Is Earth Running Out of Power? Google and NVIDIA Start Moving Compute to Space

Key Points

Orbital compute is emerging as a solution to Earth’s power limits for AI.

Power limits on Earth are throttling AI growth — grid strain, rising demand

Power shortages have become a key bottleneck for building AI data centers.

FTI Consulting (Fùshìgāo zīxún 富事高咨询) forecasts that U.S. data‑center energy demand could almost double by 2027.

The surge of large‑scale compute requests is straining utilities and grid capacity.

With electricity on Earth becoming a limiting factor, several Silicon Valley players are exploring an obvious alternative: putting compute where the sun never sets—into orbit.

Big tech plans: Suncatcher, SpaceX, Amazon — who’s saying what

Google (Gǔgē 谷歌) CEO Sundar Pichai (Sāngdá’ěr Pícháyī 桑达尔·皮查伊) said on social channels that Google has launched a project called “Suncatcher” to study scalable machine‑learning compute systems in space.

The company plans early test flights—two satellites targeted for launch in early 2027—to experiment with running next‑generation TPUs (tensor processing units) off‑planet.

Other major tech leaders are talking about the same idea.

SpaceX (SpaceX) CEO Elon Musk (Āilóng Mǎsīkè 埃隆·马斯克) said SpaceX will build data centers in space by scaling up Starlink V3 satellites equipped with high‑speed laser links.

Amazon (Yàmǎxùn 亚马逊) founder Jeff Bezos (Jiéfu Bèisuǒsī 杰夫·贝索斯) has also said that over the next 10–20 years humans should be able to construct gigawatt‑scale data centers in orbit.

Early hardware in orbit: TPUs and H100 GPUs

Some AI chips will be among the first to go to space.

Google is partnering with satellite imagery company Planet Labs on the two early satellites planned for 2027 to evaluate large‑scale orbital compute clusters and TPU operation in orbit.

Meanwhile, space‑compute company Starcloud plans to launch a satellite this November carrying an NVIDIA (Yīngwěidá 英伟达) H100 GPU—the first advanced data‑center GPU to be deployed in outer space, according to the company.

The Starcloud‑1 satellite weighs about 60 kg (roughly the size of a small refrigerator) and the company claims it will deliver up to 100× the GPU compute performance of previous in‑space compute installations.

Starcloud says it ultimately plans a 5‑gigawatt (5 GW) orbital data center equipped with ultra‑large solar arrays and radiator/cooling panels stretching on the order of kilometers.

Why space? Abundant, continuous solar power and simplified cooling

Energy is the main attraction.

In orbit, solar arrays receive sunlight nearly continuously.

No night cycle, no clouds, no rain.

That produces significantly more energy per unit area than on Earth.

Google’s analysis notes that, on the right orbit, solar panels can generate roughly eight times the output compared with typical ground installations.

Continuous production reduces dependence on batteries.

Starcloud argues its orbital data center would not need water for cooling and would dramatically reduce lifecycle CO₂ emissions compared with terrestrial data centers.

The company claims a roughly 10× reduction across the lifecycle, with the primary environmental cost being launches.

Jeff Bezos has echoed the point that plentiful orbital solar could allow orbital AI clusters to outperform equivalent Earth facilities because they can draw uninterrupted solar power.

Economics and engineering: falling launch costs, remaining technical hurdles

High launch cost has long been the biggest barrier to large orbital systems.

Google’s recent paper (summarizing historical trends and launch‑price forecasts) estimates that, by the mid‑2030s, launch costs might fall below $200 per kilogram (¥1,460 RMB ($200 USD) per kg using an exchange approximation).

At that threshold, start‑up and operating costs for an orbital data center could become comparable with the energy costs of a similarly sized ground data center.

But substantial engineering challenges remain.

Early tests indicate promising radiation tolerance: Google reports that its next‑generation TPU survived simulated low‑Earth‑orbit radiation in particle‑accelerator testing.

Still, issues such as thermal management, on‑orbit systems reliability, communication latency and bandwidth, maintenance/repair, long‑term solar‑panel lifetime, and orbital debris management are significant hurdles.

Google’s forthcoming 2027 satellite launches are designed to validate models and evaluate TPU hardware performance in orbit.

Starcloud’s near‑term H100 mission will similarly test GPU performance and operations in space.

Outlook: a decade for meaningful scale?

Industry proponents suggest orbital compute could become a mainstream option within 10–20 years if launch prices continue to fall and technical challenges are solved.

Some executives predict that in roughly a decade many new data centers might be constructed in space.

Others caution that costs, logistics, and space‑specific risks mean the path will be gradual and technically demanding.

What this means for investors, founders, and builders

• Investors: look for bets on launch cost reduction, high‑efficiency solar, thermal radiators, and in‑orbit repair/robotics.
• Founders: focus on modular, radiation‑hardened compute units and communication stacks that tolerate latency and bandwidth constraints.
• Engineers: prioritize thermal design, radiation testing, and life‑cycle maintenance strategies for long‑duration missions.
• Marketers: tell the story around sustainability gains (the claimed 10× CO₂ reduction) and continuous power from orbital solar.

Quick take — what to watch next

• Google’s Suncatcher two‑satellite tests in early 2027.
• Starcloud’s H100 mission launching this November and its performance claims.
• Trends in launch pricing toward the ¥1,460 RMB ($200 USD) per kg threshold.
• Radiation and thermal test results from TPU and GPU hardware in orbit.
• Regulatory and debris‑management rules that will shape feasible scale and timelines.

Orbital compute could reshape where we run AI, but the decade ahead will decide whether orbit becomes a niche testbed or the next major frontier for compute.

References

• Earth’s power running out? Google, NVIDIA begin moving compute to space – Yicai (第一财经)
• Suncatcher and Google’s space TPU research – Google
• Starcloud announces H100 GPU satellite and orbital data‑center plans – Starcloud
• Global data‑center energy outlook and analysis – FTI Consulting (富事高咨询)