SpaceX 的 S-1 显示出马斯克式的太空雄心：把运算能力搬到轨道上，建立 AI 资料中心。公司认为，长期成败部分取决于能否在地球以外部署算力，以避开电力与基础设施限制；但其文件同时承认，这一路径受制于 GPU 供应不足、晶片短缺与脆弱的供应链，且「未来的 AI 将由对实体堆叠的控制决定」。

产业瓶颈已相当明显：Jensen Huang 表示供应链每年「超过翻倍」，甚至可能每年四倍成长，但至少十年内仍难追上需求。较旧的 Nvidia H100 系统交期可逼近 1 年，而多数较新的 Blackwell 系统已被预订到 2027 年；问题不只在 GPU，HBM 记忆体与先进封装也高度集中于少数供应商。

SpaceX 也在推进 Terafab，这是与 Tesla 相关、旨在为未来资料中心制造自家晶片的计划，但文件承认可能失败。同时，公司已建成 Colossus 与 Colossus II 资料中心，合计约 1 GW 运算能力，部分依赖自建天然气发电与 Tesla Megapack 电池；下一阶段要训练 Grok 5。尽管 Nvidia、TSMC、SK Hynix、Samsung 等正扩产，短期内 AI 竞赛仍受晶片、记忆体与电力限制。

SpaceX’s S-1 shows a Musk-like ambition to move compute into orbit and build AI data centers in space. The company says its long-term future partly depends on deploying computing power beyond Earth to escape power and infrastructure limits, while also acknowledging that GPU shortages, fragile supply chains, and the control of the physical stack may define AI’s future.

Industry bottlenecks are already severe. Jensen Huang said the supply chain is still growing by more than 2x a year, perhaps 4x, yet will struggle to keep up for at least a decade. Older Nvidia H100 systems can take close to a year to arrive, and most newer Blackwell systems are already spoken for through 2027; the constraints also extend to HBM memory and advanced packaging, which are concentrated among a few suppliers.

SpaceX is also advancing Terafab, a Tesla-linked plan to make in-house chips for future data centers, though the filing says it may fail. Meanwhile, Colossus and Colossus II already provide about 1 GW of computing power, helped by self-built gas generation and Tesla Megapack batteries, and the next phase is meant to train Grok 5. Even with capacity expansions from Nvidia, TSMC, SK Hynix, Samsung and others, AI remains constrained by chips, memory and electricity today.