Do Orbital Data Centers Make Sense?
What are the economics and physics behind constructing orbital data centers and will this make sense in the future?
In the future, is everything computer, or is everything in space? Perhaps its both. Everything is space computer.
Orbital Data Centers are as sci-fi as it gets and something that looks within reach, but its a somewhat polarizing topic between the general techno-optimists and the hard-nosed engineers. Future-oriented thought leaders like Eric Schmidt and Jeff Bezos have both made predictions that ODC’s will be a big thing in the future.
Yet the industrial empire most well-suited to carry out, and likely the most pragmatic in terms of balancing engineering trade-offs hasn’t pursued it at all. SpaceX owns all the launch and xAI builds data centers by far the fastest.
Then this post came out from NVidia prompted many to say “this is the top”
https://x.com/nvidia/status/1980757719809138854
Most especially critical were internet-favorite hard-nosed engineers like Andrew McCalip and Casey Handmar
Here is my best-attempt at breaking down the argument for, and against data centers, looking at economics, engineering considerations, first-principles physics, and then possible second and third order effects in the future that might undermine these assumptions.
Roughly:
- Arguments in favor: Land usage, permitting, energy costs, falling launch costs
- Engineering challenges: Deployment of orbital mega-structures, maintenance, station-keeping, bandwidth and therefore use-cases
- 2nd and 3rd oder: New and better model architectures, computing substrates, regulatory reform, and embodied learning.
So, does space computer make beep boop? Lets find out.
Trends Favoring Orbital Data Centers
The core idea behind orbital data centers comes down to a simple bottleneck: putting up new ones on the ground gets messy fast. You need land, permits, and more importantly lots of stable cheap power. Energy eats up the largest slice of ongoing costs—often 40% or more—and it gets worse if your company pushes for low-carbon operations to hit sustainability goals.
Starcloud’s take is that space sidesteps all that. Picture a 4-gigawatt facility up there, pulling constant sunlight with solar arrays spanning 16 square kilometers. Sounds cool, right? Wouldn’t it be even cooler if this prints massive amounts of cash? Would it?
Here are the arguments for why it might.
Surging demand for compute, especially AI workloads, projected to hit 100 gigawatts globally by 2030.
Solar panel prices dropping below $0.20 per watt, making orbital PV setups cheaper per kilowatt-hour than terrestrial grids in sunny spots.
Regulatory hurdles staying put—no quick fixes for zoning or transmission lines on Earth.
A big issue with this is the interconnection queue for new energy projects, which is growing horribly and means bringing new energy producing capacity to the grid is extremely difficult, to the extent new data centers often have their own behind-the-meter power supplies:
Launch costs halving every few years, now around $1,000 per kilogram with reusable rockets.
Project those out 10 years, and the total cost of ownership dips 90% below ground-based equivalents, per their models. Power becomes near-free after deployment, and cooling is, oddly enough, not such a big issue in space (more on that later). Avoid all the land and permitting issues, grid bottlenecks, etc.
Space computers here we come. Or do we?
Now for the current or near-term engineering complications before moving on to second and third order effects, things that would undermine the fundamental assumptions around data center growth, energy consumption, AI model architecture, and so on.
Engineering Difficulties
Deployment of Large Orbital Structures
Assembling large structures in orbit is extremely challenging, in fact its a whole field you can get PhD’s in solving brutal partial differential equations for things like control stability (in other words does your entire thing fly apart into tiny pieces).
The ISS measures 109 meters tip to tip, about the length of a football field, and required a decade of shuttle missions, robotic arms like Canadarm2, spacewalks, and coordinated international crews to piece together. An orbital data center would need to scale that up by a few orders of magnitude: a 4x4 kilometer solar array is about 3,000 times the area of solar panels on the ISS.
It would be the first true space mega-structure.
In practice this means deploying kilometer-long trusses, inflatable springs, unfolding panels, mechanically complex systems with many degrees of freedom which make controllability and mechanical stability a big problem. For example, the James Webb Space Telescope’s sunshield layers snagged during deployment in 2021, delaying weeks of troubleshooting, while ROSA (Roll-Out Solar Array) tests on the ISS showed hinges binding under microgravity loads. Close cable loop mechanisms, as explored in recent multibody dynamics models, aim to link panels for even rollout, but they introduce cable fatigue from repeated cycles and add points for micrometeoroid punctures.
