Casey Hammer, a Caltech PhD and founder of Terraform Industries, posits that the future of AI isn't just about algorithms—it's a colossal industrial race. The winner won't be who writes the best code, but who can build the most solar panels, batteries, and GPUs. In a deep-dive conversation, Hammer lays out a compelling, first-principles case for why a solar-powered future is not just possible, but inevitable.
The Short-Term Reality: Why AI Runs on Natural Gas Today
When building gigawatt-scale data centers, AI hyperscalers like Meta and xAI are currently turning to natural gas. The reason is simple: speed and availability. In the race to deploy trillions of dollars worth of GPUs, the cost of electricity is almost a rounding error.
"The hyperscalers are not power cost-sensitive, they are power availability-sensitive," Hammer explains.
For a project like xAI's "Colossus One" in Memphis, tapping into an existing gas line and deploying truck-mounted gas turbines is the fastest way to get a gigawatt of power online. This approach works for the first few data centers, but Hammer argues it will quickly hit a manufacturing wall.
- Supply Bottlenecks: The global production capacity for large gas turbines, transformers, and other grid components is limited. You can't simply 10x production overnight.
- Grid Constraints: The US electrical grid is aging and notoriously difficult to expand due to regulatory and permitting hurdles that can take years.
This reliance on natural gas is a tactical move for a battle, but not a strategy for winning the war.
The Long-Term Inevitability: Solar's Unstoppable Learning Curve
Hammer's core thesis rests on a powerful economic principle: Wright's Law, or the learning curve. For solar panels, this is an almost magical force.
"The Wright's solar coefficient is 43%. So every time we double production, we get a 43% reduction in cost."
This relentless cost reduction, which has held true for decades, means that solar energy has fundamentally different economics than any other power source. While a gas turbine is a complex piece of mechanical engineering with a relatively flat cost curve, a solar panel is a semiconductor product that gets exponentially cheaper with scale.
This creates a virtuous cycle:
- Lower prices create explosive demand.
- Explosive demand drives a doubling of production every ~2.5 years.
- Doubled production triggers another 43% cost reduction.
This is why, Hammer argues, that betting against solar is betting against one of the most powerful and consistent technological trends in history.
The Blueprint for the Future: Off-Grid AI Super-Factories
The solution to the grid bottleneck is to bypass it entirely. Hammer envisions a future where massive AI data centers are built as self-contained, off-grid systems co-located with their own power source.
Here’s the breakdown for a 1-megawatt data center rack aiming for 99.99% uptime:
- Power Source: 10 acres of solar panels. This includes a significant overbuild to handle cloudy days and winter months.
- Storage: Six truckloads of batteries (e.g., Tesla Megapacks) to provide around 24 hours of backup power.
- Location: Remote, undeveloped land in places like Texas or Nevada, where land is cheap and sunlight is abundant.
For a 5-gigawatt facility, this scales to approximately 50,000 acres. While this sounds enormous, it's comparable to the land set aside for historical megaprojects like the Manhattan Project's Hanford and Oak Ridge sites. The key is that the data center becomes an island, connected to the world only by fiber optic cables, free from the constraints of the public grid.
America's Real Challenge: Overcoming Regulatory Drag
While China is mass-producing solar panels at 20 times the rate of the US, Hammer doesn't believe America's weakness is a lack of manufacturing know-how. The real bottleneck is self-inflicted regulatory drag.
"The regulatory environment around solar is just insane... You end up having to go through a more stringent environmental review process than if you just wanted to grade the whole thing and cover it in concrete."
Laws like NEPA, originally intended to protect the environment, are now paradoxically hindering the deployment of the very technology needed to decarbonize the industrial stack. For the US to win this industrial race, Hammer contends, it needs a World War II-level of motivation to slash red tape and prioritize the domestic manufacturing and deployment of solar energy.
The Ultimate Endgame: A Silicon-Based Civilization
Looking further, Hammer's first-principles thinking leads to a mind-bending conclusion. The most efficient way to convert stellar energy into cognition is silicon. A solar panel converts photons to electrons, and a GPU uses those electrons for computation.
The final form of civilization might not be human at all, but vast swarms of solar sails, each a thin wafer of silicon with integrated processing, floating in space, directly converting starlight into thought. This represents the ultimate collapse of the energy stack—from a star's fusion core directly to a silicon brain, with nothing in between.
Casey Hammer's vision is a stark reminder that the AI revolution is not happening in the cloud. It's happening in the physical world, and it will be built with silicon, steel, and immense amounts of energy. The future belongs to whoever can build it, fastest.
