马斯克启动史上最大造芯计划Terafab：年产1太瓦算力芯片，剑指太空算力霸主

埃隆·马斯克再次以一项令世界侧目的计划刷新了雄心的边界。2025年3月22日，他在德克萨斯州奥斯汀正式宣布启动 Terafab 项目——由特斯拉、SpaceX 与 xAI 三方联合主导的 2纳米晶圆厂制造计划，被定位为"有史以来规模最大的芯片制造工厂"。目标只有一个：每年生产 1太瓦（TW）算力芯片，相当于当前全球AI芯片年产能的整整50倍。

马斯克用"疯狂"和"物理极限"来形容这个计划，但他显然是认真的。在他的判断里，全球现有芯片产能与他未来的需求之间存在一道无法靠采购弥合的鸿沟——"要么建成Terafab，要么我们将无芯片可用"，这句话道出了这场造芯行动的根本逻辑。

Terafab究竟"大"在哪里

Terafab的核心突破，在于彻底打破了传统芯片制造的全球分工模式。当前业界普遍将光刻掩膜、晶圆制造、封装测试分散在不同地区、不同企业完成，而 Terafab将把上述全流程集中在同一厂区内，形成"制作掩膜—制造—测试—优化掩膜—再制造"的极速迭代闭环。

马斯克称，全球目前没有任何一家工厂能将逻辑、存储、封装、测试、光刻掩膜全部整合在一起，Terafab的迭代速度将比常规产线高出一个数量级。

工厂内设两个晶圆厂，分工明确：

• 边缘推理芯片：主要搭载于 Optimus 人形机器人与特斯拉自动驾驶系统。马斯克预判未来人形机器人年产能将达 10亿至100亿台，是汽车市场的10至100倍，芯片需求极为庞大。
• 太空高功率定制芯片：专为 SpaceX 轨道 AI 数据中心设计，具备抗辐射、抗老化等太空级指标，运行参数全部针对太空极端环境专项定制。

算力为何要搬上太空

Terafab最具争议、也最具想象空间的部分，在于其算力部署重心在太空而非地面。马斯克给出了一套量化逻辑：地球仅接收太阳总能量的五亿分之一，地球算力扩容存在无法突破的能源天花板；而太空无大气衰减、无昼夜更替，卫星直面太阳的太阳能获取效率是地面的5倍以上，且无需大规模储能配套，硬件成本更低。

按照他的规划，80%的太瓦级算力将部署于太空轨道，地球每年仅部署100至200吉瓦（约占总产能20%）。更关键的是，他预判在2至3年内，太空AI算力的部署成本将全面低于地面，届时"把算力放到太空将变得几乎显而易见"。

这一布局并非只是技术路线的选择，更是马斯克多年来推进多星球文明愿景的组成部分——从月球基地到星际算力网络，Terafab在他的蓝图中扮演着基础设施奠基者的角色。

值得关注的是，马斯克明确表态将继续向三星、台积电、美光等供应商采购芯片，Terafab并非取代现有供应链，而是在其之上构建一套完全不同量级的增量能力。这场造芯计划能否落地，将深刻影响全球AI算力格局、芯片制造产业分工，乃至人类在太空部署计算基础设施的进程。

Elon Musk Launches Terafab: The World's Most Ambitious Chip Manufacturing Plan Targeting 1 Terawatt of AI Computing Power

Elon Musk has once again redefined the boundaries of ambition. On March 22, 2025, in Austin, Texas, he officially unveiled Terafab — a 2-nanometer wafer fabrication project jointly led by Tesla, SpaceX, and xAI, billed as "the largest chip manufacturing facility ever built." The goal is singular and staggering: produce 1 terawatt (TW) of AI computing chips annually — 50 times the entire world's current AI chip output.

Musk described the plan as "insane" and as pushing "the limits of physics," yet he was clearly in earnest. In his view, global chip production capacity is fundamentally incapable of meeting his future demands, no matter how aggressively he sources from existing suppliers. His blunt conclusion: "Either we build Terafab, or we will have no chips." That single statement captures the core logic behind this historic chip-making venture.

What Makes Terafab Truly Different

Terafab's most radical innovation lies in dismantling the global division of labor that has long defined the semiconductor industry. Today, photomask fabrication, wafer manufacturing, and packaging and testing are spread across different companies and geographies. Terafab will consolidate the entire process within a single campus, creating an ultra-fast iteration loop: mask making → fabrication → testing → mask optimization → re-fabrication.

Musk noted that no facility in the world currently integrates logic, memory, packaging, testing, and photomask production under one roof. Terafab's iteration speed is expected to be an order of magnitude faster than conventional production lines.

The facility will house two dedicated fabs with distinct mandates:

• Edge inference chips: Designed for the Optimus humanoid robot and Tesla's autonomous driving systems. Musk projects that global humanoid robot production could eventually reach 1 billion to 10 billion units per year — 10 to 100 times the scale of the automotive market — driving enormous chip demand.
• Space-grade high-power custom chips: Engineered specifically for SpaceX's orbital AI data center network, with radiation resistance, anti-aging properties, and performance parameters tailored to the harsh conditions of space.

Why Move Computing Power Into Space

The most audacious — and most imaginative — dimension of Terafab is its strategic emphasis on space-based computing rather than ground-based infrastructure. Musk's reasoning is grounded in hard numbers: Earth receives only one five-hundred-millionth of the Sun's total energy output, placing an absolute ceiling on terrestrial computing expansion. Space, by contrast, offers satellites direct, unobstructed solar exposure with efficiency more than five times higher than ground-based solar, with no need for large-scale energy storage and lower hardware costs overall.

Under his blueprint, 80% of the terawatt-scale computing capacity will be deployed in orbital space, while Earth hosts only 100 to 200 gigawatts annually — roughly 20% of total output. More critically, Musk projects that within two to three years, the cost of deploying AI computing in space will fall below that of ground-based alternatives, at which point "putting computing power in space will become almost obviously the right move."

This strategic orientation is not merely a technical preference — it is a central pillar of Musk's long-held vision of a multi-planetary civilization, from lunar bases to an interstellar computing network. Within that blueprint, Terafab serves as the foundational infrastructure layer.

It is worth noting that Musk explicitly stated he will continue purchasing chips from Samsung, TSMC, Micron, and other existing suppliers. Terafab is not designed to replace the current supply chain, but to build an entirely different order of incremental capacity on top of it. Whether this plan materializes as envisioned will profoundly shape the global AI computing landscape, the structure of the chip manufacturing industry, and humanity's long-term capacity to deploy computational infrastructure beyond Earth.