The Earth can't "afford" NVIDIA GPUs

[Introduction] The Earth can no longer afford GPUs. Microsoft can't plug its GPUs into the data centers, while NVIDIA's H100 GPUs are being sent directly into space.

The Earth really can't "afford" NVIDIA's GPUs anymore!

Today, I came across two news stories. One is that Microsoft has hoarded countless GPUs, but they "can't be plugged in."

The other is that NVIDIA's H100 GPUs are being launched into space to build a data center.

Behind these two events lies a profound problem:

Although GPUs have been manufactured, the infrastructure that supports them, such as power supply and cooling systems, has not kept up!

Let's start with the fact that Microsoft's GPUs are sitting idle in the warehouse.

Microsoft CEO Satya Nadella revealed a shocking fact during an interview with OpenAI's Sam Altman - Microsoft has hoarded a large number of GPUs.

However, there isn't enough power to make them run.

Another more practical reason is the lack of data centers where GPUs can be immediately "plugged in."

Nadella admitted: My problem now isn't a shortage of chips, but the lack of "warm shells" where they can be plugged in.

The so - called "Warm Shell" refers to the enclosure of a data center that has power supply and cooling capabilities.

To quickly understand this concept, in architecture, the opposite of a Warm Shell is a Cold Shell.

A Cold Shell refers to a building structure/enclosure that is basically complete, but with little or no system installation indoors.

A Warm Shell is in a more ready state, with basic building systems installed and operational, such as cooling systems, HVAC (Heating, Ventilation, and Air Conditioning), lighting, basic electricity/water/fire - fighting systems, etc.

The chip competition triggered by the AI boom is now constrained by the most traditional bottleneck - power.

The US power grid is facing unprecedented pressure, and tech giants are competing to deploy small nuclear reactors to save themselves.

Meanwhile, Altman also mentioned that in the future, there may be low - power consumption consumer devices that can "run GPT - 5 or GPT - 6 locally," which may completely disrupt the existing data center business model.

The Earth can't support them, so "send them to space"

Compared with Altman's proposed low - power devices, another news story offers a new idea.

NVIDIA, with the help of Starcloud's Starcloud - 1 satellite, sent the H100 to space!

On Sunday, November 2nd, NVIDIA sent the H100 GPU into space for the first time to test how a data center operates in orbit.

This GPU, equipped with 80GB of memory, is a hundred times more powerful than any computer that has ever flown in space.

Supporters believe this idea is reasonable:

In the vast expanse of space far from Earth, data centers won't occupy precious land, won't require as much energy and water for cooling, and won't emit greenhouse gases that contribute to global warming into the atmosphere.

This three - year mission will be launched on SpaceX's Bandwagon 4 Falcon 9 rocket.

The 60 - kilogram Starcloud - 1 satellite will orbit the Earth at a very low altitude of about 350 kilometers.

There, it will receive data from a fleet of synthetic aperture radar (SAR) Earth - observation satellites operated by the US company Capella, process it in real - time, and transmit messages to the ground.

Benefits of sending GPUs to space

Another major advantage of setting up a data center in space is that only a small portion of the data needs to be sent back.

Downlinking synthetic aperture radar (SAR) data has always been a big problem because of the extremely large amount of data.

But being able to process the data in orbit means we only need to downlink the "insights."

What are insights?

Insights might be information like a ship sailing at a certain speed in a certain direction at a certain location.

That's just a small packet of about 1 kilobyte of data, rather than the hundreds of gigabytes of raw data that would need to be downlinked.

Put simply, it means getting the algorithm closer to the data source, performing screening, fusion, and inference locally, and only sending back the high - value "information summary."

To put it more simply (but not necessarily precisely), it means processing all the data in outer space and only sending back the conclusions.

This method can better achieve low - latency response, significantly save bandwidth and energy consumption, improve resilience (sustainable operation in disconnection/disaster scenarios), and reduce the risk of sensitive data leakage.

Why send GPUs to space?

Different from the troubles of Microsoft CEO Satya Nadella, Starcloud is actively exploring this data center model.

Just like their company name, Star Cloud, a data center in space.

Of course, the main driving force behind this isn't to cool the GPUs.

It's the bottleneck of Earth's energy and resources:

Earth's data centers consume too much energy!

By 2030, the power consumption of global data centers is expected to equal that of the entire Japan.

At the same time, they consume a huge amount of cooling water every day (a 1 MW - class center ≈ the daily water consumption of 1000 people).

In comparison, space has natural advantages.

Unlimited solar energy: There is sunlight 24 hours a day in orbit, eliminating the need for battery energy storage.

Zero land occupation: No ground construction is required, and the ecosystem isn't damaged.

No greenhouse gas emissions: It doesn't rely on fossil fuels.

Ultimately, it's the exploding computing power demand of AI.

As AI models become larger (such as GPT, Claude, Gemini, etc.), energy and cooling costs are skyrocketing, and companies are in urgent need of new solutions.

Therefore, space data centers are seen as a long - term scalable solution.

By leveraging low - cost, continuous solar energy and avoiding land occupation and the use of fossil fuels, Starcloud's technology enables data centers to expand rapidly and sustainably. As the digital infrastructure develops, this helps achieve growth while protecting the Earth's climate and key natural resources.

Can space "dissipate heat"?

Another thing worth mentioning is that many people think GPUs are sent to space because it's too hot on Earth and space is better for heat dissipation.

Actually, that's not the case.

Space can dissipate heat, but it's very difficult.

There is almost no air in space, so fans or liquid circulation can't be used to carry away heat (this is called convective heat dissipation).

Convective heat dissipation refers to the process of "hot fluid (liquid or gas) moving and carrying heat from one place to another."

There is only radiative heat dissipation left:

Radiative heat dissipation is the process of "an object emitting heat in the form of waves through electromagnetic/infrared waves."

The equipment releases heat into outer space through infrared radiation.

The heat dissipation efficiency depends on the radiation area, material emissivity, and temperature.

Therefore, satellites or space - based GPUs need large - area radiators, and the design is extremely crucial.

In Starcloud's project, this part is specially strengthened:

They designed a dedicated thermal radiation system for the H100, using the high - temperature difference in a vacuum and thermally conductive materials to achieve heat dissipation.

Is it feasible to build data centers in space to save electricity, land, and water on Earth?

Johnston, the CEO and co - founder of Starcloud, said:

My expectation is that within ten years, almost all newly built data centers will be built in space.

The reason is simply the energy limitations we face on land.

Johnston said the only additional cost in space is the launch fee.

The launch cost can break even at about $500 per kilogram. Calculated per kilogram, after SpaceX's Starship is fully operational, the launch price is estimated to range from $150 to just $10.

As the Starship is put into use, we expect the launch cost to be even lower.

Starcloud is already planning its next mission, aiming to send a data center with ten times the computing power of Starcloud - 1 into space next year.

The Starcloud - 2 mission will be equipped with NVIDIA's Blackwell GPUs and several H100s.

Johnston said that this mission will provide 7 kilowatts of computing power and is expected to provide commercial services to customers, including Earth - observation satellite operators.

Microsoft's "lack of warm shells" and Starcloud sending the H100 into space are essentially the same problem.

No matter how powerful AI is and how large the computing power demand is, it can't break the laws of physics.

References:

https://www.starcloud.com/starcloud-2

https://spectrum.ieee.org/nvidia-h100-space

This article is from the WeChat public account "New Intelligence Yuan", edited by Dinghui. Republished by 36Kr with permission.