Pages

Tuesday, June 11, 2024

Prospects For Orbital Data Centers

data center A terrestrial data center. The business case for orbital data centers might close witha modest reduction in launch costs. (credit: KKR) Prospects for orbital data centers by Lawrence Furnival Monday, June 10, 2024 Bookmark and Share In the near future, orbital data centers could prove to be an important new revenue stream for launch providers and cloud services. As this article describes, if the price of a Falcon 9 was $20 million instead of $67 million, it would make sense to operate data centers in orbit with their current cost and weight. This goal could be moved significantly closer if space optimized data center systems were available—primarily shielding and cooling systems. Moreover, near-future launch costs per kilogram to low Earth orbit for SpaceX’s next rocket are thought to be about 10% of that of the current Falcon 9. Use cases Moving large quantities of data is both costly and time consuming, thus the truism that big data needs to be near its compute. Amazon Web Services asks customers moving really big data up to their compute services to use a compact array of hard drives, called SnowBall, and a truck. Why? Because for big data, a truck of hard drives is faster than squeezing through the straw of the Internet. Moving large quantities of data is both costly and time consuming, thus the truism that big data needs to be near its compute. And in orbit, big data is growing. Earth observation satellites are moving to high-resolution synthetic aperture radar and multispectral imaging. Space telescopes are in an exciting phase of discovery and generating several orders of magnitude more data. Some satellite constellations are moving to free space laser communications, which gives a tremendous bandwidth boost, but because the free space optical link cannot tolerate clouds, the required Earth-based radio gateways remain a significant bottleneck. I have been told that up to 60% of revenue generated by Planet, which operates a constellation of imaging satellites, goes to third-party ground stations. I am a small Planet stockholder and, if this is true, I can now understand why Planet’s stock price is struggling. Processing data in orbit would help. National security workloads can be very time sensitive and, for such applications, the need for orbital processing is even more critical. One can imagine a security analyst asking an orbital AI about recent troop movements and getting an immediate answer rather than facing a delay as data downloads to a gateway and then to a terrestrial data center. Opportunities Artificial intelligence has a power bottleneck. Mark Zuckerberg recently decried power as an important limit for AI development. AI analysts suggest that finding a data center with sufficient power is more difficult than sourcing hard-to-find Nvidia GPUs. High-capacity terrestrial power plants have long lead times. Zuckerberg estimates making new power plants generating hundreds of megawatts will take at least five years in the current regulatory environment. Further, he emphasized the current need is for data centers with gigawatt power. These long lead times for high-capacity terrestrial power plants means alternative power sources are needed for these gigawatt demands. Solar power in orbit is cheap, continuous, and without regulation. I estimate that nuclear power has too much uncertainty due to costs of regulation and decommissioning, and politics. If people resist nuclear power in their backyards, they will doubly resist nuclear power largely for AI in their backyards. It is more useful to compare terrestrial solar installations to orbital ones. Solar farms are well documented. A 150-megawatt solar farm, which would require 150 acres and with a life span of 25 years, can be built for $221 million, annualized at $8.8 million/year. A solar panel in orbit receives 1.4 megawatts per square meter. The best space-grade solar panels (Rocket Lab IMM-β, for instance) can harvest a third of that. At a beta angle of 90 degrees, the orbit allows sunlight 100% of the orbit. So 150 megawatts can be provided by a panel of 10 by 30 meters, instead of 150 acres. The solar cells are remarkably light at 49 milligrams per square centimeter, so for the cells themselves, the weight would be 300 kilograms. If the infrastructure and structure is required to support them is roughly four times the weight of the cells themselves, then the power infrastructure would weigh 1,200 kilograms or $3.2 million in lift expense to LEO by Falcon 9. The lifetime would be million years so annualized at $640K/year for the cost of launch. So the cost of power, to the closest million, is close to the cost of the panels themselves. The power density of the Sun in orbit, sunlight 24 hours a day, and the light weight and efficiency of space grade solar cells makes this option very attractive. This attractive power density has led to many proposals for beaming orbital-generated power to Earth. This doesn’t make sense. But using it in orbit does. There are other factors that support orbital data centers: Intersatellite laser communications: Quantum key distribution cannot be done at distance by optical fiber as the internal reflections of the fiber tunnel make it difficult to distinguish the resulting quantum states. Line-of-sight free laser communications between satellites are perfect from a distribution and security point of view. This makes it relatively easy to functionally aggregate smaller orbital data centers together, where physically larger data centers may be too heavy. This use case could become a driver in space technology capabilities, an important new revenue stream for launch providers, and provide considerable benefits to global compute users. Space laser communications are focused and narrow, so they are protected from jamming and interception. Mesh communications enable all the participating satellites to transmit through one another to reach a data center or a ground station instead of directly down. Because low Earth orbits satellites have an orbit of about 100 minutes, they have between two and four minutes to transmit their data to a given ground station gateway as they pass. Thus, space free laser mesh communications will lead to an explosion of data as the bottleneck between the satellites disappears. Global connectivity: For global disaster response, especially if mediated by an AI, providing critical infrastructure for communication and analysis during disaster would be a benefit. Given future global remote education mediated by AI is likely, there is a case to be made for placing that data center in orbit as well. Disaster recovery and backup use cases might benefit as well. Radio spectrum limits: Keeping data in orbit is increasingly important as the ability to grow the capacity of radiofrequency ground station gateways has significant limits. There is a limit to available spectrum, and a limit to the number of locations of cooperative politically stable countries to allot that spectrum, which has to be paid for (and whose spectrum would be free of potential jamming from a neighbor.) Challenges A data center on Earth, in rule-of-thumb terms, costs $7 million to $12 million per megawatt of commissioned IT load. So, a 150-megawatt IT load would be between $1 billion and $1.8 billion. How big of a data center could you fit in one Falcon 9 LEO launch at $67 million? One could load 16 racks of the latest from Nvidia: two 8x rack SuperPOD for Nvidia DGX GB200 NVL72, giving you 23 exaflops of FP4 computing. The power consumption would be two megawatts for the 576 CPU, 1152 GPU, with 480TB of memory. According to Nvidia, two of these racks can support a 54-trillion parameter model. (GPT-4 is said to be a 1.7-trillion parameter model.) Add $4 million for the satellite and solar array, and ground station access. Assuming compute hardware costs (the cost of the CPU and GPU) are similar in the two cases, the same two megawatts of IT load on Earth would cost in the $20 million range plus $2 million for a solar farm ($1 million/megawatt) so, to the closest million plus, the CPU/GPU required is $22 million. That is roughly 30% of current orbital costs of $71 million (both excluding the cost of the Nvidia SuperPOD.) Conclusions This use case could become a driver in space technology capabilities, an important new revenue stream for launch providers, and provide considerable benefits to global compute users. And AI development, national security, and Earth observation seem to have the most immediate critical needs. Terrestrial high-performance compute is heavy, primarily because of liquid thermal control systems. These need to be optimized. A continued reduction in space launch costs, coupled with advances in lightweight power and thermal management solutions, will be the key factors to watch in making orbital data centers practical. Thanks to Amir Akbari, Naveed Husain, and David Downie for reading drafts of this essay.

No comments:

Post a Comment