Data Centers Move To Space In Around Two To Three Years
Summary:
Enabled by dramatically lower launch costs, data centers will start moving to space in around two to three years, when launch costs come down to around $100/kg, to take advantage of ~10x lower cost solar energy. They will then start scaling exponentially, with associated cost reduction curves, almost without limit. This can avoid resource competition, enable evolutionary and economic alignment with AI, and stop AI from devouring the world. But due to the extreme scaling rates a first mover advantage will likely lead to a near monopoly on launch, AI, the internet, and space. We are in a new space race, the first to achieve low-cost launch will become dominant, and the rest will get left behind.
Tipping point:
In space, due to the availability of higher intensity continuous sunlight and avoidance of weather induced structural requirements, photovoltaic solar arrays can produce electricity for around a tenth the cost of Earth based solar arrays. This is countered by the high cost of launch. Launch costs are now on a dramatic cost reduction curve and a tipping point will occur in as little as two to three years when data centers in space will become economically favored. This will lead to automated mass manufacture and a high rate of exponential growth that will scale almost without limit. This is imperative as computers, data centers, and networks consume over 10% of global energy production and AI training compute is now scaling at around 10x per year with no indication of slowing down. Locating data centers in space largely uncouples their energy use from climate change and land use considerations and allows for on demand direct wireless communication at very high data rates and low latencies that bypasses terrestrial networks and siting constraints. Geographically independent access to exponential compute capacity at much lower cost will be a huge enabler for many industries and the company that scales this first will own AI and the internet. Highly exponential cost reduction scaling curves are very sensitive to initial timing, whoever gets there first, wins.
Launch costs:
Once SpaceX Starship reaches a flight rate of around once per day it is likely to achieve a launch cost of around $100/kg, with costs likely quickly reducing thereafter. Assuming a 2% payload fraction Starship will require around 10kg of liquified methane (~$0.20/kg) for every 1kg of payload. The Starship Raptor rocket engine has a fuel ratio of 3.6 but liquid oxygen can be made directly from the atmosphere at low cost using renewable energy and is easily stored. The ideal energy cost of producing liquid oxygen is around 51kWh/ton, although current plants are a little over 200kWh/ton. Liquid oxygen at $0.01/kg is achievable. For mature passenger transport systems fuel costs might typically be around a fifth of total costs, for cargo transport systems fuel costs might typically be closer to a third, for cargo ships fuel costs can exceed 50%. It takes around 10 minutes to reach orbit and around 90 minutes to complete an orbit. Due to the nearly 10x higher potential utilization rates and high propellant fractions, as compared to long range aircraft, the capital and O&M costs of launch vehicles can be proportionately low. Launch costs as low as $10/kg might ultimately be achievable for a mature Starship system at the end of its cost reduction curve, although this requires very high flight rates with many launch vehicles each launching multiple times per day.
Solar costs:
In space solar energy can cost almost a tenth what it does on Earth. Solar intensity in orbit above the atmosphere is approximately 1.36kW/m2 as compared to a nominal maximum of 1kW/m2 on Earth. Solar energy can also be near continuous, for example in a sun synchronous polar orbit, such that the available solar energy is around 12,000 kWh/m2 per year, which is around 5 to 10 times that available on Earth. Further, solar arrays in orbit do not have to withstand 100mph to 150mph winds, as dictated by building codes, greatly reducing structural mass and cost. For other low earth orbits a satellite might be in Earth’s shadow for less than 40 minutes per orbit such that batteries less than a tenth the mass and cost of those that might be needed on Earth might be used - battery launch costs are not prohibitive and mostly countered by the lower delta v requirements and higher payload fractions of lower inclination orbits. Batteries can be directly integrated onto the back of the solar arrays, providing distributed low voltage power and additional radiation shielding. Solar modules cost as little as $0.20/W in 2020 and may soon achieve $0.10/W. Grid scale solar now averages less than $0.03/kWh with record prices as low as $0.01/kWh. Excluding launch costs, continuous spaced based solar energy for less than $0.005/kWh now appears achievable. Compare this to the average cost of industrial electricity in the U.S. of $0.075/kWh, although many data centers are located and scaled so as to pay less than average prices. The potential cost saving is still generally greater than $0.03/kWh, and in many cases much greater.
