The launch of SpaceX's Starship may enable the development of orbital data centers, responding to terrestrial infrastructure challenges.
The anticipated launch of SpaceX's Starship this March may mark a significant turning point for the development of orbital data centers, according to insights from The Economist. This shift is prompted by ongoing challenges in terrestrial infrastructure, with Sightline Climate reporting that 30% to 50% of planned capacity for this year could face delays due to permitting issues, grid connections, and local opposition.
As demand for electricity surges, especially driven by artificial intelligence, some U.S. states have enacted construction moratoriums. One potential solution is to relocate servers to orbit, where solar energy is abundant. In November 2024, Starcloud successfully launched the Starcloud-1 satellite, which utilized a standard Nvidia H100 chip to effectively train the NanoGPT model using Shakespeare's works and to run Google's open model, Gemma.
Key Challenges
The primary hurdle for establishing orbital data centers is the high cost of launching payloads into space. Currently, sending one kilogram via the Falcon 9 rocket costs $3,400, while the Falcon Heavy charges $1,500 per kilogram. Additionally, two critical factors influence feasibility:
- The power-to-weight ratio, which measures the computing power available per kilogram of satellite; and
- The cost of satellite production, evaluated in dollars per watt of computing power.
These metrics are closely tied to the efficiency of solar panels and cooling systems employed in the satellites. Another significant concern is the effect of cosmic radiation on the reliability of AI chips in space, necessitating thorough assessments of these variables.
Andrew McCalip, an engineer with the startup Varda, has developed a web calculator to compare the costs of orbital versus terrestrial data centers of equivalent capacity. His analysis suggests that constructing a 1 GW terrestrial data center and operating it over five years would require an investment of $15.9 billion. In contrast, the orbital equivalent, factoring in various assumptions, could reach approximately $51.1 billion. This estimate does not include the cost of the AI chips, which could range from $15 billion to $30 billion for both setups.
Potential Solutions
Starcloud, established in 2024 to explore orbital computing, is gathering practical data from its satellite experiment. The initial results from Starcloud-1 have provided insights into chip performance in space, aiding in refining calculations for future projects. The power-to-weight ratio noted in preliminary calculations is based on Starlink satellites, but AI satellites could optimize this by using lighter components, such as laser links instead of heavy antennas.
Elon Musk has expressed optimism that achieving a power-to-weight ratio of 100 W/kg is possible, with future projections even suggesting 150 W/kg. Starcloud is setting a more conservative goal of 70 W/kg. Regarding satellite costs, Starlink's expenses are around $22/W, but AI satellites could be designed to be much cheaper due to the absence of costly communication hardware. Starcloud's CEO, Philip Johnston, aims for a target cost of under $5/W, excluding GPUs, which would bring the price of an orbital data center down to $16.7 billion—only slightly more than a terrestrial facility.
Nevertheless, launch costs remain a critical factor. If the projected launch cost of $500/kg can be reduced to $100-200/kg with the operational Starship, the orbital data center could cost as little as $12.1 billion, making it a more economical option than its terrestrial counterpart.
The first Starcloud satellite experienced overheating issues, but an upcoming launch will feature a radiator designed to improve cooling efficiency significantly. Early data also suggest that chip failure rates are lower than expected, with a decrease from 9% to 5% potentially leading to further cost reductions.
A pivotal moment for the industry is anticipated with Starship's 12th test flight in March, which could signal the onset of affordable launch capabilities and the viability of orbital data centers. Additionally, Musk's recent merger of SpaceX and xAI, now valued at $1.25 trillion, includes plans for an orbital data center comprising one million satellites.
