Musk "highly confident": In 2027, NVIDIA Corporation (NVDA.US)'s most powerful AI computer will be powered on in space.

Title context: Musk "highly confident": In 2027, NVIDIA Corporation (NVDA.US)'s most powerful AI computer will be powered on in space. Text: Recently, Musk dropped another bombshell. The head of SpaceX (SPCX.US), replying to a netizen on X, made it clear: "I am highly confident that SpaceX will deploy NVIDIA Corporation (NVDA.US) VR NLV72 AI computers in space next year." For a CEO known for giving aggressive timelines, the weight of this wording should not be underestimated. It is understood that the VR NLV72 he mentioned is the Vera Rubin NVL72 that NVIDIA Corporation has been mass-producing this year. This rack-scale AI supercomputer integrates 72 Rubin GPUs and 36 Vera CPUs, with single-rack inference compute reaching 3.6 EFLOPS and training compute at 2.5 EFLOPS. The core Rubin GPU is based on Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR (TSM.US)'s 3nm process, integrates 336 billion transistors, carries 288GB of HBM4 memory with 22TB/s bandwidth, and delivers single-card inference performance 5 times that of the previous-generation Blackwell. The total memory plus VRAM capacity of the entire rack reaches 74.7TB, roughly equivalent to the combined memory of about 4,500 mainstream phones. NVIDIA Corporation's own claim is that compared with GB200 NVL72, the inference cost per million tokens is only one-tenth. Not just putting a chip up there Sending such a piece of equipment into space is a completely different order of magnitude from previous experiments that "run a GPU in orbit." SpaceX's roadmap is more specific than outsiders imagine. According to CFO Bret Johnsen at a Goldman Sachs Group, Inc. meeting, the company will launch the first batch of Starmind AI1 satellites in the fourth quarter of 2027 and greatly expand deployment scale in 2028. These satellites are essentially "racks in space" - reusing the Starlink V3 satellite platform, removing the communications phased-array antenna, replacing it with computing payloads and larger CECEP Solar Energy arrays, and adding a set of 110-square-meter deployable liquid-cooling radiators. The first-generation AI1 satellite has a deployment altitude of about 20 meters, a wingspan of 70 meters, is equipped with a 210-kilowatt CECEP Solar Energy battery array, and has average computing power of 120 kilowatts and peak power of 250 kilowatts. The production side is also advancing. SpaceX's AI satellite factory in Bastrop, Texas, aims to achieve large-scale mass production by the end of 2027, with a long-term plan to deploy about 1 million AI satellites. At the chip level, SpaceX is already a "diehard" of NVIDIA Corporation. Musk said very bluntly on the earnings call: "We believe the Vera Rubin architecture is the best architecture, the best AI computer, so we only choose NVIDIA Corporation." Johnsen added that the partnership with NVIDIA Corporation helps SpaceX obtain scarce capacity allocations at a time when GPU supply is tight. Heat dissipation: a wall in a vacuum But the real bottleneck of orbital computing may not be whether there are enough chips, but where the heat goes. This problem seems counterintuitive - the background temperature of space is minus 270 degrees Celsius, so how could it be hot? The reason is that space is a vacuum, with no air and no water, so heat cannot be dissipated through convection or conduction and can only be slowly "radiated away" through infrared radiation. The numbers are brutal. A white paper from Starcloud, a company focused on space data centers, estimates that a double-sided radiator at around 20C can radiate only about 633 watts per square meter, more than 1,000 times slower than ground-based liquid-cooling systems. In other words, an orbital data center of just 1 megawatt would require about 1,600 square meters of heat-dissipation area, nearly the size of an ice hockey rink. Ground-based hyperscale data centers, by contrast, often run into the hundreds of megawatts, and applying that ratio in space would mean the weight of the radiators would directly crush launch economics. Use the back of the CECEP Solar Energy panel for heat dissipation? Sounds clever, but the high temperature on the side directly facing the sun would severely reduce heat-dissipation efficiency, making it unworkable in practice. SpaceX's solution is to have the satellites operate in a sun-synchronous orbit so that the radiators always remain in shadow. But this also limits solar power-generation efficiency, amounting to a trade-off between heat dissipation and power supply. Radiation is another hurdle. High-energy particles hitting semiconductors can cause bit flips, instantly contaminating data during training. NVIDIA Corporation has specifically launched the Space-1 Vera Rubin module for this, using radiation-hardened designs such as lockstep processing and ECC error correction, with inference capability reportedly 25 times that of the H100. However, these figures currently have no independent third-party verification and should be regarded as vendor marketing claims. The track is gradually getting crowded SpaceX is far from the only one eyeing space data centers. Alphabet Inc. Class C (GOOGL.US)'s Project Suncatcher plans to use CECEP Solar Energy satellites equipped with TPUs to build an orbital AI cloud, with prototype satellites expected to launch around 2027, and it is in talks with SpaceX about launch cooperation. Bezos's Blue Origin directly submitted an application to the FCC for 51,600 data center satellites, under the code name "Project Sunrise." But Bezos himself poured cold water on the timeline. He once said in an interview that the economic advantages of space data centers "may take another 20 years to materialize," in sharp contrast with Musk's claim of "two to three years." Startups are not idle either. Washington-based Starcloud already sent a satellite carrying NVIDIA Corporation H100 into orbit in November 2025 and successfully ran AI workloads in orbit. The company has reached a valuation of $1.1 billion and is applying to the FCC for a constellation license for as many as 88,000 satellites. Market research institutions predict that the global orbital data center market was worth about $2.08 billion in 2025 and is expected to expand at a compound annual growth rate of 25.15% through 2036. The DRIVE is clear: ground-based data centers face increasingly severe bottlenecks in power supply and land approval, and 70% of people across the United States already oppose building AI data centers near their homes. The financial logic and the reality gap To understand why Musk is in such a hurry to send computing power into orbit, one must first look at how big SpaceX's AI business is right now. SpaceX already operates one of the world's top five AI infrastructure businesses. A hosting agreement with just one customer, Anthropic, contributes about $3.75 billion in quarterly revenue. In September this year, SpaceX signed another new hosting contract with a monthly payment of $1.11 billion, or about $13 billion annualized, with billing starting on December 1. Johnsen said these contracts give management more confidence in achieving the $100 billion ARR target - after all, SpaceX's revenue in the second quarter of 2026 was only $7.8 billion, and although the AI segment grew 247% year over year, it still had an operating loss of $1.3 billion. Against this backdrop, the significance of orbital computing is not just a technological vision. SpaceX listed on Nasdaq in June this year at a valuation of $1.77 trillion, setting the record for the world's largest IPO, and Musk needs a sufficiently grand narrative to support that valuation. Space data centers - a blue ocean whose market size could reach hundreds of billions of dollars - are clearly far more attractive than "rocket launch services." But SpaceX also admitted in its IPO filing that the space AI computing business "depends on unproven assumptions." The laws of physics are there, and engineering challenges cannot be filled in by confidence alone. As Igor Bargatin, a professor of mechanical engineering at the University of Pennsylvania, said: "The technology exists, but applying it to space data centers is not realistic."