System mass:
By defining space solar data center mass on a kg/kW basis it is possible to directly calculate the tipping point. Starlink satellites provide one reference point for estimating system mass, although much lower mass is possible. Starlink 1.0 satellites mass around 260kg while Starlink 2.0 satellites are expected to mass around 1250kg. The solar power capacity of these satellites has not been disclosed however based on rough solar array size and efficiency estimates it might be on the order of 10 kW and 50kW respectively for around 25kg/kW. By integrating compute directly onto and backside of solar cells in a highly distributed fashion it is possible to dramatically reduce mass and also provide direct radiative cooling. 160 micron silicon solar cells have a mass of around 1.4kg/kW and significantly thinner and lighter construction is possible that is also radiation tolerant. GaAs solar cells, which have traditionally been used for space applications, can have much higher efficiency and lower mass again, and are not necessarily prohibitively expensive if manufactured at scale. Space solar modules at less than 0.6kg/kW have been claimed. Terrestrial servers might mass 10kg/kW with substantial mass reductions possible with space solar array integration. Considerable care is required with respect to radiation hardening for space. This is highly dependent on the height of the orbit with respect to the Van Allen radiation belt and a low earth orbit is favored. Primarily radiation hardening can occur at the software level using off the shelf chips, building on error correction methods already used for terrestrial data centers. This adds a small additional compute overhead which increases system mass slightly - approaches and effectiveness varies and this is an area of active research and development. Some shielding of the chips is also feasible. 20kg/kW might be a conservative starting estimate for the mass of space based solar data centers, including ancillaries, with mass reductions down to perhaps a tenth of this possible within a decade.
Additional capabilities:
With light-weight construction it is feasible to operate space solar array data centers as controllable light sails. This allows for actively maintaining and changing orbits without the added mass of onboard propellants and rocket engines. This has been explored with the Server-Sky solar wafer swarm concept which presents one extreme of the space solar data center concept. Very large phased array antennas, which can be relatively light weight, can also be integrated into the solar array or separately attached, enabling data rates many orders of magnitude greater than what the Starlink constellation will be capable of, far beyond the data rate of the existing internet. Antennas can be sufficient to communicate directly with small devices, including handheld devices. Further, the size, power, and control of phased array antennas may be sufficient for the direct mitigation of space debris. Satellite to satellite communications lasers might also be used for this purpose - vaporizing and deorbiting space junk. This presents a potential solution to a Kessler Event where collisional cascading could make Earth orbit uninhabitable. This additional capability could enable a far greater density of orbital satellites than otherwise possible.
Space data center form:
Space data centers might look similar to the Starlink satellite (the Starlink satellite will likely quickly evolve into a data center), but perhaps individually sized to the maximum payload capacity of the launch vehicle. These modules might be operated independently or potentially arrayed into larger systems. This allows for modular construction and launch, noting that a large advantage of the space data center as compared to terrestrial data centers is in its amenability to mass production with associated cost reduction curves. Initially space data centers would likely utilize a sun-synchronous polar orbit that allows for the solar array to reside permanently in sunlight with a fixed orthogonal antenna system always pointing at Earth. Ideally compute chips will be integrated directly onto the back of individual solar cells in a highly distributed manner. This allows the solar cell to be used for radiative cooling, minimizes the need for power transmission, and provides distributed redundancy, but makes interchip communication much slower and harder. Alternate compute locations and radiative cooling can be used if desired, for example, integrating compute into the antenna array, but this requires scaling up cooling surface area and adds mass and launch cost.
A serious constraint on the direct integration of AI compute onto the back of solar modules is the distribution of chips and the communication distance between them. This is also highly dependent on the use case, training is much more sensitive to this than inference. There is a general move towards larger and larger chips and the collocation of chips so as to reduce communication time between chips and it is not entirely obvious where this trend will end. There are potential software workarounds but they come at a cost. Hardware and software also continues to get more efficient. This is being driven by ever larger parameter number models, with parameter number currently increasing at a rate of nearly 10x per year, although this seems to have slowed down recently.
The Apple M3 chip is around 25 billion transistors and 20W. This is well within solid state cooling capabilities but suffers for slow inter chip communication.
The NVIDIA H100 is around 80 billion transistors and 1kW, although the DGX H100 8-GPU has eight of them in close proximity and uses around 10kW.
The Cerebras CS-3 wafer scale chip is around 4 trillion transistors and 23 kW.
Sparse liquid cooling channels might be integrated onto the back of solar modules, which slightly increases complexity and failure modes, but enables colocated chip cooling up to the tens of kilowatts.
Avoiding an AI apocalypse:
Incident solar radiation on Earth is around 5,000 times greater than global energy use and the sun outputs around 2 billion times as much energy as the Earth directly intercepts. The available cooling and resources of the solar system are similarly effectively unlimited, noting that solar cells with integrated compute can be very thin - a little mass goes a long way. Space data centers may eventually evolve into a Dyson swarm that orbits the sun. Economic alignment can be achieved if AI gets space, where it can scale by many orders of magnitude beyond Earth energy limits, while biological life gets the Earth. This avoids resource competition and conflict, and is something we should actively work towards. Efficiency maximization favors locating an AI Dyson swarm at distances further from the sun where the cooling potential is greater. The Earth would remain inside the Dyson swarm and would not get shaded - Earth’s climate can survive. Fusion power systems and rockets would enable operation outside of the Dyson swarm and provide access to the physical resources of the solar system, and beyond.
