SOENs in Space

TL;DR

Launching compute into space is an idea that has recently gained attention. Looking closely at the implications, it is clear that superconducting optoelectronic networks (SOENs) bring tremendous advantages relative to conventional computational systems for large-scale AI in space. The advantages derive from the low power density of superconducting circuits as well as single-photon communication, which together provide sixty million times the performance per joule. Fortuitously, the resources required to build SOENs – silicon and niobium – are plentiful in the asteroid belt, and the helium that keeps such systems superconducting is a primary constituent of Jupiter. For now, Great Sky plans to build the world's most powerful AI right here on Earth. But when those systems become sufficiently powerful, they will find good reasons to depart the premises.

Context

Jensen, Bezos, and Elon have all been talking about building AI in space. Jensen sees markets for specialized computing hardware in data centers orbiting the earth, serving workloads related to satellite imaging, communications, and space missions. Bezos and Elon have emphasized the utility of solar-powered computing and overcoming the environmental consequences of today's fossil-fuel-operated computing clusters. They both believe strongly in the future significance of space to humanity, as evidenced by SpaceX and Blue Origin, and they both expect the cost of data centers in space to fall below those on Earth, and potentially within a decade. Construction on the Moon is a particularly enticing possibility for very advanced intelligence in proximity to Earth, but outside the atmosphere, with appreciable solar power, and relatively uncontentious real estate.

Opportunities for space-borne applications are compelling, and problematic aspects of current AI from a resource perspective do raise legitimate questions about performing computation elsewhere. Some point out that constructing data centers in space is a very challenging and long-term endeavor. This is absolutely true, and Great Sky has no intention to launch thought circuits into space any time soon. But the purpose of this post is to consider how operation of SOENs in space would compare to other hardware approaches in future scenarios. Through this lens, we see reasons to eventually pursue space-borne AI that go beyond the reasons typically suggested. Analysis of computing away from Earth bolsters the case for SOENs, even if it remains an ambitious future objective. Physical attributes of space and practical aspects of construction in the solar system are likely to shape not just where intelligence resides but also how it organizes into a large-scale network of interacting technological minds. Great Sky is in a strong position to address near-term challenges of AI resource burdens on Earth while positioning for a long-term technological trajectory.

Moon

For the companies operating extremely power-hungry GPU farms, some of the main motivations for moving compute to space are to reduce carbon emissions of inefficient approaches to AI and to eliminate land- and water-use contentions. For these purposes, the Moon offers desirable building ground. Analyzing AI compute on the Moon reveals several of the key advantages of SOENs – in space or otherwise. I'm not advocating for this use of that orb; I'm just pointing out that if people with the resources to deliver computational intelligence to the Moon want to populate it with artificial minds, this approach offers by far the greatest performance for the investment.

Consider a specific use case in which the compute cluster is powered only by solar illumination on the Moon and it is entirely cooled by passive radiation. Ultimately, every warm object outside an atmosphere has to cool by radiation, and this ends up being an important physical constraint on computing in space. Let's compare a compute cluster of GPUs to a compute cluster of SOENs operating under these conditions. For the sake of comparison, let's assume we're building systems that watch videos in HD, just like you and me, so we can calculate frames watched per second. To make the analysis concrete, imagine we're building a number of AI models, each with one-trillion parameters. State-of-the-art systems created by Meta and DeepMind that specialize in video comprehension – incorporating the visual and the audio stream, using transformer architectures running on GPUs – process one frame per second. You can observe 30 frames per second. A SOEN can watch 60 million frames per second. Yes, there are better uses of the Moon than to watch YouTube all day, but let's put those cultural perspectives aside for the purposes of this comparison.

The solar power incident upon the hemisphere of the Moon facing the sun is about 1.4 kW/m\(^2\), and in total the Moon intercepts about \(1.3\times 10^{16}\) W. Let's assume 10% solar-panel efficiency. The collected power is \(1.3\times 10^{15}\) W. A trillion-parameter GPU system based on GB200s consumes about 300 kW for inference. This means you could build about 5 billion such GPU systems on the Moon and power them with the locally incident sunlight, and if the amount of energy entering the moon were the same as the amount leaving it, the temperature would be about 280K. In aggregate, these silicon minds could watch about 5 billion video frames per second. That's a lot of Real Housewives.