Even when correctly assembled one of the big issues with large mechanical structures is the susceptibility to mechanical oscillations. In space no one can hear you scream, because there’s no air, and for the same reason there’s no fluid medium to damp vibrations, meaning any mechanical perturbations can dump energy into resonant modes that continuously builds up. Basically, in space everything is a high quality-factor mechanical resonator.
The energy for oscillations comes from what we would call a ‘noise source’ in engineering lingo - something that produces unwanted energy that messes up our system dynamics in the frequency spectrum. There are different kinds of noise sources with different amounts of energy at different frequencies. Now, a mechanical impact is a good approximation of what’s called the ‘impulse response’ which in an ideal circumstance has a flat frequency energy spectrum, meaning there is a constant amount of energy in each frequency interval. The bad news is that energy is linear in amplitude and goes like the square of frequency, so for a flat noise spectrum you have larger amplitudes at lower frequencies. Low frequency large-amplitude oscillations in a mechanical structure are really bad, because it means there is large displacements or strains, which increases the metal fatigue and so increases chances of failure and breakage. This isn’t an issue for smaller sized things because the low-frequency waves can’t effectively ‘fit’ inside the bounds of the structure, in other words, there isn’t a resonance at those frequencies. For large structures like 16 square kilometer solar array however this is a big issue.
There is another subtlety that makes this even worse - a lot of noise sources have whats called a ‘pink noise’ spectrum which goes like 1/f, meaning more energy is in lower frequencies to begin with. Examples of things that produce pink noise are stuff like mechanical pumps, like the kind that would be operating liquid-based heat exchanges from GPU racks to radiative cooling fins.
Vacuum offers no atmospheric drag to absorb energy, so inputs like thruster pulses for attitude control, uneven solar heating causing panel flex, or docking impacts from resupply craft propagate as waves. These couple into the array’s lowest bending modes, often below 0.1 Hz for ultra-large designs, where amplitudes grow and grow until fatigue cracks form in composite frames or welds fail.
Thin-film photovoltaics, stretched over lightweight spars, behave like a low-rigidity beam: the second moment of area shrinks with slimmer cross-sections, reducing resistance to out-of-plane bending by factors of 10 or more as span-to-thickness ratios hit 10,000:1.
This is already an issue for the ISS, 3,000 times smaller than a 4 GW orbital data center.
The ISS uses viscoelastic struts and tuned mass dampers to effectively absorb mechanical oscillatory energy at these resonant frequencies. For a super-large solar array you can think about de-coupling some of the mechanical structures from each other with things like flexible couplings, but then these add more degrees of freedom and make it harder to control overall.
Materials like carbon fiber are good at damping vibrations, and are stiff and light, which makes them popular for bicycle frames, but carbon fiber fails under large thermal swings that you would see in space environments where temperatures can swing by hundreds of degrees between sunlit and the shade.
So, even if you can assemble your space-based mega-structure, the fact that its shaped like a giant thin wafer isn’t good. It supports low frequency resonant modes, where amplitudes are higher and more energy is concentrated from the kinds of mechanical noise you’d get from thrusters and heat pumps and docking.
But wait, it gets worse:
Station-Keeping
Station-keeping just means keeping your satellite or orbital data center in the right place, facing the right way, but mostly it means combatting the effect of atmospheric drag that slows a satellites orbital velocity eventually leading it to burn up in the Earth’s atmopshere. Generally speaking the useful lifetime of a satellite is determined by the amount of propellant it can keep on board to fight atmospheric drag and point it in the right way. Things like reaction wheels work to re-orient and keep it pointed in the right direction, but can only spin up to such-and-such maximum RPM (torque is produced by changes in angular moment, so changes to RPM, not absolute RPM).
Extending satellite lifetime means getting extremely high specific-impulse thrusters, which is why satellites mostly use ionic propulsion: accelerating gas ions through an electric field.
But wait, there’s air in space? Yes, just little bits though.
