align with the very hardware they referencing. Even SpaceX's TERAFAB project, infamous for historic levels of over-projection only assumes that space data center hardware would reach 100 watts per kilo. Starcloud numbers are four times higher than even Elon Musk's dream numbers. Then there's the solar panels. The top-end advanced lightweight ROSA arrays get 100 watts per kilogram in space. Designs that have never left
the lab that use micrometers-thick silicon panels can achieve 300 watts per kilogram. Starcloud's white paper quotes 1000 watts per kilogram. For 5 gigawatts, a more realistic weight would be 50 thousand tons instead of the 5 thousand tons they quote. Starcloud's white paper has no mass figures for their radiator. We calculate it would need to be 4 kilometers tall and 840 meters wide. Current generation panels weigh 10 kilograms per square meter. Resulting in a total mass of 33600 tonnes. With the most advanced, and very expensive carbon radiators achieving 4 kilograms per square metre, they may hit 13,440 tons.
Even ignoring the pumps, coolant, radiation shielding, fuel, inertia wheels, structures and other stuff, Starcloud's station exceeds a 113 million kilograms. More than an aircraft carrier sitting in orbit, more than six times the total mass launched into space in history. So how much would this cost to launch if Starship becomes operable at the scale SpaceX envisions? SpaceX's Starship does not have a public price menu but one of its first customers that we know of, Voyager Technologies, has signed a deal for a 90 million dollar delivery into orbit. That's a launch cost of $900 per kg.
That's 102 billion dollars in launch costs alone, but they don't quote $900, they quote $30 dollars per kilogram. I don't know where they got this figure from. That would barely even cover the cost of the fuel and launch operations. Reality seems to have become less valuable in the last few years. It's easier than ever to throw an idea at a large language model, and do some handwavey gestures in front of a cost comparison slide in your pitch meeting and get funded. These data centres are under immense competitive pressure. Data centres work their chips as hard as possible to get as much value out of them before they need to be upgraded. This is currently happening on a 2-4 years lifespan.
An AI data center in space, even with measures to protect against radiation and damage, can only be expected to have more failures and a shorter useful lifespan. It's a matter of fact that some AI companies are going to fail to return a profit for their investors, some analysts expect OpenAI to run out of runway as soon as 2028. But if that happens, at least all that spending leaves something useful here on earth. The computational power and power generation created to power it doesn't just vanish, it is still on earth and available for other uses.
Starcloud imagines an option to recover the server containers at the end of their life. But the cost of sending up an empty rocket to orbit doesn't come at much of a discount. If a serious fault happens, which may be as small as a power connection rattling loose during launch or a docking failure, the whole satellite may be written off immediately. So when, not if, a space data center fails, it leaves nothing behind. But what of the other concepts out in the world? This is just one early rushed concept to fundraise, and move on. In the ever evolving world of tech, first movers are being heavily rewarded.
Google "Suncatcher" corrects many of Starcloud's design mistakes. First they have split the system up into a constellation of smaller satellites. These data centres need to form a network to combine their computation power together. To do that they need communication, and laser networking is the new hot topic in ai datacentre expansion, with companies like Coherent Corp, a laser semiconductor manufacturer, gaining momentum. But the inverse square law makes communicating with lasers without the help of fibre optic cables difficult. Light spreads out from where it was emitted,
lowering in intensity with the square of the distance travelled. So laser signal power drops off drastically, even in space. Keeping the satellites near each other is beneficial and the more connections they have with neighbouring satellites the better. So Google calculated this unique bounded orbit that allows pods of 81 satellites to orbit together. But to do this they need to nail the initial conditions that allows this unique undulating disk of satellites to navigate the lumpy complicated gravity of earth without any heavy support structures binding them together. But even if they nail the initial conditions, Google wants this orbiting in a Sun Synchronous
orbit, surfing Earth's twilight and harvesting the perpetual solar energy, and that's a busy orbit, and if data center constellations start going up it's only going to get busier. These satellites will also need to avoid space debris. [REF] SpaceX recently reported to the FCC that the Starlink constellation performed a total of 300,000 collision avoidance manoeuvres in 2025. A problem that will only increase with more material being launched into orbit. And flying a 81 satellite bounded orbit means that one satellite moving could cause a knock
on effect that requires all 81 satellites to adjust. Multiplying the burden of a single collision avoidance. IEEE even published a paper on this exact problem last year. Other technologies may mature too, one technology IEEE Spectrum magazine reported on in their article was liquid-droplet radiators. Which proposes getting rid of the pipes and heavy radiator structure completely and instead spray a stream of coolant oil directly into the vacuum of space. With each droplet's surface now acting as a radiative surface, it drastically changes the physics of this problem.
So it's clear that a lot of really smart people are working on this problem, and a lot of not so smart people too. The economics of the challenge may improve and the business as a whole may become slightly less of a massive gamble, but it still just leaves the question. Why would anyone want computers in space? Space is becoming a new layer in the internet. With terabytes of data being passed around up there. With a lot of it being sensitive military data. When Iran was first struck there was a frenzy of satellites from countries all over the world,
all desperate to get eyes on the ground in a chaotic political environment. That's just one event that likely generated gigabytes of data. And it's all being fed into data centres for analysis. This is an incredibly valuable military asset and in the event of a war, whoever can process that information quicker is going to win. Ukraine recently reported that fully autonomous drones were used in the battlefield for the first time, patrolling an area and attacking without human guidance. Autonomous warfare, powered by satellite intelligence, is the future of warfare. If we are going to see space based data centres,
I believe intelligence processing will be the application that gets them off the ground. And as we have seen, data centres are already attractive targets in the event of a war. An economic and potentially soon a strategic target. Placing them in space with uninterruptible power makes them a near impossible target for most adversaries. And the military has famously deep pockets. This video is brought to you by IEEE Spectrum, an award-winning engineering magazine, from the institute of electrical and electronics engineers. In a world full of hype and bubbles, Spectrum is one of the few voices I trust to deliver clear
technical science journalism. This helps me stay on top of emerging technologies for this channel, but it has also helped me immensely with how I have invested my savings. Everyone loves investing in high growth tech stocks, and being ahead of the curve on where legitimate breakthroughs are happening can give you a small advantage. And it's truly great writing. Not everything related to ai is terrible. There are plenty of incredible cases where it provides real value to the world. IEEE covered one of them in their March issue. We have built massive particle
accelerators and detectors around the world trying to unravel the mysteries of particle physics, and we have generated Petabytes worth of data from them. But in recent years, everything fits the standard model we have created and because of that new discoveries have slowed down. Perhaps machine eyes can see a pattern that human eyes miss and find a clue to the mysteries the universe is still hiding from us. The very next article shows how these particle physics developments can help the world, with a new cancer treatment under development with CERN,
that uses particle acceleration technology to deliver incredibly precise single doses of ultrahigh power radiation to the exact point in the body it's needed, which researchers have shown can cause significantly less damage to healthy tissue. This isn't the layman magazine you see on a store shelf. When you subscribe, you're joining a global community of some of the most accomplished engineers on the planet. They of course have a digital only subscription, but there is so much to be said for having a physical magazine in your hands. It's tactile. It's collectible, but most importantly for me.
It's free from the constant notifications and distractions of working online. Getting a coffee in the morning and slowly ramping into my day away from screens and distractions with a magazine or book is my favourite way to get my brain kickstarted for a day of research and writing. IEEE are offering 20% off their usual subscription fee, which you can get by visiting spectrum.ieee.org/realengineering or by scanning the QR code on screen