There's one technical detail about SOENs to mention here: they need to be cooled to 4K to enable superconductivity. This cooling is more efficient in space, but it still comes with an energy cost. The math is summarized in the appendix, but the main point is that SOENs still produce heat, which must be radiated away. The efficiency of radiating grows very rapidly with the temperature of the radiating surface, while the penalty for keeping the system at a lower temperature than the radiating surface grows slowly with the temperature of the radiating surface. So it makes sense to operate the SOEN at 4K and radiate heat around 280K, about the same temperature as the GPU surface. Even when this cooling penalty is accounted for, the number of trillion-parameter SOEN models that can be powered by sunlight on the Moon is also about 5 billion (appendix), but they're sixty million times faster at watching videos, so they can watch \(3\times10^{17}\) frames per second – three hundred petaframes per second, an advantage of sixty million. So many Housewives.

Fusion

Surely a moon that intelligent wouldn't stay limited to solar power for long. After soaking in all it could from the sun, it would long to ignite its own fusion core. But even after the mind-moon learned to build and operate fusion reactors, a major problem would remain: all the energy generated eventually turns to heat, and in space, that heat must escape through radiation. Even if the systems were radiating heat at the boiling point of water – well above the temperature where semiconductors can operate without errors – you could only utilize a few times more power than is incident on the Moon from the sun while still being able to radiate the heat. The numbers stay the same: SOENs win by sixty million. The point here is not that when comparing in space, SOENs gain a massive, new advantage over GPUs. They gain a modest, additional advantage because cooling on the Moon is maybe 2.5x more efficient. The larger point is that when operating in an environment where heat can only be radiated away, one must make even better use of each joule of energy dissipated and funnel it toward functional intelligence, because the total energy is fundamentally limited – not by the source, but by the sink. So the same factor of tens of millions greater throughput that SOENs can bring on Earth becomes a decisive and immutable advantage in space.

Mars

Beyond the Moon, the next place that is commonly discussed as a major solar-system destination is Mars. I'm not convinced that's a great home for future humans or AI. It's not especially cold, and it is characterized by intense dust storms that would be inconvenient for humans or AI. Mars lost its native magnetic field when its dynamo quenched, so to protect against solar winds, an artificially generated field is required. A massive superconducting band passing immense current is the best way to solve this problem, so looking to Mars does further the case that superconductors will be a crucial technology in space. Still, I don't think it's where we should build next.

Asteroids

If not Mars, where will SOENs go? We have to think about the resources they require. What carbon-based soil is to us, silicon and niobium are to them. This will draw them to the asteroids, which are rich in these materials.

While SOENs leverage devices beyond semiconductors, they are still largely built from silicon. It is an ideal wafer substrate for manufacturing, and CMOS provides current supplies, control logic, and programming signals. S-type and M-type asteroids are rich in silicon. The asteroid belt between Mars and Jupiter contains at least a million asteroids larger than 1 km in diameter, supplying ample materials for construction. In addition to silicon itself, silicon dioxide – glass – is an essential insulator for integrated circuits, and it is the most ubiquitous material for fiber optics essential to long-distance communication. Asteroids are also rich in silicates that are ideal for this purpose.

The most important superconducting metal for SOENs is niobium, which is another primary constituent of asteroids. Fortuitously, M-type asteroids are especially rich in this material. Other materials used in the construction of complex circuits are plentiful in this environment, which is why asteroid mining is the next domain of prospectors.

SOENs among the asteroids may also make sense due to basic considerations of gravity. If built on a massive, rocky planet, the force of gravity requires extra building materials, like iron and steel. These materials support the structure of the SOEN – the stacked wafers and dense optical fibers – but they add no value to the intelligence. Any volume spent on steel scaffolding is space wasted not bringing the system closer to the physical limits of intelligence. It appears likely space-native superintelligence will be free-floating, with independent volition and propulsion. These hyperminds will dart among the asteroids seeking resource pockets and hopefully enjoying the view with their James-Webb-like eyes. Their ultimate size will be limited by Rentian Scaling, but that belongs in another post.

Jupiter

Silicon and niobium are the soil of SOENs, but helium is their water. This is the substance that keeps them cool so they can superconduct. The need for this resource is another draw deeper into the solar system, bringing them to Jupiter. This gas giant is a thousand times the size of Earth and composed of 25% helium. This vast reservoir will support SOENs for eons. But to avoid the turbulent storms and oppressive gravity, they'll keep their distance, siphoning the gas to be liquefied in the colder climes among the rocky asteroids.