In low Earth orbit, the atmosphere thins gradually rather than stopping at a hard boundary, following an exponential decay in density with altitude. Residual molecules at these heights create measurable forces, especially on extended surfaces like solar arrays. Starcloud’s design, with its 16 square kilometers of panels across solar and radiator wings, amplifies this effect. Their white paper specifies a dawn-dusk sun-synchronous orbit in low Earth orbit, typically around 600-800 kilometers altitude, to prioritize continuous sunlight exposure. This setup precesses the orbital plane once per year to track the sun’s seasonal shift, keeping the spacecraft near the day-night terminator for over 95% solar capacity factor and avoiding the thermal stresses of full eclipses.
Drag scales with the cross-sectional area facing the direction of travel. For the ISS at 400 kilometers, the baseline solar arrays total about 2,500 square meters per wing and contribute up to 95% of the drag profile when oriented face-on. Starcloud’s arrays, at 4,600 times that area, would see drag forces in the tens of newtons continuously, demanding frequent boosts to hold altitude. Density drops sharply above 500 kilometers—halving every 50-100 kilometers in nominal models—but solar activity flares can spike it temporarily, accelerating decay rates by 20-50%.
The ISS manages this through deliberate orientation shifts. During eclipse phases, which cover about 35% of each 90-minute orbit, it turns the arrays to “night glider” mode: edges aligned with the velocity, reducing the projected area and drag by 20-30%. This tactic conserves around 1,000 kilograms of xenon propellant annually for its ion thrusters, which deliver impulses at 3,000-4,000 seconds specific impulse for efficient delta-v.
Without it, reboosts every few weeks would triple the mass budget. Fortunately for Starclouds proposed sun-synchronous orbit, their solar panels are always edge-on to the direction of travel, but it is still a massive area, and rotating the panels on the ISS only reduced the panel drag by 20% - so there’s still drag.
Okay, so you can avoid some of the drag issues and thermal cycling in a higher altitude sun-synchronous orbit
But, that higher altitude means you need more delta-vee to launch into it.
Even worse, at those higher altitudes less drag means space junk last a lot longer. In fact, you’re basically entering the graveyard for orbital satellite debris, and your 16 square kilometer wafer-thin solar array basically starts to act like a giant broom, sweeping up all this space junk. Most of the space-junk isn’t in a sun-synchronous orbits and will be hitting the solar panels face-on at orbital velocities of several kilometers per second, 20-50 times faster than a rifle bullet. There are tracked objects, and then there are the millions of shards and fragments produced by Chinese anti-satellite missile tests. Thanks, China, for pissing in the pool of humanity’s future in space.
More space debris, higher launch costs, any other downsides to a high altitude orbit? Yes, actually - its also a lot harder to do any kind of maintenance operations.
But why would you need to do maintenance on an Orbital Data Center? Because under heavy AI workloads, GPUs fail constantly.
Maintenance
Until GPUs are made of diamonds they unfortunately don’t last forever, and even if they did they become obsolete after a few years.
NVIDIA’s H100 GPUs rate around 1.3 million hours—about 150 years—in light office use, but heavy training reduces that to 1-3 years in practice. Meta’s Llama 3 run on 16,384 H100s logged a cluster-wide interruption every three hours, with 30% tied to GPU or NVLink faults and another 17% to HBM3 memory issues—faulty silicon from early production batches.
From a friendly LLM:
Operators like Meta see 9% of H100s drop offline yearly under sustained loads, compounding to 25-27% over three years. Broader surveys peg data center GPUs at 0.1-2% per year in mixed duty, but AI pushes the high end, with early-hour burn-in failures spiking another 5-10% in fresh deployments.
A 1-gigawatt ODC equips about 1.4 million H100s at 700 watts apiece. At 9% annual failures, expect 350-400 units per day to fail, or 10% offline after 13 months and 20% after 26 months (this percentage would scale to a 4 gigawatt ODC because, hey, its a percentage).
On Earth this means sending in a technician that periodically goes in an replaces faulty or burned out units, which just the marginal cost of that unit, labor, perhaps lost revenue from downtime. Technicians swap cards in under an hour, recycling via specialized firms that refurbish 70-80% for resale. Costs run $500-2,000 per replacement, folded into 5-10% of op-ex.
Putting those in space makes this difficult.
