Shadow Foundry: Technical Brief Running a frontier-scale AI model on simple hardware, from 1990s silicon to a lunar cold trap

Shadow Foundry is an order-of-magnitude design study of how simple the hardware can be and still run a frontier-scale AI model, if speed is given up. The working case is a 2-trillion-parameter mixture-of-experts model with 200 billion parameters active per token, stored at 4 bits per weight (1.2 TB with error correction). The study finds that storage, not computation, sets the technology floor. Two machines are worked through. The first is an Earth machine built on 350 or 180 nm chip processes, with weights held in read-only or one-time-programmable memory. It runs about 20 tokens per second on 10–30 kW. The second is a machine in a permanently shadowed crater at a lunar pole, using niobium-nitride superconducting logic at 8–10 K, patterned without photoresist by stencil or laser scribing, with weights stored as laser-programmed inductive links on about 14,000 m² of tile. It produces one token every 10 seconds per stream, across about 100 streams, on roughly 0.5 kW. The crater machine is not the efficient choice: silicon made on Earth and landed on the Moon would give more tokens per watt and per kilogram. Its case is that it could be made there, from local materials. The brief derives every figure in the open and tags each with its basis and uncertainty. It closes with a five-phase, gated research program of about $6–17B over 15 years, of which the first three phases, all on Earth, total $130–260M. Progress toward self-sufficiency is measured by the mass imported from Earth per year. Nothing described here has been built, and no specialist has yet reviewed the work. This brief was written with Claude, Anthropic’s AI model. Carl Quillen directed the work, reviewed it and takes responsibility for the content. “影子铸造厂”(Shadow Foundry)是一项以数量级为尺度的设计研究,要回答的是:若不再追求速度,运行前沿级大模型的硬件究竟能简单到什么地步。研究对象是一个两万亿参数的混合专家模型,每个词元激活约两千亿参数,权重以4比特精度存储(算上纠错开销共1.2 TB)。研究发现,决定技术下限的是存储,而不是计算。 文中推演了两种机器方案。第一台建在地球上,采用350或180纳米制程,权重存放在掩膜只读存储器或一次性可编程存储器中,功耗10至30千瓦,每秒约20个词元。第二台位于月球极区永久阴影区的撞击坑内,采用在8至10 K下工作的氮化铌超导逻辑,图形经镂空掩膜成形,或由激光刻划成形,全程不用光刻胶;权重以激光熔断编程的电感连接形式,存于约1.4万平方米的基板上。在约100路并发会话下,每路每10秒生成一个词元,整机功耗约500瓦。这台机器并不高效:把地球上制造的硅芯片运上月球,每瓦、每千克能产出的词元都更多;它的意义在于可以用月球本地材料就地制造。 文中每个数字都给出推导,并注明依据与不确定度。文末给出一份分五个阶段、逐阶段设关口的研究计划,15年内约需60亿至170亿美元;其中前三个阶段全部在地球上完成,合计1.3亿至2.6亿美元。自给程度以每年仍需从地球运入的质量来衡量。上述一切均尚未建成,也未经相关领域专家评审。 本文由 Carl Quillen 与 Anthropic 的人工智能模型 Claude 共同撰写。Carl Quillen 主持并审阅了这项工作,并对内容负责。