No One Sees What SpaceX and Nvidia Are Building

On SpaceX’s first earnings call on Aug. 4, 2026, Elon said the company had decided to build exclusively on Nvidia. He called Vera Rubin the best AI computer and said the companies now have a close partnership on many levels. SpaceX also said it is working with Nvidia on the compute payload inside StarMind, its planned AI satellite, using an optimized Vera Rubin NVL72. First launches are targeted for next year. That is a company target, not a booked manifest.

An NVL72 is a rack-scale AI computer: dozens of GPUs and CPUs, memory, network switches, and the cooling that makes them one machine. SpaceX wants that whole system rebuilt as a satellite, flown on a rocket, and put in orbit. Three commitments land at once. Nvidia becomes the exclusive AI platform. Nvidia helps design the first orbital AI computer. SpaceX plans to fly that same design on Earth and in space.

This is not Elon buying a pile of GPUs from Jensen. Nvidia is helping define the system. SpaceX is helping define where that compute lives. Each company needs the other for the full stack to work.

The scale they are talking about

SpaceX says it expects to finish this year with more than 2 GW of AI compute. By the end of next year, Elon said the total could be closer to 10 GW. A gigawatt is roughly a large power plant. Those are SpaceX and Elon targets for their own build, not a measured inventory today.

Even a fraction of that needs more than GPUs. Memory, CPUs, networking, fiber, power electronics, packaging, cooling. In space you also need the satellite structure, the launch cadence, and a network that can move the results back to Earth. Nvidia’s side of that stack is the compute core: GPUs, networking, CUDA, and with Vera Rubin a coordinated AI factory. SpaceX’s side is rockets, satellites, Starlink’s laser mesh, giant ground clusters, and internal demand through xAI. It can also rent capacity to other AI companies.

The flywheel is simple. Better Nvidia chips make orbital capacity more valuable. More orbital capacity needs more Nvidia systems. More systems mean more compute. More compute funds more satellites, more launches, and more chips.

Orbit kills every data-center assumption

A ground rack sits on concrete. Power comes from the grid. Water moves heat. A tech can pull a dead board. Orbit removes all of that. Every kilogram rides a rocket. The computer has to survive launch vibration. Radiation flips bits and kills electronics. Heat cannot blow into air that is not there. It has to leave through radiators. The satellite needs solar arrays. Nobody drives a repair truck to a dead chip.

So the orbital design needs a different balance of performance, power, weight, heat, reliability, and redundancy. Nvidia owns the compute architecture. SpaceX owns the environment. They have to cut those trade-offs together. Nvidia gets operating data from the hardest deployment of its hardware. SpaceX gets direct access to the leading AI platform. Feedback moves both ways. A design that gets lighter, cooler, or more reliable for orbit can also get cheaper on Earth.

On the Aug. 4, 2026 earnings call, Elon said SpaceX expects to deploy the optimized NVL72 design on the ground as well as in orbit, because it thinks that design is simpler and cheaper than a standard rack. Analysis, not a company program: that reuse could let orbit debug a terrestrial factory design, and the factory debug the next satellite.

Then the wall: chips

That loop hits a physical limit. Where do the chips come from? Farzad’s read of the demand curve is orders of magnitude beyond what fabs make today. SpaceX said orbital AI at scale will need significantly more AI processors than it can currently access, and named fab capacity, raw materials, geopolitics, and natural disasters as risks. The pointed answer was Terafab: the proposed SpaceX and Tesla chip project, with Intel expected on design, fabrication, and packaging expertise. The long-term goal named in the video is 1 terawatt of compute hardware a year.

A leading-edge fab takes years, costs tens of billions, and depends on thousands of suppliers. Extreme ultraviolet lithography tools for leading-edge chips come from one commercial supplier, ASML in the Netherlands. ASML's 2025 annual report recognized 48 EUV systems (44 NXE and 4 EXE); it does not publish a lifetime installed-base total. On the July 15, 2026 Q2 call, ASML said it expects to ship about 65 Low-NA EUV machines in 2026, plans a 30% add to that 2026 Low-NA capacity in 2027, and is studying a further 30% for 2028. Those are shipment and capacity plans, not a cap of about 60 a year. Terafab cannot summon hundreds of them. It does not have to replace TSMC to matter. It can cut steps that depend on one supplier, add packaging, speed iteration, and become a second source for critical parts.

Nvidia designs chips. It does not own the fab that prints them. Partners manufacture, assemble, test, and package. The advanced supply chain is still concentrated in Asia, with TSMC and Samsung on wafers and SK Hynix, Micron, and Samsung on memory. That model let Nvidia move fast without spending tens of billions on fabs. It also means Nvidia does not fully control how many chips get made, where, or how fast capacity grows. Localizing more production in the United States reduces that risk. It does not create independence. Terafab would still need ASML tools, German optics, Japanese materials, Korean and U.S. memory expertise, and a global supplier web.

Orbit changes the size of the order

SpaceX’s long-term goal in the video is up to 100 GW of compute to orbit each year. Early satellites are described around 100 kW of compute per metric ton. Hitting the full goal would take thousands of launches a year and on the order of 1 million metric tons to orbit annually. Those are planning numbers. At that scale, chip supply matters as much as propellant. SpaceX cannot plan a giant orbital network around leftover GPUs after every other customer is served. Nvidia cannot treat orbit as a major market while the supply chain is sized only for today’s ground demand.

Wall Street is being taught to finance the rack

Nvidia recently signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR for financing platforms that could mobilize more than $500 billion for AI infrastructure over time. The stated goal is to make Nvidia-based data centers easier to finance. Jensen is packaging compute the way airlines finance planes and utilities finance plants: long-lived assets with known users, cash flow, and residual value.

SpaceX fits that model unusually well. It is a giant buyer of Nvidia systems. It operates compute and can sell access. It controls a network that can distribute output. It can eventually put hardware where solar energy is constant. And it is building Terafab to attack supply directly. Paths deepen from there: investment, prepaid capacity, equipment finance, licensed designs, demand guarantees beyond Elon’s own companies.

The real tension

Elon wants control. Jensen wants Nvidia to stay the standard. If SpaceX eventually builds a general-purpose AI processor that replaces Nvidia and asks developers to leave CUDA, the relationship turns competitive fast. Analysis, not a signed deal: a hybrid would fit both sides' incentives better, with Terafab making some custom SpaceX/Tesla chips and, only if the parties later agree, some Nvidia-designed hardware. SpaceX has not announced a contract to manufacture Nvidia chips at Terafab.

That is the Apple and TSMC split applied to AI factories: design, manufacturing, and software can sit in different companies and still form one system. The exact Nvidia–Terafab relationship will move as each side proves yield and demand.

This Exclusive is from the long-form at https://www.youtube.com/watch?v=4VDAzWeC44k.