Starship resupply burns $50-100 million per ton delivered, so ferrying a pallet of 100 GPUs matches the build price of a new pod—$5-10 million at scale. A single pod has tons of GPUs in it, but only some have failed. Do you replace the whole pod? Is there a way to eject individual GPUs from the pod and replace them? Hard to say.
Is every single mechanical operation under extreme scrutiny for the risk of introducing runaway mechanical oscillations into your ODC megastructure that might tear it apart because of runaway resonance? Yes.
The intrinsic failure rate of GPUs isn’t the only issue, however, because space is what we call a high radiation environment. This is bad for electronics. Quite bad. High energy particles from the solar wind or cosmic rays punch through the charge depletion regions in semiconductors turning them into regular conductors, so, they become useless for computing. Radiation-hardening and testing electronics sent to space is an entire industry unto its own. Unfortunately, the highest-performing chipsets use smaller die sizes, down to 4nm, and this means they are even more susceptible to radiation damage.
Basically, you don’t put high-performance computing assets in space because they’re the most sensitive to this kind of radiation damage.
From another friendly LLM:
Low Earth orbit delivers 1-10 krad of total ionizing dose yearly behind basic aluminum shielding, plus sporadic solar events that double it. Protons and heavy ions carve single-event upsets through silicon, flipping bits at 10^3-10^5 per day per chip without error-correcting codes—ECC catches most, but permanent latch-ups fry gates after 10-100 hits. For GPUs, this degrades thresholds by 2-5% annually, hastening electromigration in interconnects and boosting failure rates another 20-50% over ground baselines. Commercial parts lack rad-hard variants; space-qualified siblings like BAE’s RAD750 lag H100 perf by 10-20x. Shielding—tantalum or polyethylene—adds 200-500 kg per rack, hiking launch mass 15-25%, while active cooling for heat-trapped particles draws from the power budget.
There was an important point buried in all that generic and terribly written LLM-speak: in general radiation hardened electronics for space applications lag behind terrestrial computing performance by about 10 years, meaning they are about 10-20x worse in energy performance per operation.
So, perhaps energy is 10x cheaper in space, but you have to use chipsets that are 10x less efficient, and you end up with the same cost per compute, stacked on top of the mechanical nightmare of assembling and operating orbital megastructures along with the inability to maintain or service them.
Now, a lot of people balked at Orbital Data Centers right off the back because they think cooling it will be impossible. This isn’t as bad as you think, surprisingly.
(Space Daddy Elon chiming in on this)
Cooling and Power Budget
Compute hardware converts nearly all electrical input to thermal output. A single H100 GPU draws 700 watts under load, with over 99% emerging as heat through Joule losses in transistors and interconnects. An orbital data center with 1GW of computing power is really just a 1GW heater, and you need to dissipate all that heat into space or else everything melts.
Ground facilities cycle this through chillers or evaporative towers, where the coefficient of performance sits at 2-3: each watt of heat removal costs 0.3-0.5 watts of compressor or fan power, plus water for wet cooling that runs 1-2 liters per kilowatt-hour in arid sites.
In space there is no atmosphere to conduct heat into, so you have to just rely on radiative cooling. This doesn’t work as well on Earth because the background temperature is quite high, and the net power flux via radiative cooling follows the Stefan-Boltzmann law: everything is constantly radiating blackbody emissions, so the net power emitted is the temperature of your thing minus the temperature of whatever its facing it e*(T^4 - T_sink^4), where e is the emissivity of the surface. A high emissivity surface would be like black rubber, a low emissivity surface would look like a mirror (emissivity is 1-reflectivity by energy conservation).
With T_sink near zero in space, it simplifies to proportional to T^4, yielding effective COPs of 20-50 since the process draws no active power beyond transport.
Lets run the numbers for 1 gigawatt. Ideal blackbody at 300 kelvins—surface temperature around room conditions—emits 460 watts per square meter, so you need 2.18 square kilometers of radiator area.
Put a realistic emissivity of 0.85, and it climbs to 2.56 square kilometers. But if you increase the temperature of the radiating surface to 350 kelvins (77 degrees Celsius, within derated chip envelopes), and output jumps to 850 watts per square meter, shrinking area to 1.18 square kilometers.
