Musk is “highly convinced”: in 2027, Nvidia (NVDA.US)'s most powerful AI computer will be launched in space

The Zhitong Finance App learned that recently, Musk threw another bombshell. The head of SpaceX (SPCX.US) clearly stated in response to netizens on X: “I am highly convinced (highly convinced) SpaceX will deploy the NVDA.US (NVDA.US) VR NLV72 AI computer in space next year.” For a CEO who is used to giving aggressive timelines, the weight of this phrase should not be underestimated.

According to information, the VR NLV72 he mentioned is the Vera Rubin NVL72, which Nvidia fully mass-produced this year. This rack-level AI supercomputer integrates 72 Rubin GPUs and 36 Vera CPUs, with a single cabinet inference computing power of 3.6 EFLOPS and a training computing power of 2.5 EFLOPS. The core Rubin GPU is based on the TSM.US (TSM.US) 3nm process, integrates 336 billion transistors, and is equipped with 288GB HBM4 memory, 22Tb/s bandwidth. The single-card inference performance is 5 times that of the previous generation Blackwell. The total memory and video memory capacity of the entire cabinet is as high as 74.7TB, which is almost equivalent to the total memory of 4,500 mainstream mobile phones. Nvidia's own statement is that compared to GB200 NVL72, the inference cost per million tokens is only one-tenth.

It's not just about putting a chip on it

Bringing such a device to heaven and the previous “running a GPU in orbit” experiments are entirely on two levels.

SpaceX's roadmap is more specific than the outside world might think. According to CFO Bret Johnsen at the Goldman Sachs conference, the company will launch the first Starmind AI1 satellites in the fourth quarter of 2027 and significantly expand the deployment scale in 2028. These satellites are essentially “racks in space” — multiplexing the Starlink V3 satellite platform, removing the communication phased array antennas, replacing them with computational payloads and larger solar arrays, plus a 110-square-meter expandable liquid-cooled radiator. The first AI1 satellite is deployed at a height of about 20 meters, has a wingspan of 70 meters, and is equipped with a 210 kilowatt solar array with an average computational power of 120 kilowatts and a peak of 250 kilowatts.

The production side is also progressing. SpaceX's AI satellite factory in Bastrop, Texas aims to achieve large-scale mass production by the end of 2027, and plans to deploy about 1 million AI satellites in the long term.

At the chip level, SpaceX is already a “die-hard fan” of Nvidia. Musk put it bluntly during the earnings call: “We think the Vera Rubin architecture is the best architecture and the best AI computer, so we only chose Nvidia.” Johnsen added that the partnership with Nvidia helps SpaceX obtain scarce production capacity quotas at a time when GPU supply is tight.

Heat dissipation: a wall in a vacuum

However, the real bottleneck in orbital computing is probably not that the chips are not enough, but rather where the heat is arranged.

The question seems counterintuitive — the background temperature of space is minus 270 degrees Celsius, how can it be hot? The reason is that space is a vacuum. There is no air or water, heat cannot be dissipated through convection or conduction; it can only slowly “glow” and dissipate heat by infrared radiation.

The numbers are cruel. The white paper of Starcloud, a space-focused data center company, estimates that a double-sided radiator can only radiate about 633 watts/square meter at around 20°C, which is more than 1,000 times slower than liquid cooling systems on the ground. In other words, a single megawatt orbital data center requires about 1,600 square meters of cooling area, about the size of a hockey rink. However, large-scale data centers on the ground can easily reach 100 megawatts. Placing this ratio in space means that the weight of radiators will directly crush the launch economy.

Using the back of a solar panel to dissipate heat? It sounds great, but the high temperature on the direct surface of the sun can seriously reduce heat dissipation efficiency, and it doesn't work in actual operation. SpaceX's solution is to keep the satellite in sun-synchronous orbit, so that the radiator remains shaded at all times. However, this also limits the efficiency of daily power generation, which is tantamount to making a trade-off between heat dissipation and power supply.

Radiation is another hurdle. High-energy particles impacting semiconductors may cause bits to flip, and data during training will be instantly polluted. Nvidia specially launched the Space-1 Vera Rubin module for this purpose, using radiation-resistant reinforcement designs such as lock-step processing and ECC error correction. The reasoning ability is said to be 25 times that of the H100. However, this data has not yet been verified by an independent third party and should be regarded as the manufacturer's propaganda.

The racetrack is getting crowded

SpaceX is far more focused on space data centers.

Google (GOOGL.US)'s Project Suncatcher plans to use a solar satellite equipped with TPU to form an orbital AI cloud. It is expected to launch a prototype satellite around 2027 and is in negotiations with SpaceX for launch cooperation. Bezos' Blue Origin directly submitted an application for 51,600 data center satellites to the FCC. The project code is “Project Sunrise.”

But Bezos himself threw cold water on the schedule. He once said in an interview that the economic advantages of space data centers “may take another 20 years to be realized”, in stark contrast to Musk's claim of “two to three years.”

Startups aren't idle either. Starcloud in Washington state has already launched a satellite carrying the Nvidia H100 into orbit in November 2025 and successfully operated AI workloads in orbit. The company has reached a valuation of 1.1 billion US dollars and is applying for constellation permits for up to 88,000 satellites from the FCC.

Market research institutions predict that the global rail data center market will be worth about US$2.08 billion in 2025 and is expected to expand at a compound annual growth rate of 25.15% by 2036. The driving force is clear: the electricity supply and land approval bottlenecks faced by terrestrial data centers are becoming increasingly severe, and 70% of people across the US are already opposed to building AI data centers near their homes.

The gap between financial logic and reality

To understand why Musk is in a hurry to put computing power into orbit, we must first look at how big SpaceX's AI business market is right now.

SpaceX currently operates the top five AI infrastructure businesses in the world. Anthropic's client's hosting agreement alone contributed approximately $3.75 billion in quarterly revenue. In September of this year, SpaceX also signed a new hosting contract with a monthly payment of 1.11 billion US dollars, with an annualized rate of about 13 billion US dollars, and billing will begin on December 1. Johnsen said these contracts gave management more confidence to achieve the $100 billion ARR target — know that SpaceX's revenue for the second quarter of 2026 was only $7.8 billion. Although the AI sector grew 247% year over year, operating losses were still $1.3 billion.

In this context, the significance of orbital computation is more than just a technical vision. SpaceX landed on the NASDAQ with a valuation of 1.77 trillion US dollars in June this year, setting a record for the world's largest IPO. Musk needed a sufficiently grand narrative to support this valuation. Space data centers — a blue ocean with a market size that could reach 100 billion dollars — are clearly far more appealing than “rocket launch services.”

However, SpaceX also acknowledged in the IPO file that the space AI computing business “relies on unproven assumptions.” The laws of physics are there, and engineering challenges cannot be filled with confidence. As Igor Bargatin, a professor of mechanical engineering at the University of Pennsylvania, said, “The technology exists, but it's not realistic to apply it to space data centers.”