马斯克首次披露太空AI算力的工程细节，做云计算、卫星通信或AI基础设施的团队值得关注——这可能是未来算力格局的颠覆性方向。
马斯克首次公开详解SpaceX的太空数据中心计划，核心是AI1轨道AI卫星。该卫星峰值功率150kW，持续计算功率约120kW，相当于一个NVIDIA GB300机架。采用可展开太阳能板（250W/m²）和双面散热器（1400W/m²），通过激光链路实现1Tbps连接，轨道高度600-800公里，延迟仅6-8毫秒。卫星设计比星链更简单，无需相控阵天线，由Starship发射，计划2027年底实现合理量产，长期目标达到千兆瓦级轨道AI算力。
For the very first time Elon Musk explains the "sp…
For the very first time Elon Musk explains the "space data center plan" of @SpaceX in detail and its AI1 orbital AI data center satellite - and suddenly it looks so much closer than I thought.
He says "There’s not some magic necessary that doesn’t exist for AI satellites. As Ian said this is a lot of this is technology we’ve already made for the… we basically don’t think this is a super hard problem compared to things that we already do."
📌 Power and compute capacity: - 150 kW peak power - ~120 kW sustained/average compute power - Roughly equivalent to one full NVIDIA GB300 (or upcoming Rubin) rack in a typical data-center operating envelope (~140 kW peak is possible but 120 kW average is more realistic for sustained workloads).
📌 Solar array: - Assumed efficiency: 250 W/m² (expected to improve beyond this). - Large, deployable solar panels (evolutions of the solar arrays already flying on Starlink V3 satellites).
📌 Radiators (thermal management): - Double-sided design, oriented “knife-edge” to the Sun to minimize solar heating. - Heat rejection: ~1,400 W/m² (expected to improve). - Radiator panels are roughly the same size/scale as the Starlink V3 solar arrays (~70 m wingspan class).
📌 Design philosophy: - Significantly simpler than a Starlink satellite — no massive phased-array antennas or complex communications hardware. - Core elements: solar panels + radiators + compute chips + laser links. - Larger overall than Starlink sats but described as “the easier one to design for.”
📌 Connectivity: - ~1 terabit/s via inter-satellite laser links. - Can mesh with the existing Starlink constellation or link directly to ground. - Low latency: satellites planned for ~600–800 km altitude → light-travel time yields only ~6–8 ms round-trip (light travels ~300 km per millisecond).
📌 Deployment and operations: - Launched by Starship (the only vehicle capable of the required millions-of-tons-to-orbit scale). - Part of a future large constellation (potentially up to ~1 million satellites). - Orbital data centers can be networked together or routed through Starlink for terrestrial users.
📌 Manufacturing and timeline: - Production in Bastrop, Texas. - Solar manufacturing facility already under construction. - Dedicated AI satellite production building to follow. - Reasonable-volume production targeted by end of next year (2027). - Initial chips will use existing NVIDIA GB300/Rubin designs with SpaceX reference hardware; future scaling via a new “Terra Fab” chip factory (~100 million sq ft, 10× the size of Tesla Giga Texas).
📌 Scalability notes: - Near-term goal: gigawatt-scale orbital AI compute. - Longer-term: terawatt-scale and beyond, eventually using lunar mass drivers (electromagnetic rail-gun style) to launch photovoltaics and radiators from the Moon (no atmosphere + 1/6 g makes this feasible). - Starship is expected to increase annual mass-to-orbit from today’s ~2,500 tons to millions of tons per year within a few years.