At 400 kelvins, closer to 110 degrees, it hits 1,300 watts per square meter and 0.69 square kilometers. Higher temperatures accelerate degradation of electronics by making it easier for atoms in conductors to migrate around, leading to short circuits, open circuits, etc, and this doubles every 10 degrees above 85 Celsius.
Starcloud’s 5-gigawatt target scales this fivefold: 10-13 square kilometers at 300 kelvins, covered by their paired solar and radiator wings, each notionally 4 kilometers across for 16 square kilometers total deployment. Their white paper baselines passive emission with no active chillers, leaning on the orbit’s stability for steady flux.
Radiative cooling by itself draws no power, but you do need to run heat exchanger pumps that take heat from the GPUs out to the radiative fins. The ISS’s external loops, for example, handle 100 kilowatts with only ~1.5 kilowatts of electrical power consumption. It’s actually a beautifully cool system, then again everything NASA makes is Picasso-like engineering, even if expensive and behind schedule.
In terms of energy efficiency this space-based cooling beats the comparables on Earth, where to good approximation the surrounding environment is at 300 kelvin, meaning radiators would only dump 50-100 watts per square meter.
So the cooling issue isn’t so bad. I should mention a tweet here doing another analysis:
https://x.com/terrorproforma/status/1981318788806300058
Notably this plot here:
The really big issue besides deployment is, in my mind, the difficulty in getting large amounts of data to and from space-based assets. This is where some more fundamental physics comes into play.
Downlink / Uplink Make ODC’s Niche
Space platforms face hard limits on pulling results down, in fact overall the orbital satellite market is heavily downlink-bandwidth constrained.
Starlink by itself accounts for a significant fraction of all bandwidth to and from space and is adding something like 5 Tbps/week as they launch more and more Starlink V3’s.
Data centers need a lot of bandwidth. How can we get more?
Transmitting information means modulating electromagnetic waves. Inside a server, traces on printed circuit boards carry signals as voltage oscillations, but copper’s resistance caps frequencies around 50-100 gigahertz before losses swamp the signal, and crosstalk from capacitance or inductance also start to cut into the edges of digital signals (the edges have high frequency components, and inductance / capacitance don’t like that so much).
Fortunately there’s a better way of transmitting information through physical conduits: fiber-optic. FO operates at visible light frequencies in the 100-400 THz range, orders of magnitude higher than can be supported by copper traces, and is nearly lossless even over many kilometers.
In fact, the history of Data Center design in some ways can be seen as the gradual migration of fiber optic data connectivity eating up more and more of the total interconnect: first between data centers, then between server racks, then between GPUs in the racks, then between components in a single GPU unit, and now, between individual chips on single boards (I wrote about silicon photonics and why they are cool in my early twitter days here). Just to note: NVIDIA’s NVLink doesn’t use fiber optic because copper traces have lower latency, since there’s no need to convert between digital copper voltage signals to the analog signals that travel along fiber optic cables, and it communicates over relatively short distances.
So, how does this all apply to ODCs?
Orbital bandwidth primarily relies on radio frequencies, which are a few orders of magnitude lower than fiber-optic frequencies and so have comparably smaller bandwidth (they are also not multi-modal, but whatever). This is constrained by spectrum allocations and atmospheric interference.
Standard bands span C (4-8 gigahertz) for broad coverage, Ku (12-18 gigahertz) for TV relays, K (18-27 gigahertz) for mobile, and Ka (26-40 gigahertz) for peak speeds up to 10-20 gigabits per beam. Higher bands squeeze more bits per hertz through wider channels, but water vapor absorbs at 22 and 60 gigahertz, called ‘rain fade’ (something I once worked on for Ka-band satellites when I was an aspiring RF engineer) that drops links 20-50% in storms.
Free-space optical laser communications are the future of downlink and uplink for satellite communications, but they face even more serious issues with atmospheric interference and thus reliability. They are, however, already a frequent choice for crosslink: when satellites talk to each other. Starlink constellations use free space optical lasers to communicate with one another, letting them re-route data to other satellites when their Ka-band phased array antennas are blocked by atmosphere or for heavy data loads that need to be distributed across multiple Ka-band antennas.
