AI 的无尽纵向一体化 2011 年，Marc Andreessen 在《华尔街日报》上发表《软件正在吞噬世界》（Software is Eating the World）时，我认为大多数科技圈的人并没有按字面意思去理解这句话。当然，亚马逊击败 Borders 成为图书零售霸主、Netflix 取代 Blockbuster、Spotify 统治音乐行业，这些几乎是板上钉钉的事。人们也很容易想象，软件会为航空、金融服务、农业等传统行业赋能。但当时几乎没人能想到，科技公司会纵向整合到亲自下场开采和提炼铜、锂等矿产的程度——而现在，这正是我们所处的现实。 亚马逊正在购买十多年来美国首次新产出的铜；特斯拉则在德州建成了一座设施，可以把原始矿石直接转化为电池级氢氧化锂。特斯拉于 2023 年 5 月为其内部锂精炼项目破土动工，目标是每年生产足以支持 100 万辆新电动车所需的电池级氢氧化锂。特斯拉去年交付了 165 万辆汽车，因此它仍然需要其他供应商，但如此迅速地实现这种程度的纵向一体化，依然令人震撼。而关于对中国电池供应链过度依赖的警告，其实也只是最近几年才真正被广泛重视。 如今，人们已经普遍意识到：稀土矿物其实并没有那么“稀有”。一家名为 Ionic Mineral Technologies 的公司在去年年底于犹他州发现了16 种不同类型的矿物。这对所有担心中国供应全球 90% 稀土的人来说，无疑是个好消息。矿藏存在固然重要，但真正改变格局的关键在于：能否将它们精炼并转化为可用产品。而科技公司正越来越多地扮演“锚定买家（anchor buyer）”的角色，为超级矿业项目提供需求保障（或者像特斯拉那样直接自己做）。 亚马逊与力拓（Rio Tinto）的合作正是这一逻辑的体现。力拓此前一直在研究一种提炼低品位铜矿的新方法。随着数据中心建设规模爆发式增长，对工业金属的需求急剧上升，铜价在过去两年中几乎翻了一倍。铜在前沿 AI 的关键原材料中或许排不上前列，但如今最大的 data center 单个设施就要使用数以万计的公吨铜来分配和传输电力——电线、电路板、变压器，以及大量各类电气部件，都离不开铜。 从 AWS 的体量来看，与力拓的这笔交易其实并不算大：四年 1.4 万公吨，对于每隔几个月就要组装多吉瓦级设施的公司来说，并不足以改变整体格局。但这个项目可能会解锁全新的铜开采机会。力拓计划采用一种新方法，利用细菌和酸从矿石中提取铜。这种方法过去在经济上不可行，但在铜价上涨、并且有超大规模云厂商提供稳定需求的背景下，他们决定再试一次。如果成功，可行的铜储量总量可能会在事实上增加两到三倍。力拓估计，全球约 70% 的铜储量目前被锁在“不值得花成本去精炼”的矿石中。 软件创新会制造繁荣周期。这些繁荣会带来价格错位，而价格上涨又会创造新的机会。谁能想到，一家起家于在线卖书的公司，竟然会去资助新的采矿技术？而采矿，不正是“吞噬世界”这一过程的又一次自然延伸吗？ AI’s Endless Vertical Integration When Marc Andreessen published “Software is Eating the World” in the WSJ in 2011, I think most people in tech did not interpret it literally. Sure, Amazon would beat Borders in book selling, Netflix would overtake Blockbuster in video, Spotify would dominate in music. That was a given, and it was exciting to think about software powering legacy industries like airlines, financial services, and agriculture, but it was hard to imagine tech companies vertically integrating to the point where they started mining and refining copper and lithium, but here we are. Amazon is buying the first new copper produced in America in over a decade, and Tesla now has a Texas facility that can convert raw ore directly into battery-grade lithium hydroxide. Tesla broke ground on the in-house lithium refinery operation in May of 2023 and is aiming to produce enough battery-grade lithium hydroxide to supply a million new electric vehicles every year. Tesla shipped 1.65 million vehicles last year, so they will still need other suppliers, but it’s remarkable to see this level of vertical integration happen so quickly. The warning signs about overdependence on Chinese battery supply only started flashing a few years ago. It’s now common knowledge that rare-earth minerals aren’t actually all that rare. A company called Ionic Mineral Technologies made a large discovery of 16 different types of minerals in Utah late last year. This was great news for everyone worried about the fact that China supplies 90% of the world’s rare earths. Minerals in the ground are great, but we need to actually refine them and turn them into useful products to shift that balance, and tech companies are increasingly stepping in as anchor buyers to underwrite megaprojects (or doing them in-house like Tesla). Amazon’s partnership with Rio Tinto follows this playbook. Rio Tinto had been working on a new method of refining low-grade copper deposits. As the data center buildout has exploded, demand for industrial metals has spiked. The price of copper has nearly doubled in the past two years. Copper is pretty far down the list of key ingredients for frontier AI, but the biggest data centers are now using tens of thousands of metric tons of copper to move electricity around within a facility. Wires, circuit boards, transformers, and tons of miscellaneous electrical components all need copper. Now the Rio Tinto deal is pretty small for AWS. 14,000 metric tons over four years isn’t enough to move the needle if you’re assembling multi-gigawatt facilities every few months, but the project might wind up unlocking new opportunities in copper extraction. Rio is planning to use a new methodology that extracts copper from ore using bacteria and acid. It was uneconomical before, but with higher copper prices, and more stable demand from the hyperscalers, they’re giving it a shot. If the project works, the total supply of viable copper reserves could effectively triple. Rio Tinto estimates that 70% of global copper reserves are locked away in ores that aren’t worth the refining expense. Software innovation creates booms. These booms can create dislocations in price, but higher prices create new opportunities. Who would have thought an online bookstore would be financing new mining technologies? And doesn’t mining feel like a clear extension of eating the world?