SPCX's Space Data Center in 2027! **The cooling solution has been finalized: allowing chips to operate at higher temperatures, with each satellite carrying a complete NVL72 rack.** SpaceX has disclosed that AI1 has a peak power of 250 kW and an average power of 175 kW, equipped with a 160-square-meter heatsink, but has not yet released its operating temperature. Based on our heatsink parameter table released in June, this means the heatsink surface temperature is approximately 65-75°C, and the chip junction temperature is 80-90°C. Our industry research also confirms this: the space version of the chip will indeed operate at a junction temperature of 80-90°C. From an engineering perspective, this is feasible because chips in ground racks often exceed 80°C when operating at full load. This temperature itself is not a problem for the chip; the real limiting factor is the memory. In orbit, increasing the operating temperature is more helpful for heat dissipation: in a vacuum, heat can only be dissipated through thermal radiation, and the radiation heat dissipation capacity is proportional to the fourth power of the absolute temperature. If a typical terrestrial cooling temperature of approximately 40°C is used, a heatsink of 240-290 square meters would be required. By increasing the operating temperature, a satellite weighing approximately 2 tons can carry a complete NVL72 system, and a Starship can carry approximately 40 such satellites, without requiring on-orbit assembly. **The computing payload has been finalized: the space version of NVL72, codenamed "Kessel."** In August of this year, SpaceX announced a collaboration with NVIDIA to design the AI1 computing payload, using Vera CPUs and Rubin GPUs. The architecture of the space rack was finalized in October: retaining the NVLink interconnect domain consisting of 72 GPUs, compressing the 48U vertical rack on the ground into a flat chassis of approximately 9U, and reducing the weight to a few hundred kilograms. The first-generation product is primarily used for inference; large model training remains on the ground. Both the chips and HBM use standard products. Radiation issues are addressed through an external shielding layer, the HBM's built-in ECC error correction function, and system-level fault tolerance mechanisms; heat dissipation is achieved through a liquid-cooled loop directly connected to an external heatsink. Therefore, the demand for advanced materials and lightweight alloys will primarily focus on shielding layers, heat sinks, and structural components. **Supply chain orders have been placed: 6,000 racks, corresponding to 1.05 gigawatts, to be procured in 2027.** SpaceX has ordered 6,000 space-specific racks for delivery in 2027 from a leading server ODM manufacturer. This manufacturer is also the project's sole JDM (Joint Design Manufacturer), planning to complete NPI (New Product Introduction) in May 2027 to catch up with the test satellite schedule and complete production in December. These 6,000 racks correspond to 432,000 GPUs; based on the price of ground racks, the hardware value is approximately $42-45 billion. Assuming one rack per satellite, the total power is 1.05 gigawatts. The total rack requirement for the entire project lifecycle is approximately 30,000 racks, consistent with the plan to deploy approximately 30,000 satellites in the first phase. Based on 72 GPUs per rack, this totals approximately 2.16 million GPUs, corresponding to an average power consumption of 5.25 gigawatts and a peak power consumption of 7.5 gigawatts. The initial 6,000 racks represent one-fifth of this total. **Restrictions on terrestrial construction are driving computing power towards space.** The Democratic Party won both houses of Congress in the November 3rd midterm elections, and the incoming Texas governor has pledged to halt new data center construction until the state legislature enacts relevant rules. Terrestrial data centers are facing increasing pressure regarding approvals, grid connection, and cost sharing, primarily impacting new capacity added in 2028 and beyond—precisely the period when Starmind begins to scale. Orbital computing requires no land, grid connection, or water, and does not require state or local government data center construction permits. As terrestrial expansion becomes more restricted, the attractiveness of orbital computing will increase. **Optimistic Scenario: By the end of 2028, the power consumption of on-orbit computing power will be approximately 1 gigawatt.** **In our optimistic scenario, Starship will have approximately 220 available launch opportunities in 2028: Starbase will expand to 100 launches per year, plus 44 from Launch Complex 39A and 76 from Launch Complex 37. About 70%, or 154 launches, will be used to launch computing satellites. Assuming 40 satellites per launch, this is sufficient to put all 6,000 racks into orbit. Therefore, we project that, under the optimistic scenario, by the end of 2028, the on-orbit computing power will correspond to approximately 1 gigawatt of power, calculated based on an average computing payload power of 175 kilowatts per satellite.** **There are two key validation milestones.** The first is the 15th test flight, expected as early as late November, which will attempt to capture both stages of the spacecraft using the launch tower. Successful capture and subsequent flight are crucial to supporting 30-40 launches by 2027. Secondly, there's the test satellite, scheduled for launch aboard Starship in May 2027. It will provide the first on-orbit thermal performance data and verify whether an average power output of 175 kilowatts per satellite can be achieved in orbit. This forms the basis for all our gigawatt-scale projections.