Starlink’s optical crosslinks hit 100 gigabits for mesh routing, which is awesome. To get optical lasers through the atmosphere however means correcting for things like turbulence, which scatters photons. You can do this with cool things like deformable mirrors that correct 1,000 times per second, but it ain’t easy. NASA’s 2025 demos with partners like SEAKR and Tesat reached 200 gigabits over 1,000 kilometers, and startups like Skyloom target hybrid RF-laser terminals for 1 terabit by 2027.
Global orbital bandwidth capacity is around 500-800 terabits total in 2025, with Starlink claiming two-thirds at 300-500 terabits after 8,000+ satellites and weekly Falcon launches adding 2-5 terabits a week. Starship’s V3 batches, starting late 2025, add on 60 terabits per flight with 1-terabit sats, but this is distributed across the entire network to end-level consumers.
How does this orbital bandwidth compared to whats used in ground based data centers? Single stand-alone DC’s might have on the order of 5-20 Tbps to the external world, a multi-DC ‘region’ (as Meta and Google say) might have ~100 Tbps of bandwidth between DCs. Meta’s proposed new 10X backbone that intends to train large models across multiple DCs within a single geographic area extends this to around 100-400 Tbps.
A single lone orbital gigawatt-class data center could feasibly consume a vast quantity of the available orbital bandwidth currently deployed, depending on how its used.
The more practical use-cases are therefore:
Lone training runs that require less total bandwidth, on the order of 10 Tbps
Inference, which has lower bandwidth requirements overall
Insitu processing of Earth observation data, where satellites can route processing jobs via free space laser to the ODCs, and processes results are then beamed down to Earth. But, this isn’t a “1 GW ODC” problem.
So, bandwidth limitations come down to physics, which puts ODC’s into a particular niche. If the idea is to deploy multiple ODC’s as part of some constellation that support large distributed AI training runs, you may need a Starlink-sized constellation just to support the data backhaul between them.
As far as engineering considerations go, the bandwidth is likely more solvable, but will probably require things like dedicated auxiliary satellites with dedicated laser crosslink that routes to dedicated ground stations.
Everything I’ve written so far, however, rests on a few key assumptions that I think are unlikely to pan out.
AI training runs will continue to be massive by the time ODCs are practicable in 10 years, in the hundreds of millions of dollars each, an assumption based on the current LLM paradigm.
Energy costs will dominate the cost of AI training and inference, despite new computing substrates or transformer-specific chip architecture dropping energy costs by many orders of magnitude.
Regulatory burden on new data center construction is still a big barrier.
Lets look at these possible higher order impacts in turn:
Second and Third Order Disruptions
New Model Architectures are Needed.
There’s a growing sense of awareness within the AI community that LLM’s are, well, not that great. They are exceedingly data-inefficient and also compute-inefficient, and they don’t actually generalize, they just memorize text. Dwarkesh covers this well in a couple episodes with Karpathy and Sutton.
This is still enough to provide massive economic benefit to end-level users, but mean that operating a foundation model provider company is yet to be profitable despite massive growth in revenue overall.
I reached a similar conclusion more than a year ago with much simpler argument on an information-theoretic basis (called empiricism), which is simply, you can’t get smarter than your environment. For humans the environment is the physical world, but for LLMs training on text their environment is human knowledge. This is largely correct if you see the saturation in model competency and also scaling laws: logarithmic improvement for additional compute means its constantly declining marginal returns.
In short, the current paradigm of ‘making a giant model read the entire internet until its memorized everything’ is brutally inefficient and the model doesn’t actually learn generalizations of the information, it just memorizes it, and so inference is similarly brutally inefficient.
This means all the extrapolations on energy demand and compute required to scale AI to industrially relevant quantities to the entire world are likely to collapse, perhaps by orders of magnitude. Once again we are in the midst of a massive infrastructure over buildout, which has happened time and time again from canals, to trains, to the original internet providers. I’ve also written about this general trend, where AI fits into the pattern, and what happens next.
Even if we stick to LLMs, though, there are other trends that can undercut the energy demand-growth assumptions justifying ODCs:
New Computing Substrates and Specialized ASICs
Generative AI runs on deterministic gates emulating probabilistic processes, which is pretty energy inefficient. It takes many many gate operations to generate a random sample. But, physical systems like an electron trapped in a tunable potential well can generate random samples natively, up to a million times cheaper in energy. This is the premise behind my friend Guillaume Verdon’s company Extropic. I’ve written about this explaining its significance here.