标题：AI能源终局：Terraform CEO Casey Hammer谈为何太阳能而非天然气将主导未来
摘要：Terraform Industries首席执行官Casey Hammer深入剖析了中美之间的AI产业竞赛。他认为，尽管当前的超大规模数据中心为追求速度而青睐天然气，但太阳能无情的学习曲线和供应弹性使其成为必然的长期赢家。本文将揭示他为离网、吉瓦级AI数据中心设计的蓝图，以及美国为赢得竞争必须克服的监管障碍。
正文:
加州理工学院博士、Terraform Industries创始人Casey Hammer提出了一个观点：AI的未来不仅是算法的较量，更是一场巨大的工业竞赛。赢家将不是编写最佳代码的人，而是能制造最多太阳能电池板、电池和GPU的人。在一场深度对话中，Hammer从第一性原理出发，为我们描绘了一个令人信服的、由太阳能驱动的未来为何不仅可能，而且必然的图景。
短期现实：为何今天的AI依赖天然气
在建设吉瓦级数据中心时，像Meta和xAI这样的AI巨头目前正转向天然气。原因很简单：速度和可用性。在这场部署价值数万亿美元GPU的竞赛中，电力的成本几乎可以忽略不计。
“超大规模数据中心对电力成本不敏感，但对电力的可用性极为敏感，”Hammer解释道。
对于像xAI在孟菲斯的“巨像一号”这样的项目，接入现有的天然气管道并部署车载燃气轮机是让一吉瓦电力上线的'最快方式。这种方法对最初的几个数据中心或许有效，但Hammer认为它很快就会撞上制造业的“墙”。
- 供应瓶颈：大型燃气轮机、变压器和其他电网组件的全球产能是有限的。你不可能在一夜之间将产量提高10倍。
- 电网限制：美国电网正在老化，而且众所周知，由于需要数年时间的监管和许可障碍，扩建极其困难。
因此，依赖天然气是一场战役中的战术选择，但并非赢得整场战争的战略。
长期必然：太阳能不可阻挡的学习曲线
Hammer的核心论点基于一个强大的经济原则：赖特定律（Wright's Law），即学习曲线效应。对于太阳能电池板来说，这几乎是一种神奇的力量。
“太阳能的赖特系数是43%。这意味着产量每翻一番，我们就能获得43%的成本降低。”
这种持续了几十年的无情成本下降，意味着太阳能的经济学与其他任何能源都根本不同。燃气轮机是复杂的机械工程产品，其成本曲线相对平坦；而太阳能电池板是一种半导体产品，其成本会随着规模的扩大而呈指数级下降。
这创造了一个良性循环：
- 更低的价格创造了爆炸性的需求。
- 爆炸性的需求推动产量大约每2.5年翻一番。
- 产量的翻番再次触发43%的成本下降。
Hammer认为，这就是为什么做空太阳能，就等于是在对抗历史上最强大、最持久的技术趋势之一。
未来蓝图：离网的AI超级工厂
解决电网瓶颈的方案是完全绕过它。Hammer构想了一个未来，巨大的AI数据中心将作为独立的离网系统，与其自身的电源共存。
以下是一个追求99.99%在线时间的1兆瓦数据中心机架的配置方案：
- 电源：10英亩的太阳能电池板。这包括了显著的超额配置，以应对阴天和冬季。
- 储能：六辆卡车装载的电池（例如特斯拉Megapacks），提供约24小时的备用电力。
- 地点：位于德克萨斯州或内华达州等土地廉价、阳光充足的偏远未开发地区。
对于一个5吉瓦的设施，这个规模将扩大到约5万英亩。虽然听起来很庞大，但这与历史上为曼哈顿计划的汉福德区和橡树岭等大型项目预留的土地面积相当。关键在于，数据中心变成了一座“孤岛”，仅通过光纤与世界相连，摆脱了公共电网的束缚。
美国的真正挑战：克服监管阻力
尽管中国生产太阳能电池板的速度是美国的20倍，但Hammer并不认为美国的弱点在于缺乏制造技术。真正的瓶颈是自我施加的监管阻力。
“围绕太阳能的监管环境简直是疯了……你最终需要经历比直接平整土地并铺上混凝土更严格的环保审查程序。”
像NEPA这样的法律，最初旨在保护环境，现在却ناقض地阻碍了工业体系脱碳所需关键技术的部署。Hammer坚称，如果美国想赢得这场工业竞赛，就需要拿出二战级别的决心，削减繁文缛节，优先发展国内的太阳能制造和部署。
终极结局：一个硅基文明
看得更远，Hammer的第一性原理思维导向了一个令人脑洞大开的结论。将恒星能量转化为认知最有效的方式是硅。太阳能电池板将光子转化为电子，而GPU则利用这些电子进行计算。
文明的最终形态可能根本不是人类，而是巨大的太阳帆群，每一个都是一片集成了处理器的薄硅片，漂浮在太空中，直接将星光转化为思想。这代表了能源堆栈的终极坍缩——从恒星的聚变核心直接到一个硅基大脑，中间没有任何环节。
Casey Hammer的愿景尖锐地提醒我们，AI革命并非发生在云端。它发生在物理世界，并将由硅、钢铁和巨大的能量建成。未来属于能够最快将其变为现实的人。