They Are Building an Earth Computer, Board by Board

Data centers going up now are not random warehouses full of boxes. They are being laid out like giant motherboards. The racks are the chips. The aisles, fiber trays, and power bars are the traces. Cooling is the heatsink. The whole room is meant to act as one machine.

That pattern does not stop at the loading dock. The same builders are wiring those rooms to each other, to the electric grid, and toward a layer of computers in orbit. The working name for the result is simple. An Earth computer, with an orbital grid plugged in as the next slot.

A hall is already a board

A PC motherboard exists to do four jobs. Place parts close together. Feed them clean power. Move data on short paths. Dump heat before it cooks the silicon.

A data hall does the same four jobs at building scale. Identical racks stand in parallel rows. Hot aisles and cold aisles keep exhaust off the next intake. Busways and PDUs act like power planes. Overhead fiber acts like printed traces. The 19-inch rack is a slot. Swap a server the way you swap memory.

Google engineers Luiz Barroso and Urs Hölzle named this in 2009. The computer that mattered was no longer one server. It was the warehouse, designed as one system. In 2012 Urs Hölzle said an individual server means nothing if you measure power as an abstract pool. On April 28, 2025, Vertiv said the quiet part in industry language. The data center should be treated as a machine, not a building, and the facility now resembles a printed circuit board. NVIDIA chief executive Jensen Huang has said the data center itself is the unit of compute. In 2026 he called the building a factory that makes tokens, the small chunks of output an AI model produces.

In 2014, Silvio d’Ascia Architecture even designed a Shanghai Futures Exchange campus in Zhangjiang so the buildings read as chips on a board. That was architecture. The operating halls are engineering. Same drawing.

The next traces run between cities

One board is not enough for the models now being trained. So the builders started connecting boards.

On September 18, 2025, Microsoft opened Fairwater in Wisconsin as one AI supercomputer spread across about 1.2 million square feet and hundreds of thousands of NVIDIA GPUs. On November 12, 2025, a second Fairwater site near Atlanta came online and was joined to Wisconsin on a private fiber network Microsoft calls an AI WAN. Microsoft said it added 120,000 miles of dedicated fiber in a year. The first span is about 700 miles. A protocol built with OpenAI and NVIDIA, called MRC, is meant to keep those distant chips on the same job. Microsoft’s own phrase was planet-scale AI superfactory. Sister sites are planned in Loughton in the United Kingdom and Narvik in Norway.

That is not backup. That is one computer whose traces happen to cross state lines.

Meta uses the same map. Scale-up inside a rack. Scale-out inside one factory. Scale-across between factories when one site runs out of power or chips. Public figures include a 2024 cluster of about 129,000 H100-class GPUs and a fleet near 1.3 million by the end of 2025, with gigawatt-class systems still being added.

On August 17, 2026, NVIDIA put $1.5 billion behind a Pike County, Ohio campus that OpenAI will lease for 20 years. First phase load: 4.25 gigawatts, with an option to 8. Huang’s point matches the motherboard logic. The scarce pins are land, power, and a shell, not only chips.

AWS and other operators say AI traffic between sites is heavier and more two-way than old cloud traffic. Meshes replace hub-and-spoke. The Earth computer, if it exists, is a fabric of boards, not one glowing cube.

The grid is the power plane

A motherboard without power planes is dead. The Earth version of those planes is the electric grid.

A Goldman Sachs note dated May 20, 2026, put U.S. data center load at 31 gigawatts in 2025 and 66 gigawatts in 2027. Other 2026 studies warn that AI halls could take a much larger share of U.S. electricity by 2030. New grid hookups often take three to seven years.

So the same operators are teaching the halls to behave like a dimmer, not a brick. On December 5, 2025, Nature Energy published a Phoenix test on 256 GPUs. Software cut power about 25 percent for three hours at peak and the work still hit its quality target. Later tests described a U.K. cluster following more than 200 grid events, and live inference sliding from Virginia to Illinois during a winter peak. EPRI, NVIDIA, and partners planned a 96 megawatt flexibility site in Manassas, Virginia, for late 2026. Google signed about 1 gigawatt of demand-response deals with utilities.

That is not courtesy to the power company. That is the Earth computer learning to share one power plane. When the grid is tight, a job slows, pauses, or moves. When the grid is slack, the board runs hot.