Probabilistic computing natively handles uncertainty with p-bits—superpositions between 0 and 1—ideal for Monte Carlo sampling in diffusion models or Bayesian neural nets. Extropic’s thermodynamic chips, taped out in Q2 2025, use noise from thermal fluctuations to drive computations, yielding 10-100x speedups on generative tasks at 1/10th the energy of digital equivalents. Their alpha hardware, rolling out summer 2025, baselines on energy-based models that converge in fewer iterations than autoregressive LLMs.
But wait, there’s more.
In general there’s an energy floor to any kind of computation, Landaur’s limit, which current GPUs are still several orders of magnitude higher. But this only applies to irreversible computation. You can actually design reversible computation and implement it physically, either with custom logic or more simply, superconducting computers. This can also drastically undercut the energy assumptions we have extrapolating into the conventional LLM/GPU dominated future. Another friend of mine, Rudolfo Rosini, runs a company pursuing this and they’ve already demonstrated results (Vaire).
Even if we stick to LLMs there are reasons to believe we can massively beat the current energy costs by at least an order of magnitude.
Etched is a startup building application specific integrated circuits (ASICS) optimized for transformers, the architecture behind LLMs. They’ve claimed they’ve already achieved a 20x energy reduction compared to NVIDIA, which makes a GPUs, a much more general purpose chip architecture (oddly enough the founder of Etched, Gavin, was almost my roommate, but thought it was a bad deal until he talked to a bunch of other people and realized it was an awesome place to live, but by then he has missed out. lol.) Since then they’ve raised a ridiculous $420 million dollars to continue developing transformer ASICs.
More generally, I don’t actually think we’ll find AGI in a data-center. More likely it’ll be found the same way we found it- by navigating physical environments directly.
Embodied Robotics
I’ll keep this short because this article is already super long.
The rollout of humanoid robotics will probably be massive beyond our imagination in the coming years. It will be, in short, completely awesome. Just like Tesla has made massive strides in self-driving, not by simulating driving, but by driving, its likely humanoid robotics will achieve a much higher degree of generalized intelligence by learning directly by physical interactions with the environment.
This is likely a much bigger economic opportunity than a purely digital AI, since it does stuff in the physical world. It also means that a lot of AI will be running on edge-platforms in the robot itself.
This last point is stupidly simple and only deserves one sentence.
Regulatory and Grid Reforms
What’s easier, changing laws to permit more data centers, or building a 16 square kilometer solar array in space? Exactly.
Conclusion
I think ODC’s are cool, and perhaps inevitable in the far future, but I don’t think they’ll be relevant to the nearer term future of AI in the next 10 or 20 years. The big deal breakers for me:
Engineering complexity of massive mega-structure deployments, maintenance, and lifetime costs. Cooling issues are usually what people balk at first, but I don’t think these are actually the barrier.
Upsets in model architecture and energy efficiency of computing platforms, which enable more on-edge AI computing.
Useful AGI is more likely to come from physical world interactions that involve continuous learning (from different model architectures than transformers)
As a civilization, if we can’t update things like permitting and regulatory oversight to make it faster and easier to build needed infrastructure, we’re all screwed anyway. AI is a geopolitical competitive front, so I expect there will be lots of land allocated to conventional data centers in the coming years.
All of this adds up to: we are already likely in a massive data center over-build-out, which is a common pattern in new tech development that demands lots of new infrastructure. This cycle has happened more than a few times already.
So it goes!
All that being said, kudos to Starcloud for thinking big and reading the room on investment appetites. It’s likely there will be demand for orbital computing clusters, but they wont be 1 GW AI hyperscalers serving terrestrial applications, but rather serving in-space needs where optical laser links make them unconstrained by bandwidth.
Love this!
This article comes at the perfect time, especially with everyone talking about scaling AI, and it really builds on the forward-thinking topics you often explore. It's so fasing how you manage to break down something as sci-fi as orbital data centers into such clear economic and engineering considerations.