Space-based compute: Earth’s Power Not Enough? Google, NVIDIA Start Moving Compute to Space

Key Points

Why tech companies are eyeing orbit for space-based compute

Space-based compute is gaining serious attention because Earth’s power and cooling limits are creating a hard cap on large-scale AI growth.

Rising energy demand for large-scale AI data centers is creating a hard physical limit on Earth.

Consulting firm FTI Consulting forecasts that U.S. data-center energy demand could nearly double by 2027.

Utility companies and grid capacity are already strained by requests to power ever-larger computing facilities.

Faced with terrestrial power limits, several Silicon Valley firms are exploring the same basic idea: put parts of the compute stack in space where sunlight and cooling advantages are far greater.

Who’s building what in orbital compute

Several major players have publicly signaled concrete plans or research into orbital compute.

• Google (Gǔgē 谷歌) — CEO Sundar Pichai (Sāngdá’ěr Pícháyī 桑达尔·皮查伊) announced a program called “Suncatcher,” an effort to investigate scalable machine-learning compute systems in space.

  Google says it will test TPU (TPU, Zhāngliàng Chǔlǐqì 张量处理器) hardware and related systems in orbit to better understand performance, radiation tolerance, thermal management and control.

• SpaceX — CEO Elon Musk (Āilóng Mǎsīkè 埃隆·马斯克) has said SpaceX intends to build data-center capabilities in orbit, leveraging and scaling its Starlink constellation with high-speed laser links.

• Jeff Bezos (Jiéfū Bèi suǒ sī 杰夫·贝索斯) and Amazon have also publicly suggested that, within 10–20 years, it will be possible to build gigawatt-scale data centers in space using continuous sunlight.

• Planet Labs (Xīngxíng Shíyànshì 行星实验室) — Google is reportedly partnering with the satellite imagery company to launch two test satellites in early 2027 to experiment with TPU-based ML workloads and validate system models.

• Starcloud — a space-computing company, plans to launch a satellite containing an NVIDIA (Yīngwěidá 英伟达) H100 GPU later this year.

  The planned Starcloud-1 satellite weighs about 60 kg and is roughly the size of a small refrigerator.

  Starcloud claims that satellite will deliver up to 100× the GPU capability of previous orbital compute facilities.

  The company says it ultimately aims to construct a 5-gigawatt orbital data center equipped with very large solar arrays and radiative cooling structures spanning kilometers.

What makes space attractive for orbital data centers

There are three often-cited technical advantages for moving compute to orbit.

1. Abundant, continuous solar energy:

   In the right orbit, solar irradiance is higher and uninterrupted by weather, clouds or night cycles.

   Google noted solar energy in space can be many times more productive than on the ground.

   Proponents say continuous sunlight reduces reliance on batteries and backup generators.

2. Easy radiative cooling:

   In vacuum, large radiators can dump heat to cold space without water-based cooling loops.

   This can potentially simplify thermal design for high-density compute racks and reduce complex chill-water systems used on Earth.

3. Reduced carbon footprint (potentially):

   Proponents argue that, across the facility lifecycle, orbital systems powered directly by solar arrays could emit far less CO2 than equivalent Earth-bound data centers—once launch and manufacturing emissions are amortized.

Key technical steps and near-term tests for space-based compute

Industry teams emphasize the immediate next step is hardware validation.

Google says its next-generation TPUs have survived early radiation tests that simulated low-Earth-orbit particle environments.

The 2027 dual-satellite launch planned by Google and Planet Labs will stress-test TPU hardware, communications links, power generation and thermal controls in real orbital conditions.

Separately, Starcloud’s planned satellite carrying an NVIDIA H100 GPU would mark one of the first times a modern data-center GPU has been tested in orbit.

If that launch succeeds, it will provide real-world measurements of compute density, performance-per-watt, radiation tolerance and cooling strategies for GPU-based AI training or inference workloads.

Costs, economics and the critical threshold for orbital compute

Historically, high launch costs have been the main barrier to large-scale orbital systems.

Google’s analysis suggests that if launch prices fall below roughly $200 USD per kilogram (≈ ¥1,440 RMB per kilogram), then the cost to launch and operate large orbital compute systems could approach the energy cost profile of an equivalently sized ground data center.

That price point would make the economics of orbital compute more plausible, though many engineering and operational risks would remain.

Major engineering and operational challenges

• Thermal management at extreme scale:

  Moving heat from multi-megawatt or gigawatt compute clusters in vacuum requires very large radiators and careful thermal architecture.

• Radiation and reliability:

  Although some TPU and GPU hardware demonstrates promising radiation tolerance, long-term reliability for sustained AI training workloads is unproven.

• Communications and latency:

  High-throughput, low-latency links (laser or RF) are necessary to move training data and model updates between ground and orbit, and to serve inference with acceptable performance.

• Launch, manufacturing and end-of-life:

  The environmental and financial costs of launches, plus orbital debris and decommissioning, are important considerations for any lifecycle assessment.

Outlook: timeline and scale for orbital compute

Optimistic industry voices suggest that within a decade, space could become a practical expansion layer for AI compute.

Google calls space “a potential best place to scale AI compute.”

Some founders predict most new data-center capacity could be built in orbit in the long run.

Others are more cautious, noting that substantial engineering, logistics and regulatory hurdles must be solved before orbital data centers become routine.

Bottom line on space-based compute

Space-based compute is moving rapidly from science fiction toward operational experiments.

In the next few years, expect to see targeted satellite launches carrying ML accelerators (TPUs, GPUs) and focused tests on power generation, heat rejection and radiation tolerance.

Whether orbital data centers ever become a mainstream complement to Earth-based infrastructure will depend on continued reductions in launch cost, breakthroughs in orbital thermal and systems engineering, and clear economic advantages once lifecycle emissions and launch footprints are included in the calculation.

Space-based compute could be the expansion layer AI needs if economics and engineering align.

References

• Earth’s power not enough? Google, NVIDIA begin moving compute to space – 第一财经 (Yicai)
• Suncatcher: space-based ML compute research – Google AI Blog – Google
• NVIDIA data-center GPUs and space compute initiatives – NVIDIA Newsroom – NVIDIA