Subscribe to Our Digest
In a first-of-its-kind experiment, Google plans to place a prototype AI computing system in orbit as the tech giant explores new approaches to powering and cooling data centers in space.
By Filippo Di Giovanni, Contributing Editor at Data Centre Digest
In the spasmodic quest to build and deploy data centers that can sustain the stellar growth of AI, diverse solutions are being explored to solve the most fundamental problem: where and how to generate all the electricity required to run these mega factories? Moreover, cooling data centers efficiently to accommodate processors and accelerators, each now consuming 1 kW or more of power, implies massive use of water, even if modern techniques, like liquid and free-air cooling, can mitigate excessive water usage and thermal waste.
As data center farms increasingly rely on these precious resources and directly weigh on our environment, communities in the US and other countries begin to express concern and make their voices heard, in an effort to influence political decisions.
Training and running increasingly complex AI models requires enormous parallel-processing capacity based on GPUs, TPUs and other accelerators. If 10 kW of power was once sufficient to supply a single rack, power densities exceeding 100 kW for dedicated AI hardware are today necessary, while some AI infrastructures are quickly moving to several hundred kW, or even 1 MW by the end of this decade.
AI facilities that once consumed tens of megawatts are increasingly being designed for hundreds of megawatts, while gigawatt-scale campuses are being planned.
Looking at the history of industry, this is the first time energy demand is rising so rapidly to require new energy sources to complement current supplies. In an effort to cut carbon footprints, data center operators and hyperscalers are signing long-term contracts to secure renewable energy supply. Meta, the parent company of popular social media and messaging platforms, has contracted 1,800 MW of solar and wind power but this is not an isolated case.
Nordic countries with abundant natural green resources, like hydro-rich Norway, are becoming attractive sites to locate data centers. A wider spectrum of diversified renewable resources (solar, wind, biogas-powered fuel cells, hydro) improves resilience by decoupling from conventional grid connections and fossil fuel-based backup generators. Yet this effort may not be enough to keep up with fast-growing energy demand.
Here comes Google’s initiative to move data centers into space. Recently, at a dedicated laboratory near San Francisco, a team of engineers inspected a refrigerator-sized satellite equipped with solar panels and computing chips after subjecting it to mechanical stress designed to simulate the intense vibrations and unusual acceleration it will experience during launch and itsjourney into orbit.
The test is part of an ambitious project, dubbed Suncatcher, aimed at exploring the feasibility of operating data centers beyond Earth. The satellite will be launched aboard a SpaceX Falcon 9–a fittingly ironic choice, given that SpaceX is itself pursuing the idea of orbiting data centers after expanding into AI through xAI. Google is effectively relying on the rocket of a companylikely to become a strong competitor in the nascent market.
As terrestrial resources, whether land or energy, become increasingly scarce, high-tech visionaries, including Elon Musk, Jeff Bezos, Sam Altman, and others, have turned their gaze toward space as a potential long-term solution. Building orbital data centers could provide plentiful solar energy: according to Google, up to eight times more solar power than on Earth.
At the same time, the vacuum of space offers an environment in which heat can ultimately be rejected only by radiation. In fact, a vacuum provides no air convection to cool electronics, meaning that heat must instead be transported through heat pipes or other thermal interfaces to radiators and rejected via radiation, something Google engineers are now testing on Earth and willfurther investigate in space.
Some start-ups, such as Starcloud, have already tested early demonstrators consisting of a few AI accelerators (including an NVIDIA H100 GPU) hosted on orbiting computing platforms. Others, such as Axiom Space, are evaluating computing modules for deployment near the International Space Station.
Despite the enormous challenges–including exposure to high-energy cosmic particles, communication latency that can limit real time inference, and potentially prohibitive launch costs, at least initially – orbital data centers demonstrate how far the industry is prepared to push the boundaries of computing infrastructure.
Google’s satellite–named MVP–will host four Trillium TPUs (Tensor Processing Units), representing the approximate compute capacity of a single high-performance rack server installed in terrestrial data centers. MVP is also equipped with a solar cell array to generate 1 kilowatt of power. The four TPUs are intended to demonstrate that the basic compute hardware and its cooling system can operate in orbit; Google says the spacecraft will run the hardware only intermittently to let chips cool off.
A rocket ascent into low Earth orbit (LEO) takes approximately 10 minutes. During this time, the whole rocket is subject to extremely high vibration and heavy acceleration loads up to 10 times the force of gravity, or 10 g. Due to mechanical resonance and high-frequency shock during launch, onboard TPU chips can experience even greater forces as high as 50 to 100 g. Once the TPU reach final orbit, beyond the shielding effect of the atmosphere, the level of external radiation becomes another big challenge.
Solar particles and cosmic rays, “can wreak havoc on electronics, so our team tested TPUs in a proton beam facility at UC Davis’s Crocker Nuclear Laboratory while running AI workloads” reads one dedicated Google blog 1. According to the company, chances are high that during the test errors – like a bitflip – arise, altering binary logic and affecting compute workloads. “Initial results have shown that our Trillium TPUs hold up remarkably well and can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission “ explains the blog.
Although MVP has been designed to survive for six years, the main goal is to operate it for one year while executing simple AI queries.
Future design iterations will aim to accommodate dozens of TPU chips, while a swarm of satellites will orbit Earth and communicate with one another using precise laser beams. Large-scale machine learning workloads require distributing tasks across different accelerators with high-bandwidth, low-latency connections. Ensuring performance comparable to terrestrial data centers requires links between satellites that support tens of terabits per second. Google plans to test this capability by sending two satellites into orbit in 2027.
Testing space as a possible location for scalable AI compute will take time and significant effort. This first trial is intended to demonstrate that the hardware can handle space’s very different conditions, identify potential failures, and implement corrective actions in future missions. Whether this vision can ultimately compete with terrestrial data centers will depend not only on computing performance, but also on the economics, launch infrastructure, and ability to manage heat in space.
1) Learn about Google’s Project Suncatcher to put ML infrastructure in space