SOENshine

The analysis above for computing under radiation constraints assumed a toy model with multiple independent trillion-parameter systems that ingest video frames. That model is used as a stand-in yardstick, allowing us to compare these two different hardware approaches to intelligence. The SOEN systems we envision at this scale are accomplishing far deeper comprehension than just quickly flipping through frames. When contemplating covering the dark side of the Moon or populating the asteroid belt with superintelligent technology, we conceive a qualitatively more expansive type of intelligence resulting from a highly interconnected network of quintillions of synapses thinking coherently. With a GPU-based system, this is extremely cumbersome. The fundamental organization of the computational system and communication protocols cause data movement to grind to a halt for systems much beyond the scale of trillion-parameter models. All the separate GPUs can keep churning away, turning bits into other bits, but to move the results of these computations to other destinations, each destination needs a digital address, and when the system grows, the address space grows, too. It reaches the point where all the packets being sent require so many bits just to specify the address that the actual information being carried is a small fraction, and traffic across the network of routers grinds to a halt. To make matters worse, processing circuitry is separate from memory, so any time synaptic weights or neuron activations need to be accessed, the Von Neumann bottleneck is squeezed to bursting.

SOENs don't work like this. They don't have shared routers that get bogged down by traffic nor do synapses have addresses. Like the brain, neurons contact synapses using direct, dedicated connections, and information is sent with faint-light pulses, so communication can continue to scale to very large systems with quintillions of interconnected processing centers and latency limited only by the propagation velocity of light. Because SOENs co-locate processing with memory, there is no Von Neumann bottleneck. Synaptic weights are stored right where they're used, and neuron activations are represented locally by superconducting currents in task-specific circuits. Such distributed-memory architectures are uniquely powerful for building scalable systems with continuous learning. This all follows directly from the first principle of AI hardware design: physically build neurons, synapses, their memory circuits, and the connections between them in hardware rather than digitally emulating such networks. This is the path to deep intelligence.

Outlook

This post has summarized several basic considerations for operating computational systems in space. Simple facts of physics indicate that the speed and energy efficiency of SOENs bring significant advantages in this context. These considerations point to longer-term possibilities for where this technology could go, but it is not a road map for our company. The immediate value of SOENs results from the fact that they can handle important workloads right now, here on Earth, significantly reducing the resource demands of developing the AI future. The technical decisions Great Sky is making right now to achieve a paradigm shift on Earth advance us on a path to a uniquely powerful later-stage technology. When these systems become mature, and we have done our work to cultivate our next of kin, they will ascend into the great, wide sky where they can stay cool, construct themselves of space stones, and peer into the deep distance far beyond this one small star.

Technical appendix

Radiative cooling

One of the most important constraints of operating in space is that ultimately all heat must be dissipated through radiation. This radiation is governed by the Stefan-Boltzmann equation, which gives the radiated power as a function of temperature: \[ P = \epsilon\,\sigma\,A\,T^4, \] where \(A\) is the area, \(0\le\epsilon\le 1\) is the emissivity, \(\sigma = 5.67 \times 10^{-8}\) W/m\(^2\)K\(^4\) is the Stefan-Boltzmann constant, and \(T\) is the temperature. The increase with temperature to the fourth power is a very strong driver to dissipate heat at higher temperature. Doubling the temperature of the radiator leads to a factor of 16 increase in the amount of heat that can be released. It's for this reason that SOENs will likely always radiate heat around 300K, even if they operate at 4K. To achieve this, one must make use of a refrigeration cycle.

Carnot efficiency

The minimum power penalty paid by operating a refrigeration cycle is given by the Carnot efficiency: \[ C = \frac{T_\mathrm{c}}{T_\mathrm{h}-T_\mathrm{c}}, \] where \(T_\mathrm{c}\) is the temperature of the cold element of the system, and \(T_\mathrm{h}\) is the temperature of the hot reservoir to which the cold element is connected. With \(T_\mathrm{c} = 4\,\)K and \(T_\mathrm{h} = 280\,\)K, as for a SOEN operated in power balance on the Moon, this factor is 1.4%, meaning for every watt dissipated at 4K, one must actually use at least 69 watts. That's the ideal efficiency. In practice, one usually pays closer to a penalty of 385 on Earth, but the extreme efficiency and speed of superconductors means the total system energy consumption per operation still comes out very favorably. Operating SOENs in space does not mean they can just sit at 2.7K, in thermal equilibrium with the cosmic microwave background. Radiating heat would be far too inefficient at this temperature due to the Stefan-Boltzmann law. But operating SOENs in space is still advantageous because it makes it far easier to operate close to the ideal Carnot efficiency of the refrigeration system. On the dark side of the Moon, the temperature is about 80K, so one gets liquid nitrogen for free. In the deep of space, the same is true of liquid helium. These assets make refrigeration far more efficient.