Orbit is the next slot

Earth is running out of easy substations, cheap land, and spare water. The documented answer is to add a slot above the atmosphere.

Google and Planet’s Project Suncatcher aims to fly two satellites with Google AI chips by early 2027. A November 4, 2025 Google writeup sketched a later pack of 81 satellites flying one or two hundred meters apart, using an orbit where sunlight almost never stops. Planet chief executive Will Marshall said a working system would need thousands of satellites.

SpaceX has shown an AI1 satellite at about 150 kilowatts peak compute, talked in June 2026 about a huge Gigasat factory in Bastrop, Texas, and filed for a very large orbital computing constellation. Those are filings and targets, not a finished sky computer. Starcloud has already run a data-center-class NVIDIA GPU in orbit. China has described a Three-Body computing constellation, with early nodes on the record and a much larger mesh on the plan. Jeff Bezos said in 2025 that gigawatt-scale space data centers could arrive in 10 to 20 years.

Orbit gives continuous solar power and a heat dump into empty space. It also gives junk, radiation, and a long, fragile talk-back path to the ground. The design still fits the board. Ground halls are the dense chips. Fiber and the grid are the traces and power planes. Satellites are the expansion card.

What they are actually connecting

No single company owns every hall. No single operating system will boot the planet. Law, oceans, and the speed of light keep the layers from acting like two sockets on one small board.

The intent in the buildout is still clear from the drawings. Build each hall as a motherboard. Connect those halls across regions with private fiber. Tie their power draw to the public grid so the whole thing can flex. Test a solar-powered layer in orbit so the board can grow after the ground is full.

That is an Earth computer under construction. Not finished. Not secret. Visible in campus maps, fiber miles, utility contracts, and satellite filings.

The old sentence was, your files live in the cloud. The new sentence is shorter. The cloud was always a computer. They are now wiring it to the size of the planet, and hanging a grid in the sky.

References

• Barroso, Luiz André, and Urs Hölzle. The Datacenter as a Computer. Warehouse-scale computer argument, 2009.

• Levy, Steven. “Google Throws Open Doors to Its Top-Secret Data Center.” Wired, October 17, 2012.

• Mingtiandi. “Architect Designs Zhangjiang Data Centre to Mimic Motherboard.” Silvio d’Ascia Architecture, Shanghai Futures Exchange campus, April 21, 2014.

• Data Center Dynamics. “If cars are now computers, why are data centers still buildings?” Vertiv and the printed-circuit-board data center, April 28, 2025.

• Microsoft. “Inside the world’s most powerful AI datacenter.” Fairwater Wisconsin, September 18, 2025.

• Microsoft. “Infinite scale: The architecture behind the Azure AI superfactory.” Fairwater Atlanta and AI WAN, November 12, 2025.

• Microsoft Source. “From Wisconsin to Atlanta: Microsoft connects datacenters to build its first AI superfactory.” November 12, 2025.

• Nature Energy. “AI data centres as grid-interactive assets.” Phoenix 256-GPU flexibility test, December 5, 2025.

• Utility Dive. Data center flexibility, Goldman Sachs load figures, Aurora Manassas plan, June 25, 2026.

• Congressional Research Service. “Data Centers and the Electricity Grid: Frequently Asked Questions.” Report R49326, update August 31, 2026.

• Forbes, Ashish Bhatia. “Why Nvidia Is Building Data Centers, Not Just The Chips Inside Them.” Pike County, SB Energy, OpenAI lease, August 20, 2026.

• SpaceNews. Planet and Google Project Suncatcher partnership, December 30, 2025.

• Google research blog and related arXiv paper on space-based TPU clusters. Suncatcher concept, November 4, 2025.

• NVIDIA Blog. “How Starcloud Is Bringing Data Centers to Outer Space.” Philip Johnston, October 15, 2025.

• Data Center Dynamics. SpaceX orbital data center constellation filing, January 31, 2026.

• Data Center Dynamics. SpaceX AI1 satellite power figures, Ian Dahl and Elon Musk, June 8, 2026.

• Tom’s Hardware. SpaceX Gigasat factory, Bastrop, Texas, June 9, 2026.

• TRT World. China Three-Body Computing Constellation, Zhejiang Lab and ADA Space, May 22, 2025.

• World Economic Forum. “How data centres in space sustainably enable the AI age.” January 16, 2026.