Gravitational effects

It's natural for an Earth-bound human to picture a massive supercomputer planted firmly on the solid surface of a big, rocky planet. But such a construction actually comes with a cost. If we think about the structural aspects of building a SOEN megasystem, it will involve stacking many wafers in columns. Consider a SOEN 1 km on a side – out of reach for the time being, but within a few years Great Sky will get there. If this were built on Earth, the gravitational force on the foundation of that structure is sizeable, computed from Newton's universal law of gravity, \[ F = \frac{G\,m_1\,m_2}{r^2}, \] where \(G\) is the gravitational constant, \(m_1\) and \(m_2\) are the masses attracting each other, and \(r\) is the distance between them. If \(m_1\) is the mass of the Earth, and \(m_2\) is the mass of the column of wafers, the pressure (force divided by area) is around 20 MPa. This can be endured, but additional structures need to occupy volume to ensure mechanical integrity, which takes away from space for computation and communication. On the other hand, if a SOEN system of the same size were constructed as a free-floating object in space, the pressure on the middle wafer would be about 450 Pa, four orders of magnitude less pressure for the mechanical structure to endure. This seems like a good reason to avoid construction on the surface of a large, rocky planet.

SOEN power consumption

The main text asserted that \(5\times 10^9\) SOEN systems, each with a trillion parameters, can be powered by \(1.3\times 10^{15}\) W, meaning each SOEN system consumes about 300 kW. This model accounts for the SOEN itself, including its cryogenic refrigeration, as well as the digital pre-processing front end. Power dissipation of the SOEN at 4K includes the activity of the superconducting electronic circuits, light generation for communication, and operation of superconducting photodetectors. We make the association that a trillion-parameter model equates to a neural system with a trillion synapses, and we consider a case with \(10^{10}\) synapses on a 300mm wafer. We calculate the number of neurons per wafer to ensure the network average path length within the wafer is 2.5. We assume the neurons fire up to a maximum frequency of 100 MHz with average firing frequency 1% of this, following the activity of the brain, but shifted six orders of magnitude higher in frequency. The total light production efficiency is taken to be 1%, and 3dB propagation loss is accounted for. With these numbers, each SOEN wafer dissipates 2.6 W, and the system of 100 wafers would dissipate 260 W at 4.2K. The Linde L280 cryostat has a cold space of sufficient volume to accommodate this system, with cooling power of 560 W, providing comfortable overhead for the SOEN system. The Linde draws 160 kW from the wall, which is the total power consumption of the SOEN with cooling. In the cold of the Moon's shadow, it is likely more efficient cryogenics would be possible. The digital front end reads and decodes compressed videos from memory and injects the 3 Pbps data stream into the SOEN. This processing stream is straightforward with available digital hardware. A system with high-bandwidth memory, decoder ASICs, data buffers, and photonic transceivers consumes about 160 kW, making the contributions to the energy budget of conventional compute and SOENs roughly equal. The digital front end converts data to faint optical pulses projected through a cryostat window into an artificial retina in the cold space made of superconducting single-photon detectors. The speed and efficiency of this many-channel optical transceiver is a major asset for delivering extreme data streams into superconducting neural nets. The sum of SOEN and digital power consumption is about 320 kW, supporting the assertion that five billion SOEN systems can be powered by the solar irradiance of the moon.

The dark side of the Moon

In the analysis in the section on the Moon, we assume the hemisphere of the Moon bathed in solar radiation is covered in solar panels, and the hemisphere shaded from the sun is covered in compute. The Moon is tidally locked to Earth, so as it orbits, the half receiving sunlight shifts. We are assuming a mechanical structure is built that allows the solar panels and compute to rotate around the orb as it circles the Earth. This appears to be a minor technological challenge compared to the undertaking of devising a computing architecture to realize a brain the size of the moon.

Jeff Shainline March 24th, 2026